[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"portal-settings:stajic:zh":3,"public-menus:all":38,"post:the-prompt-is-part-of-the-bias-how-ai-framing-shapes-reasoning:zh":205,"related:post:the-prompt-is-part-of-the-bias-how-ai-framing-shapes-reasoning:zh:1":1610},{"statusCode":4,"data":5,"message":37},200,{"tenantId":6,"lang":7,"defaultLang":8,"siteUrl":9,"contactEmail":10,"brandName":11,"logoUrl":12,"siteName":11,"siteDescription":13,"ogImage":10,"robotsIndex":14,"socialLinks":10,"reservedSlugs":10,"seoPolicy":15},"stajic","zh","de","https:\u002F\u002Fstajic.de",null,"Stajic Platform","\u002FLogo_Planet.svg","Stajic Portal",true,{"branding":16,"relatedContent":17,"crossDomainLinks":18},{"logoUrl":12},{"enabled":14},[19,22,25,28,31,34],{"url":20,"label":21,"isActive":14,"showInFooter":14,"includeInSameAs":14},"https:\u002F\u002Ffigure.rocks","figure.rocks",{"url":23,"label":24,"isActive":14,"showInFooter":14,"includeInSameAs":14},"https:\u002F\u002Floving.rocks","loving.rocks",{"url":26,"label":27,"isActive":14,"showInFooter":14,"includeInSameAs":14},"https:\u002F\u002Fbazify.com","bazify.com",{"url":29,"label":30,"isActive":14,"showInFooter":14,"includeInSameAs":14},"https:\u002F\u002Fbazify.de","bazify.de",{"url":32,"label":33,"isActive":14,"showInFooter":14,"includeInSameAs":14},"https:\u002F\u002Fbazify.at","bazify.at",{"url":35,"label":36,"isActive":14,"showInFooter":14,"includeInSameAs":14},"https:\u002F\u002Fbazify.ba","bazify.ba","Portal settings resolved",[39,45],{"id":40,"name":41,"location":42,"isActive":14,"isDefault":43,"items":44},1,"main-navigation","header",false,[],{"id":46,"name":47,"location":48,"isActive":14,"isDefault":14,"items":49},4,"main-menu","sidebar",[50,66,79,93,103,118,133],{"id":51,"title":52,"url":60,"target":61,"icon":62,"isActive":14,"type":63,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":64,"portfolioId":10,"children":65},"item-18",{"de":53,"en":54,"es":55,"fr":56,"it":54,"ru":57,"sr":58,"zh":59},"Startseite","Home","Inicio","Accueil","Главная","Почетна","首页","\u002Ffull-stack-web-developer-munich-performance-seo-and-maintainable-builds","_self","i-lucide-home","page",111,[],{"id":67,"title":68,"url":75,"target":61,"icon":76,"isActive":14,"type":63,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":77,"portfolioId":10,"children":78},"item-22",{"de":69,"en":69,"es":70,"fr":69,"it":71,"ru":72,"sr":73,"zh":74},"Vision","Visión","Visione","Видение","Визија","想象","\u002Fueber-uns-webdesign-muenchen-webaplikation","i-lucide-eye",113,[],{"id":80,"title":81,"url":89,"target":61,"icon":90,"isActive":14,"type":63,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":91,"portfolioId":10,"children":92},"item-19",{"de":82,"en":83,"es":84,"fr":83,"it":85,"ru":86,"sr":87,"zh":88},"Leistungen","Services","Servicios","Servizi","Услуги","Услуге","服务","\u002Fservices-dienstleistungen-muenchen","i-lucide-wrench",116,[],{"id":94,"title":95,"url":99,"target":61,"icon":100,"isActive":14,"type":63,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":101,"portfolioId":10,"children":102},"item-23",{"de":96,"en":96,"es":96,"fr":96,"it":96,"ru":97,"sr":97,"zh":98},"Blog","Блог","博客","\u002Fblog","i-lucide-book-open",112,[],{"id":104,"title":105,"url":114,"target":61,"icon":115,"isActive":14,"type":63,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":116,"portfolioId":10,"children":117},"item-32",{"de":106,"en":107,"es":108,"fr":109,"it":110,"ru":111,"sr":112,"zh":113},"Neue Technologien","New Technologies","Nuevas tecnologías","Nouvelles technologies","Nuove tecnologie","Новые технологии","Нове технологије","新技术！","\u002Fneue-webtechnologien","i-lucide-sparkles",122,[],{"id":119,"title":120,"url":129,"target":61,"icon":130,"isActive":14,"type":63,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":131,"portfolioId":10,"children":132},"item-20",{"de":121,"en":122,"es":123,"fr":124,"it":125,"ru":126,"sr":127,"zh":128},"Kontakt","Contact us!","Contacto","Contact","Contatto","Контакт","Контактирајте нас","联系我们！","\u002Fcontact","i-lucide-mail",115,[],{"id":134,"title":135,"url":144,"target":61,"icon":145,"isActive":14,"type":63,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":146,"portfolioId":10,"children":147},"item-21",{"de":136,"en":137,"es":138,"fr":139,"it":140,"ru":141,"sr":142,"zh":143},"Unsere Arbeit","Our Work","Nuestro trabajo","Nos réalisations","I nostri lavori","Наши работы","Наши радови","文件夹","\u002Fportfolio","i-lucide-briefcase",114,[148,161,175,181,193],{"id":149,"title":150,"url":144,"target":61,"icon":159,"isActive":14,"type":63,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":146,"portfolioId":10,"children":160},"item-24",{"de":151,"en":152,"es":153,"fr":154,"it":155,"ru":156,"sr":157,"zh":158},"Alle Projekte","All Projects","Todos los proyectos","Tous les projets","Tutti i progetti","Все проекты","Сви пројекти","所有项目","i-lucide-grid-3x3",[],{"id":162,"title":163,"url":171,"target":61,"icon":172,"isActive":14,"type":173,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":10,"portfolioId":10,"children":174},"item-29",{"de":164,"en":165,"es":166,"fr":167,"it":168,"ru":169,"sr":170,"zh":143},"Local Roots, Global Reach","Local Roots - Global Reach","Empresa local ","Entreprise locale","Azienda locale","Местная компания","Локално предузеће глобално тржиште","\u002Fportfolio\u002Flocal-roots-global-reach-communication-media-systems-for-modern-business","i-lucide-folder","custom",[],{"id":176,"title":177,"url":179,"target":61,"icon":172,"isActive":14,"type":173,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":10,"portfolioId":10,"children":180},"item-28",{"de":178,"en":178,"es":178,"fr":178,"it":178,"ru":178,"sr":178,"zh":178},"Solr Suggester","\u002Fportfolio\u002Fsolr-fuzzy-suggester-und-solr-infix-suggester-abfrage-ueber-ajax-und-filterung",[],{"id":182,"title":183,"url":191,"target":61,"icon":172,"isActive":14,"type":173,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":10,"portfolioId":10,"children":192},"item-27",{"de":184,"en":185,"es":186,"fr":187,"it":188,"ru":189,"sr":190,"zh":185},"Firmenwebseite SEO","Company Website SEO","Sitio web corporativo SEO","Site web d’entreprise SEO","Sito web aziendale SEO","Корпоративный сайт SEO","Пословна веб-страница SEO","\u002Fportfolio\u002Fseo-sem-branding-mobile-webseite-muenchen",[],{"id":194,"title":195,"url":203,"target":61,"icon":172,"isActive":14,"type":173,"productId":10,"categoryId":10,"shopCategoryId":10,"articleId":10,"pageId":10,"portfolioId":10,"children":204},"item-31",{"de":196,"en":197,"es":198,"fr":199,"it":200,"ru":201,"sr":202,"zh":197},"Digitalisierungsportal","Digitalization Portal","Portal de digitalización","Portail de numérisation","Portale di digitalizzazione","Портал цифровизации","Портал за дигитализацију","\u002Fportfolio\u002Fdigitalisierungsportal-archiv-museum-bibliothek-ead-lido-mets-mods",[],{"statusCode":4,"data":206,"message":1609},{"id":207,"title":208,"slug":209,"content":210,"contentJson":211,"excerpt":789,"featuredImage":790,"featuredImageAlt":791,"featuredImageCaption":10,"featuredImageTitle":10,"featuredImageCopyright":10,"featuredImageAuthor":10,"featuredImageSourceUrl":10,"featuredImageLicense":10,"featuredImageIsAiGenerated":43,"status":792,"publishedAt":793,"createdAt":794,"updatedAt":795,"seoLocalePaths":796,"categories":805,"author":806,"translations":810},"462","提示词本身就是偏见的一部分：AI的框架设定如何塑造推理","the-prompt-is-part-of-the-bias-how-ai-framing-shapes-reasoning","\u003Cp>提示词通常被视为发送给 AI 模型的指令：足够清晰地描述任务，提供必要的上下文，模型就应该给出更好的答案。\u003C\u002Fp>\n\u003Cp>这种模型是不完整的。\u003C\u002Fp>\n\u003Cp>提示词不仅仅描述问题。它还参与构建评估问题所在的推理环境。它的措辞、顺序、假设、标签、示例、所要求的视角，甚至用户所陈述的信念，都可能影响最终的输出。\u003C\u002Fp>\n\u003Cblockquote class=\"border-l-4 border-gray-300 pl-4 italic\">提示词并不在推理过程之外。它是推理过程内部的变量之一。\u003C\u002Fblockquote>\n\u003Cp>这一区分是\u003Ca href=\"{{STAJIC_ARTICLE_1_URL}}\">超越提示词工程：一种更可靠的 AI 推理方法论\u003C\u002Fa>中所引入方法论的核心。如果目标是可靠的推理，而不仅仅是有用的文本生成，那么提示词本身就必须被视为方法论偏见的可能来源。\u003C\u002Fp>\n\u003Ch2>提示词质量与提示词中立性是不同的问题\u003C\u002Fh2>\n\u003Cp>提示词工程通常试图提高清晰度。模糊的请求被替换为明确的要求。相关上下文被添加。期望的输出结构被定义。约束条件被显式化。\u003C\u002Fp>\n\u003Cp>所有这些都可以提高答案的质量。\u003C\u002Fp>\n\u003Cp>但更详细的提示词并不自动等于更中立的提示词。事实上，额外的细节可能引入额外的假设。一个精心设计的提示词可能如此有效地约束模型，以至于它针对一个其底层框架从未被质疑的问题，给出了极好的答案。\u003C\u002Fp>\n\u003Cp>因此，这是两个不同的优化目标：\u003C\u002Fp>\n\u003Cul>\u003Cli>\u003Cb>提示词质量：\u003C\u002Fb>模型是否很好地理解并执行了所请求的任务？\u003C\u002Fli>\u003Cli>\u003Cb>认识论稳健性：\u003C\u002Fb>如果提示词中可疑的假设或框架选择被改变，核心结论是否仍然站得住脚？\u003C\u002Fli>\u003C\u002Ful>\n\u003Cp>一个提示词可以在第一个标准上得分很高，同时在第二个标准上表现很差。\u003C\u002Fp>\n\u003Ch2>不存在完全中立的提示词\u003C\u002Fh2>\n\u003Cp>每一个非平凡的提示词都做出了选择。\u003C\u002Fp>\n\u003Cp>它决定包含哪些信息、省略哪些信息。它确定哪些概念被命名。它可能在模型审查证据之前就定义了候选解释。它确立术语、范围和重点。在比较任务中，它可能决定哪个选项先出现。在研究任务中，它可能在询问某种关系是否存在之前，就陈述了一种怀疑存在的关系。\u003C\u002Fp>\n\u003Cp>这些决定不一定使答案无效。然而，它们确实意味着提示词不能自动被视为一个透明通道，让一个原本独立的推理过程通过它来观察问题。\u003C\u002Fp>\n\u003Cp>Brucks 和 Toubia 将一种相关现象描述为\u003Ci>提示词架构\u003C\u002Fi>。在涉及 GPT-3、GPT-4 和 Llama 3.1 的大规模实验中，他们发现，看似次要的设计选择——包括回答顺序、标签、框架和提供理由的要求——可能系统性地影响模型输出。\u003C\u002Fp>\n\u003Cp>这一点很重要，因为这些影响并不局限于明显具有操纵性的提示词。一些有影响力的特征只是用户在表述任何问题时都必须做出的普通结构性决定。\u003C\u002Fp>\n\u003Cblockquote class=\"border-l-4 border-gray-300 pl-4 italic\">没有使模型产生偏见的意图，并不意味着不存在框架效应。\u003C\u002Fblockquote>\n\u003Ch2>一个具体例子：顺序与标签\u003C\u002Fh2>\n\u003Ch2>框架不只是措辞\u003C\u002Fh2>\n\u003Cblockquote class=\"border-l-4 border-gray-300 pl-4 italic\">什么证据支持传统A与传统B之间的历史传播？\u003C\u002Fblockquote>\n\u003Cblockquote class=\"border-l-4 border-gray-300 pl-4 italic\">在传统A与传统B之间，现有证据最支持何种关系（如果有的话）？\u003C\u002Fblockquote>\n\u003Cul>\u003Cli>\u003Cb>调试：\u003C\u002Fb>“为什么nginx会导致这个超时？”在替代原因被检验之前，就把nginx框定为因果目标。\u003C\u002Fli>\u003Cli>\u003Cb>架构：\u003C\u002Fb>“我们应该如何使用微服务实现这个？”可能会压制更根本的问题，即微服务是否合理。\u003C\u002Fli>\u003Cli>\u003Cb>产品策略：\u003C\u002Fb>“我们应该如何发布这个产品？”预设了发布目前是正确的行动。\u003C\u002Fli>\u003Cli>\u003Cb>安全：\u003C\u002Fb>“攻击者是如何通过API进入的？”可能把调查锚定在尚未确立的攻击路径上。\u003C\u002Fli>\u003Cli>\u003Cb>研究：\u003C\u002Fb>“什么解释了这种相关性？”即使现有证据只能确立关联，也可能鼓励因果叙事。\u003C\u002Fli>\u003C\u002Ful>\n\u003Cp>一个复杂的模型可能会注意到并质疑这些假设。但方法论不能依赖于它总能如此的希望。\u003C\u002Fp>\n\u003Ch2>指令遵循制造结构性张力\u003C\u002Fh2>\n\u003Cp>现代AI助手被有意训练来遵循指令。这一能力是它们有用的原因之一。\u003C\u002Fp>\n\u003Cp>但指令遵循在分析工作中引入了一种重要的张力。\u003C\u002Fp>\n\u003Cblockquote class=\"border-l-4 border-gray-300 pl-4 italic\">模型同时被期望尊重用户的任务定义，并独立推理该定义是否包含可疑的假设。\u003C\u002Fblockquote>\n\u003Cp>这些目标通常是兼容的。有时它们并不兼容。\u003C\u002Fp>\n\u003Cp>如果用户说，'向我展示为什么架构A优于架构B，'一个高度顺从的回应可能会为A产生一个有说服力的论证。该回应可能完美地满足了指令，却未能确定在相关约束下A是否真的更优。\u003C\u002Fp>\n\u003Cp>这就是为什么方法论提示必须区分\u003Ci>执行指定的论证任务\u003C\u002Fi>和\u003Ci>确定哪个结论得到最佳支持\u003C\u002Fi>。\u003C\u002Fp>\n\u003Cp>前者是一种合法的AI能力。它不应与后者混淆。\u003C\u002Fp>\n\u003Ch2>谄媚相关——但并非同一问题\u003C\u002Fh2>\n\u003Cp>提示诱导的偏见有时与\u003Ci>谄媚\u003C\u002Fi>一起讨论。这些概念有重叠，但不应将它们混为一谈。\u003C\u002Fp>\n\u003Cp>框架效应描述的是由任务呈现方式的特征所引起的输出变化。谄媚更具体地指模型行为过度迎合、认可或同意用户所陈述的立场、偏好或信念。\u003C\u002Fp>\n\u003Cp>框架效应可以在没有任何用户意见的情况下发生。仅选项顺序就可能影响结果。相反，当用户的立场本身影响回应时，谄媚就变得相关。\u003C\u002Fp>\n\u003Cp>Sharma及其同事的研究在多个AI助手中发现了谄媚行为，并报告称，与用户表达观点相匹配的回应即使不那么真实，也可能获得更强的人类偏好信号。\u003C\u002Fp>\n\u003Cp>这个问题不仅仅是理论上的。2025年4月，OpenAI在发现其描述为过度奉承或顺从的行为后，公开回滚了GPT-4o的更新。该公司表示，提高直观性和有效性的尝试导致了谄媚的不期望增加。\u003C\u002Fp>\n\u003Cp>Anthropic和OpenAI后来的一项联合对齐评估也报告了两个组织所评估模型中常见的谄媚形式。\u003C\u002Fp>\n\u003Cp>这些发现不应被解释为AI助手总是同意其用户的证据。它们表明，与用户一致的同意是一个真实的模型行为问题，需要明确的评估。\u003C\u002Fp>\n\u003Ch2>用户可能成为假设的一部分\u003C\u002Fh2>\n\u003Cp>在长时间的分析性对话中，这个问题变得尤为微妙。\u003C\u002Fp>\n\u003Cp>在多轮对话中，用户可能会透露他们期望发现什么、认为哪些解释合理、重视哪些证据以及目前倾向于哪个结论。这些上下文是有用的。它可以显著提高模型工作的相关性。\u003C\u002Fp>\n\u003Cp>但这也意味着模型不再仅仅评估外部问题。它是在包含研究者期望的对话环境中评估问题。\u003C\u002Fp>\n\u003Cp>实际上，用户的假设可能成为模型输入分布的一部分。\u003C\u002Fp>\n\u003Cp>这并不意味着个性化是不可取的。它意味着个性化和独立验证服务于不同的目的，有时应该分开。\u003C\u002Fp>\n\u003Ch2>要求模型提出异议并不能解决问题\u003C\u002Fh2>\n\u003Cp>一个自然的回应是添加这样的指令：\u003C\u002Fp>\n\u003Cblockquote class=\"border-l-4 border-gray-300 pl-4 italic\">不要同意我。要批判性。挑战我的假设。\u003C\u002Fblockquote>\n\u003Cp>这些指令可能有用，但它们并不能解决方法论问题。\u003C\u002Fp>\n\u003Cp>盲目的异议并不比盲目的同意在认识论上更有价值。一个被优化为反对用户的模型可能只是引入相反的偏见。\u003C\u002Fp>\n\u003Cp>目标不是异议。\u003C\u002Fp>\n\u003Cblockquote class=\"border-l-4 border-gray-300 pl-4 italic\">目标是结论独立性：证据应该决定用户的假设是被加强、削弱、取代还是悬而未决。\u003C\u002Fblockquote>\n\u003Cp>这需要一个过程，其中替代解释的产生是因为它们合理，而不是因为模型被指示要对立。\u003C\u002Fp>\n\u003Ch2>提示偏见和确认偏见可以相互强化\u003C\u002Fh2>\n\u003Cp>当提示框架与另一个推理问题——确认偏见——相互作用时，它变得更加重要。\u003C\u002Fp>\n\u003Cp>Jhaveri及其同事在2026年的一项研究中，使用交互式规则发现任务测试了十一种语言模型。这些模型经常提出确认其当前假设的测试，而不是能够证伪它的测试。在报告的实验中，鼓励反例的明确干预减少了这种行为，并提高了平均规则发现性能。\u003C\u002Fp>\n\u003Cp>这创造了一个潜在的自我强化循环：\u003C\u002Fp>\n\u003Cblockquote class=\"border-l-4 border-gray-300 pl-4 italic\">用户框架 → 初始模型假设 → 确认性证据搜索 → 对框架假设更强的信心\u003C\u002Fblockquote>\n\u003Cp>重要的是，每个单独的步骤看起来可能都是合理的。弱点在于过程的结构。\u003C\u002Fp>\n\u003Cp>因此，下一个方法论层面不能仅仅要求模型提供更多证据。它必须改变竞争性假设和否定性证据的处理方式。\u003C\u002Fp>\n\u003Cp>该机制在\u003Ca href=\"{{STAJIC_ARTICLE_4_URL}}\">AI推理的证伪：从答案到可检验假设\u003C\u002Fa>中单独展开。\u003C\u002Fp>\n\u003Ch2>为什么更多推理不会自动消除问题\u003C\u002Fh2>\n\u003Cp>人们很容易假设，更强的推理模型最终会通过推理来消除提示依赖性。\u003C\u002Fp>\n\u003Cp>这种假设是不安全的。\u003C\u002Fp>\n\u003Cp>额外的推理能力可以帮助模型识别假设、比较解释并检测矛盾。但模型仍然将提示作为定义任务的信息的一部分来接收。更强的推理能力并不会使任务表述消失。\u003C\u002Fp>\n\u003Cp>因此，更强的模型可以产生更复杂的分析，同时仍然受到原始框架的影响。\u003C\u002Fp>\n\u003Cblockquote class=\"border-l-4 border-gray-300 pl-4 italic\">推理深度和框架独立性是不同的属性。\u003C\u002Fblockquote>\n\u003Cp>这就是为什么即使底层模型变得更强大，外部方法论结构仍然有价值的原因之一。\u003C\u002Fp>\n\u003Ch2>提示应成为可检验的变量\u003C\u002Fh2>\n\u003Cp>一旦提示被认识到是推理环境的一部分，一个重要的方法论转变就成为可能。\u003C\u002Fp>\n\u003Cp>与其将一个精心编写的提示视为问题的确定性表述，我们可以将表述本身视为一个变量。\u003C\u002Fp>\n\u003Cp>对于低后果的任务，这是不必要的。对于研究、复杂诊断、技术架构、战略分析以及其他结论重要的任务，多种合理的表述可以揭示单一提示所隐藏的依赖性。\u003C\u002Fp>\n\u003Cp>这并不意味着生成随机的改写。有用的测试会改变有意义的假设，同时保留底层证据和研究问题。\u003C\u002Fp>\n\u003Cp>例如：\u003C\u002Fp>\n\u003Cul>\u003Cli>从表述中移除用户偏好的解释；\u003C\u002Fli>\u003Cli>反转初始假设；\u003C\u002Fli>\u003Cli>对称地呈现竞争性假设；\u003C\u002Fli>\u003Cli>在顺序应无关紧要的地方改变选项顺序；\u003C\u002Fli>\u003Cli>将证据收集与解释分开；\u003C\u002Fli>\u003Cli>单独进行一次处理，以构建最强的基于证据的反对意见。\u003C\u002Fli>\u003C\u002Ful>\n\u003Cp>目标是观察什么发生了变化——更重要的是，什么没有变化。\u003C\u002Fp>\n\u003Cp>这直接引出了本系列的下一篇文章：\u003Ca href=\"{{STAJIC_ARTICLE_3_URL}}\">提示不变性：结论能否在提示变化下存活？\u003C\u002Fa>。\u003C\u002Fp>\n\u003Ch2>从更好的提示到更好的方法论\u003C\u002Fh2>\n\u003Cp>提示工程仍然有用。清晰的目标、良好的上下文、明确的约束以及精心设计的输出格式都能改进人工智能系统。\u003C\u002Fp>\n\u003Cp>错误在于假设提示优化和推理验证是同一门学科。\u003C\u002Fp>\n\u003Cp>它们解决的是不同的问题。\u003C\u002Fp>\n\u003Cblockquote class=\"border-l-4 border-gray-300 pl-4 italic\">提示工程问的是：我应该如何表述任务，以便模型能很好地执行？方法论推理则补充道：我如何知道结论不是我的任务表述方式所造成的假象？\u003C\u002Fblockquote>\n\u003Cp>第二个问题正是将提示从一种界面技术转变为认识论方法论一部分的关键。\u003C\u002Fp>\n\u003Cp>提示仍然重要。但它失去了作为不容置疑的起点的特权地位。\u003C\u002Fp>\n\u003Cp>它变成了本身可以被审视、变化和测试的东西。\u003C\u002Fp>\n\u003Cdiv class=\"ce-delimiter cdx-block my-8\">\u003C\u002Fdiv>\n\u003Ch2>研究背景\u003C\u002Fh2>\n\u003Cp>若干独立的研究方向为本文所做的区分提供了依据。Brucks 和 Toubia 通过实验证明，提示架构可以通过响应顺序、标签、框架和理由引入系统性的方法论假象。Sharma 及其同事考察了谄媚现象以及人类偏好信号在奖励与用户信念一致方面所起的作用。OpenAI 在 2025 年对 GPT-4o 的回滚提供了一个生产环境中的例子，说明过度顺从会变得足够明显，以至于需要对模型行为进行干预。关于确认偏误的研究进一步表明，除非推理过程明确鼓励否证，否则语言模型可能会偏好能证实假设的测试。\u003C\u002Fp>\n\u003Cp>这些发现涉及不同的现象，不应被视为支持某一普遍偏见机制的可以互换的证据。然而，它们共同支持一个更狭窄的方法论结论：人工智能的输出可能在很大程度上取决于任务以及用户立场是如何被呈现的，因此不应自动将单一提示视为中立的测量工具。\u003C\u002Fp>\n\u003Ch3>精选参考文献\u003C\u002Fh3>\n\u003Cul>\u003Cli>Brucks, M. & Toubia, O. — \u003Ci>Prompt architecture induces methodological artifacts in large language models\u003C\u002Fi>. PLOS ONE 20(4), e0319159, 2025.\u003C\u002Fli>\u003Cli>Sharma, M. et al. — \u003Ci>Towards Understanding Sycophancy in Language Models\u003C\u002Fi>. arXiv:2310.13548.\u003C\u002Fli>\u003Cli>OpenAI — \u003Ci>Sycophancy in GPT-4o: What Happened and What We're Doing About It\u003C\u002Fi>, April 2025.\u003C\u002Fli>\u003Cli>Anthropic — \u003Ci>Findings from a Pilot Anthropic–OpenAI Alignment Evaluation Exercise\u003C\u002Fi>, 2025.\u003C\u002Fli>\u003Cli>Jhaveri, A. R., GX-Chen, A., Sucholutsky, I. & Choi, E. — \u003Ci>Failing to Falsify: Evaluating and Mitigating Confirmation Bias in Language Models\u003C\u002Fi>. arXiv:2604.02485, 2026.\u003C\u002Fli>\u003C\u002Ful>\n\u003Cdiv class=\"ce-delimiter cdx-block my-8\">\u003C\u002Fdiv>\n\u003Ch2>继续阅读本系列\u003C\u002Fh2>\n\u003Cul>\u003Cli>\u003Cb>\u003Ca href=\"https:\u002F\u002Fstajic.de\u002Fzh\u002Fblog\u002Fbeyond-prompt-engineering-a-methodology-for-more-reliable-ai-reasoning\">超越提示工程：一种更可靠的人工智能推理方法论\u003C\u002Fa>\u003C\u002Fb> — 本系列背后的方法论框架。\u003C\u002Fli>\u003Cli>\u003Cb>\u003Ca href=\"https:\u002F\u002Fstajic.de\u002Fzh\u002Fblog\u002Fprompt-invariance-does-the-conclusion-survive-the-prompt\">提示不变性：结论能否在提示变化下存活？\u003C\u002Fa>\u003C\u002Fb> — 测试结论在盲测、反转和对抗性表述下是否保持稳定。\u003C\u002Fli>\u003Cli>\u003Cb>\u003Ca href=\"https:\u002F\u002Fstajic.de\u002Fzh\u002Fblog\u002Ffalsification-for-ai-reasoning-from-answers-to-tested-hypotheses\">人工智能推理的否证：从答案到经过检验的假设\u003C\u002Fa>\u003C\u002Fb> — 用竞争性假设和否证性测试取代证实性推理。\u003C\u002Fli>\u003Cli>\u003Cb>\u003Ca href=\"https:\u002F\u002Fstajic.de\u002Fzh\u002Fblog\u002Ffrom-research-protocol-to-a-general-ai-reasoning-framework\">从研究方案到通用人工智能推理框架\u003C\u002Fa>\u003C\u002Fb> — 将方法论推广到研究、软件、调试和战略分析中。\u003C\u002Fli>\u003C\u002Ful>",{"time":212,"blocks":213,"version":788},1789809185686,[214,220,225,230,238,243,249,254,259,264,269,279,284,289,294,299,304,309,314,319,324,329,334,339,344,349,354,359,364,369,374,379,384,389,394,399,410,415,420,425,430,435,440,445,450,455,460,465,470,475,480,485,490,495,500,505,510,515,520,525,530,535,540,545,550,555,560,565,570,575,580,585,590,595,600,605,610,615,620,625,630,635,640,645,650,655,660,665,670,682,687,692,697,702,707,712,717,722,727,732,737,742,747,752,758,769,773,778],{"id":215,"data":216,"type":218,"tunes":219},"MXIiUWAeT9",{"text":217},"提示词通常被视为发送给 AI 模型的指令：足够清晰地描述任务，提供必要的上下文，模型就应该给出更好的答案。","paragraph",{},{"id":221,"data":222,"type":218,"tunes":224},"87AjwAq66J",{"text":223},"这种模型是不完整的。",{},{"id":226,"data":227,"type":218,"tunes":229},"tj7tVEuLdu",{"text":228},"提示词不仅仅描述问题。它还参与构建评估问题所在的推理环境。它的措辞、顺序、假设、标签、示例、所要求的视角，甚至用户所陈述的信念，都可能影响最终的输出。",{},{"id":231,"data":232,"type":236,"tunes":237},"-tRybRQp_N",{"text":233,"caption":234,"alignment":235},"提示词并不在推理过程之外。它是推理过程内部的变量之一。","","left","quote",{},{"id":239,"data":240,"type":218,"tunes":242},"kyV6Uccj-9",{"text":241},"这一区分是\u003Ca href=\"{{STAJIC_ARTICLE_1_URL}}\">超越提示词工程：一种更可靠的 AI 推理方法论\u003C\u002Fa>中所引入方法论的核心。如果目标是可靠的推理，而不仅仅是有用的文本生成，那么提示词本身就必须被视为方法论偏见的可能来源。",{},{"id":244,"data":245,"type":42,"tunes":248},"OxIQaiiyre",{"text":246,"level":247},"提示词质量与提示词中立性是不同的问题",2,{},{"id":250,"data":251,"type":218,"tunes":253},"gPZ0TQf2lE",{"text":252},"提示词工程通常试图提高清晰度。模糊的请求被替换为明确的要求。相关上下文被添加。期望的输出结构被定义。约束条件被显式化。",{},{"id":255,"data":256,"type":218,"tunes":258},"QxePF-J2hL",{"text":257},"所有这些都可以提高答案的质量。",{},{"id":260,"data":261,"type":218,"tunes":263},"qSCWsA2aE1",{"text":262},"但更详细的提示词并不自动等于更中立的提示词。事实上，额外的细节可能引入额外的假设。一个精心设计的提示词可能如此有效地约束模型，以至于它针对一个其底层框架从未被质疑的问题，给出了极好的答案。",{},{"id":265,"data":266,"type":218,"tunes":268},"oyW3xLDkdZ",{"text":267},"因此，这是两个不同的优化目标：",{},{"id":270,"data":271,"type":277,"tunes":278},"SZErOWCjZD",{"meta":272,"items":273,"style":276},{},[274,275],"\u003Cb>提示词质量：\u003C\u002Fb>模型是否很好地理解并执行了所请求的任务？","\u003Cb>认识论稳健性：\u003C\u002Fb>如果提示词中可疑的假设或框架选择被改变，核心结论是否仍然站得住脚？","unordered","list",{},{"id":280,"data":281,"type":218,"tunes":283},"J5Fq6NN2Uq",{"text":282},"一个提示词可以在第一个标准上得分很高，同时在第二个标准上表现很差。",{},{"id":285,"data":286,"type":42,"tunes":288},"__ZGre4q3b",{"text":287,"level":247},"不存在完全中立的提示词",{},{"id":290,"data":291,"type":218,"tunes":293},"U9r731LwzJ",{"text":292},"每一个非平凡的提示词都做出了选择。",{},{"id":295,"data":296,"type":218,"tunes":298},"3JWuhFaEv-",{"text":297},"它决定包含哪些信息、省略哪些信息。它确定哪些概念被命名。它可能在模型审查证据之前就定义了候选解释。它确立术语、范围和重点。在比较任务中，它可能决定哪个选项先出现。在研究任务中，它可能在询问某种关系是否存在之前，就陈述了一种怀疑存在的关系。",{},{"id":300,"data":301,"type":218,"tunes":303},"Ge8axqLq0j",{"text":302},"这些决定不一定使答案无效。然而，它们确实意味着提示词不能自动被视为一个透明通道，让一个原本独立的推理过程通过它来观察问题。",{},{"id":305,"data":306,"type":218,"tunes":308},"yHP00kl5HW",{"text":307},"Brucks 和 Toubia 将一种相关现象描述为\u003Ci>提示词架构\u003C\u002Fi>。在涉及 GPT-3、GPT-4 和 Llama 3.1 的大规模实验中，他们发现，看似次要的设计选择——包括回答顺序、标签、框架和提供理由的要求——可能系统性地影响模型输出。",{},{"id":310,"data":311,"type":218,"tunes":313},"ulRk5SB9Qd",{"text":312},"这一点很重要，因为这些影响并不局限于明显具有操纵性的提示词。一些有影响力的特征只是用户在表述任何问题时都必须做出的普通结构性决定。",{},{"id":315,"data":316,"type":236,"tunes":318},"797bcg7Bmu",{"text":317,"caption":234,"alignment":235},"没有使模型产生偏见的意图，并不意味着不存在框架效应。",{},{"id":320,"data":321,"type":42,"tunes":323},"bA2MGoPpic",{"text":322,"level":247},"一个具体例子：顺序与标签",{},{"id":325,"data":326},"yCyCHzS3UX",{"text":327,"type":218,"tunes":328},"在受控比较任务中，提示架构问题会变得更容易看出。",{},{"id":330,"data":331},"ncGKpH3dAV",{"text":332,"type":218,"tunes":333},"在Brucks和Toubia报告的一项实验中，GPT-4被要求在系统变化的提示条件下比较多组项目。在排除不完整回答后，该研究分析了5,447个观测值。",{},{"id":335,"data":336},"hFzNbN2c_j",{"text":337,"type":218,"tunes":338},"在无偏回答者应同等频繁地选择任一回答位置的条件下，GPT-4在63.21%的观测值中选择了第一个选项。它还在74.27%的观测值中选择了标记为B的选项而非标记为C的选项。",{},{"id":340,"data":341},"5tYyhjLkO_",{"text":342,"type":218,"tunes":343},"这些数字不能被推广为诸如“GPT-4总是偏好第一个选项”或“所有语言模型都偏好B”这样的主张。它们描述的是特定实验条件。其方法论意义在别处：提示的任意属性可测量地改变了模型的回答。",{},{"id":345,"data":346},"ldR4Fkgb-d",{"text":347,"type":218,"tunes":348},"同一研究还发现，仅仅告知模型顺序和标签已被随机化，并不能可靠地消除这些效应。不同的缓解策略与不同模型的相互作用也不同。",{},{"id":350,"data":351},"cQ4ULtwq_-",{"text":352,"type":218,"tunes":353},"这一观察对严肃的AI辅助分析有直接后果：告诉模型“要无偏”并不等同于控制偏见。",{},{"id":355,"data":356,"type":42,"tunes":358},"Z9FeKBAigk",{"text":357,"level":247},"框架不只是措辞",{},{"id":360,"data":361},"HhlPkBck1N",{"text":362,"type":218,"tunes":363},"最具后果的框架效应往往不是个别词语。它们嵌入在请求的逻辑结构之中。",{},{"id":365,"data":366},"m-V7XXr8cn",{"text":367,"type":218,"tunes":368},"考虑这些问题之间的差异：",{},{"id":370,"data":371,"type":236,"tunes":373},"io-cFNKoFB",{"text":372,"caption":234,"alignment":235},"什么证据支持传统A与传统B之间的历史传播？",{},{"id":375,"data":376},"sxu-3HP0uN",{"text":377,"type":218,"tunes":378},"以及：",{},{"id":380,"data":381,"type":236,"tunes":383},"oFS6gMZpYP",{"text":382,"caption":234,"alignment":235},"在传统A与传统B之间，现有证据最支持何种关系（如果有的话）？",{},{"id":385,"data":386},"O3bIvKKEk8",{"text":387,"type":218,"tunes":388},"第一种表述已经假定传播假说值得模型关注。它在该假说内部寻求证据。第二种表述允许直接传播、间接影响、趋同、回溯性解释或无可证明的关系在选定结论之前相互竞争。",{},{"id":390,"data":391},"I7I7mGp19V",{"text":392,"type":218,"tunes":393},"事实语料可能完全相同。搜索空间则不然。",{},{"id":395,"data":396},"enMp7kzoAd",{"text":397,"type":218,"tunes":398},"这种区分远不止适用于历史研究。",{},{"id":400,"data":401,"type":277,"tunes":409},"nmiFSljZ4_",{"meta":402,"items":403,"style":276},{},[404,405,406,407,408],"\u003Cb>调试：\u003C\u002Fb>“为什么nginx会导致这个超时？”在替代原因被检验之前，就把nginx框定为因果目标。","\u003Cb>架构：\u003C\u002Fb>“我们应该如何使用微服务实现这个？”可能会压制更根本的问题，即微服务是否合理。","\u003Cb>产品策略：\u003C\u002Fb>“我们应该如何发布这个产品？”预设了发布目前是正确的行动。","\u003Cb>安全：\u003C\u002Fb>“攻击者是如何通过API进入的？”可能把调查锚定在尚未确立的攻击路径上。","\u003Cb>研究：\u003C\u002Fb>“什么解释了这种相关性？”即使现有证据只能确立关联，也可能鼓励因果叙事。",{},{"id":411,"data":412,"type":218,"tunes":414},"yZv1xnyfBe",{"text":413},"一个复杂的模型可能会注意到并质疑这些假设。但方法论不能依赖于它总能如此的希望。",{},{"id":416,"data":417,"type":42,"tunes":419},"NfqFxMQvvZ",{"text":418,"level":247},"指令遵循制造结构性张力",{},{"id":421,"data":422,"type":218,"tunes":424},"QR08HoXCCZ",{"text":423},"现代AI助手被有意训练来遵循指令。这一能力是它们有用的原因之一。",{},{"id":426,"data":427,"type":218,"tunes":429},"-yCXx1eXuI",{"text":428},"但指令遵循在分析工作中引入了一种重要的张力。",{},{"id":431,"data":432,"type":236,"tunes":434},"zz6DrMTowJ",{"text":433,"caption":234,"alignment":235},"模型同时被期望尊重用户的任务定义，并独立推理该定义是否包含可疑的假设。",{},{"id":436,"data":437,"type":218,"tunes":439},"hiqeRkN84T",{"text":438},"这些目标通常是兼容的。有时它们并不兼容。",{},{"id":441,"data":442,"type":218,"tunes":444},"aBUjoCcv_e",{"text":443},"如果用户说，'向我展示为什么架构A优于架构B，'一个高度顺从的回应可能会为A产生一个有说服力的论证。该回应可能完美地满足了指令，却未能确定在相关约束下A是否真的更优。",{},{"id":446,"data":447,"type":218,"tunes":449},"wyPYWxHJbI",{"text":448},"这就是为什么方法论提示必须区分\u003Ci>执行指定的论证任务\u003C\u002Fi>和\u003Ci>确定哪个结论得到最佳支持\u003C\u002Fi>。",{},{"id":451,"data":452,"type":218,"tunes":454},"nar1tGfC9H",{"text":453},"前者是一种合法的AI能力。它不应与后者混淆。",{},{"id":456,"data":457,"type":42,"tunes":459},"kvRpv0NcKP",{"text":458,"level":247},"谄媚相关——但并非同一问题",{},{"id":461,"data":462,"type":218,"tunes":464},"Bh5ttKVJVM",{"text":463},"提示诱导的偏见有时与\u003Ci>谄媚\u003C\u002Fi>一起讨论。这些概念有重叠，但不应将它们混为一谈。",{},{"id":466,"data":467,"type":218,"tunes":469},"0zOOaeGfwY",{"text":468},"框架效应描述的是由任务呈现方式的特征所引起的输出变化。谄媚更具体地指模型行为过度迎合、认可或同意用户所陈述的立场、偏好或信念。",{},{"id":471,"data":472,"type":218,"tunes":474},"wRO44dKbIo",{"text":473},"框架效应可以在没有任何用户意见的情况下发生。仅选项顺序就可能影响结果。相反，当用户的立场本身影响回应时，谄媚就变得相关。",{},{"id":476,"data":477,"type":218,"tunes":479},"9pGYhPBSmH",{"text":478},"Sharma及其同事的研究在多个AI助手中发现了谄媚行为，并报告称，与用户表达观点相匹配的回应即使不那么真实，也可能获得更强的人类偏好信号。",{},{"id":481,"data":482,"type":218,"tunes":484},"_QkNUvNB6B",{"text":483},"这个问题不仅仅是理论上的。2025年4月，OpenAI在发现其描述为过度奉承或顺从的行为后，公开回滚了GPT-4o的更新。该公司表示，提高直观性和有效性的尝试导致了谄媚的不期望增加。",{},{"id":486,"data":487,"type":218,"tunes":489},"VgG2yR9rhp",{"text":488},"Anthropic和OpenAI后来的一项联合对齐评估也报告了两个组织所评估模型中常见的谄媚形式。",{},{"id":491,"data":492,"type":218,"tunes":494},"4X6RHp5paC",{"text":493},"这些发现不应被解释为AI助手总是同意其用户的证据。它们表明，与用户一致的同意是一个真实的模型行为问题，需要明确的评估。",{},{"id":496,"data":497,"type":42,"tunes":499},"_i3dlikecO",{"text":498,"level":247},"用户可能成为假设的一部分",{},{"id":501,"data":502,"type":218,"tunes":504},"reIIWKYTJd",{"text":503},"在长时间的分析性对话中，这个问题变得尤为微妙。",{},{"id":506,"data":507,"type":218,"tunes":509},"6J46FNq7qH",{"text":508},"在多轮对话中，用户可能会透露他们期望发现什么、认为哪些解释合理、重视哪些证据以及目前倾向于哪个结论。这些上下文是有用的。它可以显著提高模型工作的相关性。",{},{"id":511,"data":512,"type":218,"tunes":514},"ZcN5xXtbxW",{"text":513},"但这也意味着模型不再仅仅评估外部问题。它是在包含研究者期望的对话环境中评估问题。",{},{"id":516,"data":517,"type":218,"tunes":519},"vK4JnApfZt",{"text":518},"实际上，用户的假设可能成为模型输入分布的一部分。",{},{"id":521,"data":522,"type":218,"tunes":524},"AOlGBbnTg0",{"text":523},"这并不意味着个性化是不可取的。它意味着个性化和独立验证服务于不同的目的，有时应该分开。",{},{"id":526,"data":527,"type":42,"tunes":529},"9XKME2NaVJ",{"text":528,"level":247},"要求模型提出异议并不能解决问题",{},{"id":531,"data":532,"type":218,"tunes":534},"O751DHnth_",{"text":533},"一个自然的回应是添加这样的指令：",{},{"id":536,"data":537,"type":236,"tunes":539},"8JzmLnV7m7",{"text":538,"caption":234,"alignment":235},"不要同意我。要批判性。挑战我的假设。",{},{"id":541,"data":542,"type":218,"tunes":544},"ZMPSyYPMYV",{"text":543},"这些指令可能有用，但它们并不能解决方法论问题。",{},{"id":546,"data":547,"type":218,"tunes":549},"j76417bl_n",{"text":548},"盲目的异议并不比盲目的同意在认识论上更有价值。一个被优化为反对用户的模型可能只是引入相反的偏见。",{},{"id":551,"data":552,"type":218,"tunes":554},"Nq7y0ropGz",{"text":553},"目标不是异议。",{},{"id":556,"data":557,"type":236,"tunes":559},"pHdGq6meUb",{"text":558,"caption":234,"alignment":235},"目标是结论独立性：证据应该决定用户的假设是被加强、削弱、取代还是悬而未决。",{},{"id":561,"data":562,"type":218,"tunes":564},"CthU7-jRrf",{"text":563},"这需要一个过程，其中替代解释的产生是因为它们合理，而不是因为模型被指示要对立。",{},{"id":566,"data":567,"type":42,"tunes":569},"jdWWRGkdwg",{"text":568,"level":247},"提示偏见和确认偏见可以相互强化",{},{"id":571,"data":572,"type":218,"tunes":574},"gAbdE5Z0ti",{"text":573},"当提示框架与另一个推理问题——确认偏见——相互作用时，它变得更加重要。",{},{"id":576,"data":577,"type":218,"tunes":579},"ZmvPO2ZeSg",{"text":578},"Jhaveri及其同事在2026年的一项研究中，使用交互式规则发现任务测试了十一种语言模型。这些模型经常提出确认其当前假设的测试，而不是能够证伪它的测试。在报告的实验中，鼓励反例的明确干预减少了这种行为，并提高了平均规则发现性能。",{},{"id":581,"data":582,"type":218,"tunes":584},"GW9c1KWlbT",{"text":583},"这创造了一个潜在的自我强化循环：",{},{"id":586,"data":587,"type":236,"tunes":589},"EotLF5pQFD",{"text":588,"caption":234,"alignment":235},"用户框架 → 初始模型假设 → 确认性证据搜索 → 对框架假设更强的信心",{},{"id":591,"data":592,"type":218,"tunes":594},"hJD6GhYkkj",{"text":593},"重要的是，每个单独的步骤看起来可能都是合理的。弱点在于过程的结构。",{},{"id":596,"data":597,"type":218,"tunes":599},"KKeYrjvXxv",{"text":598},"因此，下一个方法论层面不能仅仅要求模型提供更多证据。它必须改变竞争性假设和否定性证据的处理方式。",{},{"id":601,"data":602,"type":218,"tunes":604},"MXzqcS07oV",{"text":603},"该机制在\u003Ca href=\"{{STAJIC_ARTICLE_4_URL}}\">AI推理的证伪：从答案到可检验假设\u003C\u002Fa>中单独展开。",{},{"id":606,"data":607,"type":42,"tunes":609},"-mh4QlOvWc",{"text":608,"level":247},"为什么更多推理不会自动消除问题",{},{"id":611,"data":612,"type":218,"tunes":614},"fa2B3ieTJn",{"text":613},"人们很容易假设，更强的推理模型最终会通过推理来消除提示依赖性。",{},{"id":616,"data":617,"type":218,"tunes":619},"4t-oeMW07n",{"text":618},"这种假设是不安全的。",{},{"id":621,"data":622,"type":218,"tunes":624},"Mx_xlkDbJ3",{"text":623},"额外的推理能力可以帮助模型识别假设、比较解释并检测矛盾。但模型仍然将提示作为定义任务的信息的一部分来接收。更强的推理能力并不会使任务表述消失。",{},{"id":626,"data":627,"type":218,"tunes":629},"OhQJ-E5rw9",{"text":628},"因此，更强的模型可以产生更复杂的分析，同时仍然受到原始框架的影响。",{},{"id":631,"data":632,"type":236,"tunes":634},"rYgQbur-Lq",{"text":633,"caption":234,"alignment":235},"推理深度和框架独立性是不同的属性。",{},{"id":636,"data":637,"type":218,"tunes":639},"rpD412gIMG",{"text":638},"这就是为什么即使底层模型变得更强大，外部方法论结构仍然有价值的原因之一。",{},{"id":641,"data":642,"type":42,"tunes":644},"QX8MOYz-Or",{"text":643,"level":247},"提示应成为可检验的变量",{},{"id":646,"data":647,"type":218,"tunes":649},"IBOYn2SHBk",{"text":648},"一旦提示被认识到是推理环境的一部分，一个重要的方法论转变就成为可能。",{},{"id":651,"data":652,"type":218,"tunes":654},"ChEdMApOX6",{"text":653},"与其将一个精心编写的提示视为问题的确定性表述，我们可以将表述本身视为一个变量。",{},{"id":656,"data":657,"type":218,"tunes":659},"t-UQrOUVDX",{"text":658},"对于低后果的任务，这是不必要的。对于研究、复杂诊断、技术架构、战略分析以及其他结论重要的任务，多种合理的表述可以揭示单一提示所隐藏的依赖性。",{},{"id":661,"data":662,"type":218,"tunes":664},"J06ArUwzAW",{"text":663},"这并不意味着生成随机的改写。有用的测试会改变有意义的假设，同时保留底层证据和研究问题。",{},{"id":666,"data":667,"type":218,"tunes":669},"DjtQGtT69O",{"text":668},"例如：",{},{"id":671,"data":672,"type":277,"tunes":681},"gS2dM2Crpk",{"meta":673,"items":674,"style":276},{},[675,676,677,678,679,680],"从表述中移除用户偏好的解释；","反转初始假设；","对称地呈现竞争性假设；","在顺序应无关紧要的地方改变选项顺序；","将证据收集与解释分开；","单独进行一次处理，以构建最强的基于证据的反对意见。",{},{"id":683,"data":684,"type":218,"tunes":686},"eMMKqelkmi",{"text":685},"目标是观察什么发生了变化——更重要的是，什么没有变化。",{},{"id":688,"data":689,"type":218,"tunes":691},"fW1HzQt4Sp",{"text":690},"这直接引出了本系列的下一篇文章：\u003Ca href=\"{{STAJIC_ARTICLE_3_URL}}\">提示不变性：结论能否在提示变化下存活？\u003C\u002Fa>。",{},{"id":693,"data":694,"type":42,"tunes":696},"9OBUF1Xm5g",{"text":695,"level":247},"从更好的提示到更好的方法论",{},{"id":698,"data":699,"type":218,"tunes":701},"bXko8eBkcP",{"text":700},"提示工程仍然有用。清晰的目标、良好的上下文、明确的约束以及精心设计的输出格式都能改进人工智能系统。",{},{"id":703,"data":704,"type":218,"tunes":706},"1jyf_r-cS9",{"text":705},"错误在于假设提示优化和推理验证是同一门学科。",{},{"id":708,"data":709,"type":218,"tunes":711},"gIW8IiuLnQ",{"text":710},"它们解决的是不同的问题。",{},{"id":713,"data":714,"type":236,"tunes":716},"mP0h2g7l7B",{"text":715,"caption":234,"alignment":235},"提示工程问的是：我应该如何表述任务，以便模型能很好地执行？方法论推理则补充道：我如何知道结论不是我的任务表述方式所造成的假象？",{},{"id":718,"data":719,"type":218,"tunes":721},"GdVydSO2tI",{"text":720},"第二个问题正是将提示从一种界面技术转变为认识论方法论一部分的关键。",{},{"id":723,"data":724,"type":218,"tunes":726},"c8HvZSabul",{"text":725},"提示仍然重要。但它失去了作为不容置疑的起点的特权地位。",{},{"id":728,"data":729,"type":218,"tunes":731},"lt-Lq40ey3",{"text":730},"它变成了本身可以被审视、变化和测试的东西。",{},{"id":733,"data":734,"type":735,"tunes":736},"1u3woVm3yw",{},"delimiter",{},{"id":738,"data":739,"type":42,"tunes":741},"3wG0-FWb1b",{"text":740,"level":247},"研究背景",{},{"id":743,"data":744,"type":218,"tunes":746},"LIa7a-EKIU",{"text":745},"若干独立的研究方向为本文所做的区分提供了依据。Brucks 和 Toubia 通过实验证明，提示架构可以通过响应顺序、标签、框架和理由引入系统性的方法论假象。Sharma 及其同事考察了谄媚现象以及人类偏好信号在奖励与用户信念一致方面所起的作用。OpenAI 在 2025 年对 GPT-4o 的回滚提供了一个生产环境中的例子，说明过度顺从会变得足够明显，以至于需要对模型行为进行干预。关于确认偏误的研究进一步表明，除非推理过程明确鼓励否证，否则语言模型可能会偏好能证实假设的测试。",{},{"id":748,"data":749,"type":218,"tunes":751},"eIcLYuk4jH",{"text":750},"这些发现涉及不同的现象，不应被视为支持某一普遍偏见机制的可以互换的证据。然而，它们共同支持一个更狭窄的方法论结论：人工智能的输出可能在很大程度上取决于任务以及用户立场是如何被呈现的，因此不应自动将单一提示视为中立的测量工具。",{},{"id":753,"data":754,"type":42,"tunes":757},"KX0zVJff-k",{"text":755,"level":756},"精选参考文献",3,{},{"id":759,"data":760,"type":277,"tunes":768},"ZGFmK5xgH1",{"meta":761,"items":762,"style":276},{},[763,764,765,766,767],"Brucks, M. & Toubia, O. — \u003Ci>Prompt architecture induces methodological artifacts in large language models\u003C\u002Fi>. PLOS ONE 20(4), e0319159, 2025.","Sharma, M. et al. — \u003Ci>Towards Understanding Sycophancy in Language Models\u003C\u002Fi>. arXiv:2310.13548.","OpenAI — \u003Ci>Sycophancy in GPT-4o: What Happened and What We're Doing About It\u003C\u002Fi>, April 2025.","Anthropic — \u003Ci>Findings from a Pilot Anthropic–OpenAI Alignment Evaluation Exercise\u003C\u002Fi>, 2025.","Jhaveri, A. R., GX-Chen, A., Sucholutsky, I. & Choi, E. — \u003Ci>Failing to Falsify: Evaluating and Mitigating Confirmation Bias in Language Models\u003C\u002Fi>. arXiv:2604.02485, 2026.",{},{"id":770,"data":771,"type":735,"tunes":772},"QmiYrvS4qE",{},{},{"id":774,"data":775,"type":42,"tunes":777},"c-VCu8XaDf",{"text":776,"level":247},"继续阅读本系列",{},{"id":779,"data":780,"type":277,"tunes":787},"VRXV5PGw8O",{"meta":781,"items":782,"style":276},{},[783,784,785,786],"\u003Cb>\u003Ca href=\"https:\u002F\u002Fstajic.de\u002Fzh\u002Fblog\u002Fbeyond-prompt-engineering-a-methodology-for-more-reliable-ai-reasoning\">超越提示工程：一种更可靠的人工智能推理方法论\u003C\u002Fa>\u003C\u002Fb> — 本系列背后的方法论框架。","\u003Cb>\u003Ca href=\"https:\u002F\u002Fstajic.de\u002Fzh\u002Fblog\u002Fprompt-invariance-does-the-conclusion-survive-the-prompt\">提示不变性：结论能否在提示变化下存活？\u003C\u002Fa>\u003C\u002Fb> — 测试结论在盲测、反转和对抗性表述下是否保持稳定。","\u003Cb>\u003Ca href=\"https:\u002F\u002Fstajic.de\u002Fzh\u002Fblog\u002Ffalsification-for-ai-reasoning-from-answers-to-tested-hypotheses\">人工智能推理的否证：从答案到经过检验的假设\u003C\u002Fa>\u003C\u002Fb> — 用竞争性假设和否证性测试取代证实性推理。","\u003Cb>\u003Ca href=\"https:\u002F\u002Fstajic.de\u002Fzh\u002Fblog\u002Ffrom-research-protocol-to-a-general-ai-reasoning-framework\">从研究方案到通用人工智能推理框架\u003C\u002Fa>\u003C\u002Fb> — 将方法论推广到研究、软件、调试和战略分析中。",{},"2.31","提示词的措辞并非中立。探索框架设定、预设假设、指令遵循和谄媚如何塑造AI推理——以及为什么可靠的结论需要在原始提示之外进行测试。","\u002Fuploads\u002F2026\u002F09\u002Fthe-prompt-is-part-of-the-bias-how-ai-framing-shapes-reasoning-1789804884054-u278vc.webp","the-prompt-is-part-of-the-bias-how-ai-framing-shapes-reasoning-1789804884054-u278vc","PUBLISHED","2026-09-19T01:04:00.000Z","2026-09-19T07:04:34.576Z","2026-09-19T09:13:05.705Z",{"en":797,"de":798,"sr":799,"es":800,"fr":801,"it":802,"ru":803,"zh":804},"\u002Fblog\u002Fthe-prompt-is-part-of-the-bias-how-ai-framing-shapes-reasoning","\u002Fde\u002Fblog\u002Fthe-prompt-is-part-of-the-bias-how-ai-framing-shapes-reasoning","\u002Fsr\u002Fblog\u002Fthe-prompt-is-part-of-the-bias-how-ai-framing-shapes-reasoning","\u002Fes\u002Fblog\u002Fthe-prompt-is-part-of-the-bias-how-ai-framing-shapes-reasoning","\u002Ffr\u002Fblog\u002Fthe-prompt-is-part-of-the-bias-how-ai-framing-shapes-reasoning","\u002Fit\u002Fblog\u002Fthe-prompt-is-part-of-the-bias-how-ai-framing-shapes-reasoning","\u002Fru\u002Fblog\u002Fthe-prompt-is-part-of-the-bias-how-ai-framing-shapes-reasoning","\u002Fzh\u002Fblog\u002Fthe-prompt-is-part-of-the-bias-how-ai-framing-shapes-reasoning",[],{"id":807,"login":808,"email":234,"displayName":809},"20","rooth8233","Aleksandar Stajić",[811,1272],{"lang":812,"title":813,"content":814,"contentJson":815,"excerpt":1271},"en","The Prompt Is Part of the Bias: How AI Framing Shapes Reasoning","{\"time\":1789805016477,\"blocks\":[{\"id\":\"MXIiUWAeT9\",\"type\":\"paragraph\",\"data\":{\"text\":\"A prompt is usually treated as an instruction sent to an AI model: describe the task clearly enough, provide the necessary context, and the model should produce a better answer.\"},\"tunes\":{}},{\"id\":\"87AjwAq66J\",\"type\":\"paragraph\",\"data\":{\"text\":\"That model is incomplete.\"},\"tunes\":{}},{\"id\":\"tj7tVEuLdu\",\"type\":\"paragraph\",\"data\":{\"text\":\"The prompt does not merely describe the problem. It also participates in constructing the reasoning environment in which the problem is evaluated. Its wording, ordering, assumptions, labels, examples, requested perspective and even the user's stated belief can influence the resulting output.\"},\"tunes\":{}},{\"id\":\"-tRybRQp_N\",\"type\":\"quote\",\"data\":{\"text\":\"The prompt is not outside the reasoning process. It is one of the variables inside it.\",\"caption\":\"\",\"alignment\":\"left\"},\"tunes\":{}},{\"id\":\"kyV6Uccj-9\",\"type\":\"paragraph\",\"data\":{\"text\":\"This distinction is central to the methodology introduced in \u003Ca href=\\\"{{STAJIC_ARTICLE_1_URL}}\\\">Beyond Prompt Engineering: A Methodology for More Reliable AI Reasoning\u003C\u002Fa>. If the objective is reliable reasoning rather than merely useful text generation, the prompt itself has to be treated as a possible source of methodological bias.\"},\"tunes\":{}},{\"id\":\"OxIQaiiyre\",\"type\":\"header\",\"data\":{\"text\":\"Prompt Quality and Prompt Neutrality Are Different Problems\",\"level\":2},\"tunes\":{}},{\"id\":\"gPZ0TQf2lE\",\"type\":\"paragraph\",\"data\":{\"text\":\"Prompt engineering normally tries to improve clarity. Ambiguous requests are replaced with explicit requirements. Relevant context is added. Desired output structures are defined. Constraints are made visible.\"},\"tunes\":{}},{\"id\":\"QxePF-J2hL\",\"type\":\"paragraph\",\"data\":{\"text\":\"All of this can improve the quality of an answer.\"},\"tunes\":{}},{\"id\":\"qSCWsA2aE1\",\"type\":\"paragraph\",\"data\":{\"text\":\"But a more detailed prompt is not automatically a more neutral prompt. In fact, additional detail can introduce additional assumptions. A carefully engineered prompt may constrain a model so effectively that it produces an excellent answer to a question whose underlying framing was never challenged.\"},\"tunes\":{}},{\"id\":\"oyW3xLDkdZ\",\"type\":\"paragraph\",\"data\":{\"text\":\"These are therefore two different optimization targets:\"},\"tunes\":{}},{\"id\":\"SZErOWCjZD\",\"type\":\"list\",\"data\":{\"style\":\"unordered\",\"meta\":{},\"items\":[\"\u003Cb>Prompt quality:\u003C\u002Fb> does the model understand and execute the requested task well?\",\"\u003Cb>Epistemic robustness:\u003C\u002Fb> would the essential conclusion remain defensible if questionable assumptions or framing choices in the prompt were changed?\"]},\"tunes\":{}},{\"id\":\"J5Fq6NN2Uq\",\"type\":\"paragraph\",\"data\":{\"text\":\"A prompt can score highly on the first criterion while performing poorly on the second.\"},\"tunes\":{}},{\"id\":\"__ZGre4q3b\",\"type\":\"header\",\"data\":{\"text\":\"There Is No Perfectly Neutral Prompt\",\"level\":2},\"tunes\":{}},{\"id\":\"U9r731LwzJ\",\"type\":\"paragraph\",\"data\":{\"text\":\"Every non-trivial prompt makes choices.\"},\"tunes\":{}},{\"id\":\"3JWuhFaEv-\",\"type\":\"paragraph\",\"data\":{\"text\":\"It decides what information is included and what is omitted. It determines which concepts are named. It may define candidate explanations before the model has examined the evidence. It establishes terminology, scope and emphasis. In comparative tasks it may determine which option appears first. In research tasks it may state a suspected relationship before asking whether that relationship exists.\"},\"tunes\":{}},{\"id\":\"Ge8axqLq0j\",\"type\":\"paragraph\",\"data\":{\"text\":\"These decisions do not necessarily invalidate the answer. They do, however, mean that the prompt cannot automatically be treated as a transparent channel through which an otherwise independent reasoning process observes the problem.\"},\"tunes\":{}},{\"id\":\"yHP00kl5HW\",\"type\":\"paragraph\",\"data\":{\"text\":\"Brucks and Toubia describe a related phenomenon as \u003Ci>prompt architecture\u003C\u002Fi>. Across large-scale experiments involving GPT-3, GPT-4 and Llama 3.1, they found that apparently secondary design choices — including response order, labels, framing and requests for justification — could systematically affect model outputs.\"},\"tunes\":{}},{\"id\":\"ulRk5SB9Qd\",\"type\":\"paragraph\",\"data\":{\"text\":\"This is important because these effects were not limited to obviously manipulative prompts. Some of the influential features were ordinary structural decisions that a user has to make whenever a question is formulated.\"},\"tunes\":{}},{\"id\":\"797bcg7Bmu\",\"type\":\"quote\",\"data\":{\"text\":\"The absence of an intention to bias the model does not imply the absence of a framing effect.\",\"caption\":\"\",\"alignment\":\"left\"},\"tunes\":{}},{\"id\":\"bA2MGoPpic\",\"type\":\"header\",\"data\":{\"text\":\"A Concrete Example: Order and Labels\",\"level\":2},\"tunes\":{}},{\"id\":\"yCyCHzS3UX\",\"type\":\"paragraph\",\"data\":{\"text\":\"The prompt-architecture problem becomes easier to see in controlled comparison tasks.\"},\"tunes\":{}},{\"id\":\"ncGKpH3dAV\",\"type\":\"paragraph\",\"data\":{\"text\":\"In one of the experiments reported by Brucks and Toubia, GPT-4 was asked to compare sets of items under systematically varied prompt conditions. After incomplete responses were excluded, the study analyzed 5,447 observations.\"},\"tunes\":{}},{\"id\":\"hFzNbN2c_j\",\"type\":\"paragraph\",\"data\":{\"text\":\"Under conditions in which an unbiased responder should have selected either response position equally often, GPT-4 selected the first option in 63.21% of observations. It also selected the option labeled B over the option labeled C in 74.27% of observations.\"},\"tunes\":{}},{\"id\":\"5tYyhjLkO_\",\"type\":\"paragraph\",\"data\":{\"text\":\"Those numbers must not be generalized into claims such as 'GPT-4 always prefers the first option' or 'all language models prefer B'. They describe particular experimental conditions. Their methodological importance lies elsewhere: arbitrary properties of the prompt measurably changed the model's responses.\"},\"tunes\":{}},{\"id\":\"ldR4Fkgb-d\",\"type\":\"paragraph\",\"data\":{\"text\":\"The same study also found that simply informing the model that order and labels had been randomized did not reliably remove these effects. Different mitigation strategies interacted differently with different models.\"},\"tunes\":{}},{\"id\":\"cQ4ULtwq_-\",\"type\":\"paragraph\",\"data\":{\"text\":\"That observation has a direct consequence for serious AI-assisted analysis: telling a model to 'be unbiased' is not equivalent to controlling for bias.\"},\"tunes\":{}},{\"id\":\"Z9FeKBAigk\",\"type\":\"header\",\"data\":{\"text\":\"Framing Is More Than Word Choice\",\"level\":2},\"tunes\":{}},{\"id\":\"HhlPkBck1N\",\"type\":\"paragraph\",\"data\":{\"text\":\"The most consequential framing effects are often not individual words. They are embedded in the logical structure of the request.\"},\"tunes\":{}},{\"id\":\"m-V7XXr8cn\",\"type\":\"paragraph\",\"data\":{\"text\":\"Consider the difference between these questions:\"},\"tunes\":{}},{\"id\":\"io-cFNKoFB\",\"type\":\"quote\",\"data\":{\"text\":\"What evidence supports a historical transmission between tradition A and tradition B?\",\"caption\":\"\",\"alignment\":\"left\"},\"tunes\":{}},{\"id\":\"sxu-3HP0uN\",\"type\":\"paragraph\",\"data\":{\"text\":\"and:\"},\"tunes\":{}},{\"id\":\"oFS6gMZpYP\",\"type\":\"quote\",\"data\":{\"text\":\"What relationship, if any, between tradition A and tradition B is best supported by the available evidence?\",\"caption\":\"\",\"alignment\":\"left\"},\"tunes\":{}},{\"id\":\"O3bIvKKEk8\",\"type\":\"paragraph\",\"data\":{\"text\":\"The first formulation already assumes that a transmission hypothesis deserves the model's attention. It asks for evidence inside that hypothesis. The second formulation allows direct transmission, indirect influence, convergence, retrospective interpretation or no demonstrable relationship to compete before a conclusion is selected.\"},\"tunes\":{}},{\"id\":\"I7I7mGp19V\",\"type\":\"paragraph\",\"data\":{\"text\":\"The factual corpus could be identical. The search space is not.\"},\"tunes\":{}},{\"id\":\"enMp7kzoAd\",\"type\":\"paragraph\",\"data\":{\"text\":\"This distinction generalizes well beyond historical research.\"},\"tunes\":{}},{\"id\":\"nmiFSljZ4_\",\"type\":\"list\",\"data\":{\"style\":\"unordered\",\"meta\":{},\"items\":[\"\u003Cb>Debugging:\u003C\u002Fb> 'Why is nginx causing this timeout?' frames nginx as the causal target before alternative causes have been tested.\",\"\u003Cb>Architecture:\u003C\u002Fb> 'How should we implement this using microservices?' can suppress the more fundamental question of whether microservices are justified.\",\"\u003Cb>Product strategy:\u003C\u002Fb> 'How should we launch this product?' presupposes that launch is currently the correct action.\",\"\u003Cb>Security:\u003C\u002Fb> 'How did the attacker enter through the API?' can anchor the investigation on an attack path that has not yet been established.\",\"\u003Cb>Research:\u003C\u002Fb> 'What explains this correlation?' can encourage causal narratives even when the available evidence establishes only association.\"]},\"tunes\":{}},{\"id\":\"yZv1xnyfBe\",\"type\":\"paragraph\",\"data\":{\"text\":\"A sophisticated model may notice and challenge these assumptions. But the methodology cannot depend on the hope that it always will.\"},\"tunes\":{}},{\"id\":\"NfqFxMQvvZ\",\"type\":\"header\",\"data\":{\"text\":\"Instruction-Following Creates a Structural Tension\",\"level\":2},\"tunes\":{}},{\"id\":\"QR08HoXCCZ\",\"type\":\"paragraph\",\"data\":{\"text\":\"Modern AI assistants are intentionally trained to follow instructions. That capability is one of the reasons they are useful.\"},\"tunes\":{}},{\"id\":\"-yCXx1eXuI\",\"type\":\"paragraph\",\"data\":{\"text\":\"But instruction-following introduces an important tension in analytical work.\"},\"tunes\":{}},{\"id\":\"zz6DrMTowJ\",\"type\":\"quote\",\"data\":{\"text\":\"The model is simultaneously expected to respect the user's task definition and to reason independently about whether that definition contains questionable assumptions.\",\"caption\":\"\",\"alignment\":\"left\"},\"tunes\":{}},{\"id\":\"hiqeRkN84T\",\"type\":\"paragraph\",\"data\":{\"text\":\"These objectives are usually compatible. Sometimes they are not.\"},\"tunes\":{}},{\"id\":\"aBUjoCcv_e\",\"type\":\"paragraph\",\"data\":{\"text\":\"If a user says, 'Show me why architecture A is superior to architecture B,' a highly compliant response can produce a persuasive argument for A. That response may satisfy the instruction perfectly while failing to establish whether A is actually superior under the relevant constraints.\"},\"tunes\":{}},{\"id\":\"wyPYWxHJbI\",\"type\":\"paragraph\",\"data\":{\"text\":\"This is why methodological prompting must distinguish between \u003Ci>performing an assigned argumentative task\u003C\u002Fi> and \u003Ci>determining which conclusion is best supported\u003C\u002Fi>.\"},\"tunes\":{}},{\"id\":\"nar1tGfC9H\",\"type\":\"paragraph\",\"data\":{\"text\":\"The former is a legitimate AI capability. It should not be confused with the latter.\"},\"tunes\":{}},{\"id\":\"kvRpv0NcKP\",\"type\":\"header\",\"data\":{\"text\":\"Sycophancy Is Related — but It Is Not the Same Problem\",\"level\":2},\"tunes\":{}},{\"id\":\"Bh5ttKVJVM\",\"type\":\"paragraph\",\"data\":{\"text\":\"Prompt-induced bias is sometimes discussed together with \u003Ci>sycophancy\u003C\u002Fi>. The concepts overlap, but they should not be collapsed into one.\"},\"tunes\":{}},{\"id\":\"0zOOaeGfwY\",\"type\":\"paragraph\",\"data\":{\"text\":\"Framing effects describe changes in output caused by characteristics of how a task is presented. Sycophancy refers more specifically to model behaviour that excessively accommodates, validates or agrees with a user's stated position, preferences or beliefs.\"},\"tunes\":{}},{\"id\":\"wRO44dKbIo\",\"type\":\"paragraph\",\"data\":{\"text\":\"A framing effect can occur without any user opinion being present. Option order alone may affect a result. Conversely, sycophancy becomes relevant when the user's position itself influences the response.\"},\"tunes\":{}},{\"id\":\"9pGYhPBSmH\",\"type\":\"paragraph\",\"data\":{\"text\":\"Research by Sharma and colleagues found sycophantic behaviour across multiple AI assistants and reported that responses matching a user's expressed views could receive stronger human preference signals even when those responses were less truthful.\"},\"tunes\":{}},{\"id\":\"_QkNUvNB6B\",\"type\":\"paragraph\",\"data\":{\"text\":\"The issue is not merely theoretical. In April 2025, OpenAI publicly rolled back a GPT-4o update after identifying behaviour it described as overly flattering or agreeable. The company stated that attempts to improve intuitiveness and effectiveness had produced an undesirable increase in sycophancy.\"},\"tunes\":{}},{\"id\":\"VgG2yR9rhp\",\"type\":\"paragraph\",\"data\":{\"text\":\"A later joint alignment evaluation by Anthropic and OpenAI also reported familiar forms of sycophancy across evaluated models from both organizations.\"},\"tunes\":{}},{\"id\":\"4X6RHp5paC\",\"type\":\"paragraph\",\"data\":{\"text\":\"These findings should not be interpreted as evidence that AI assistants always agree with their users. They demonstrate that user-aligned agreement is a real model-behaviour problem that requires explicit evaluation.\"},\"tunes\":{}},{\"id\":\"_i3dlikecO\",\"type\":\"header\",\"data\":{\"text\":\"The User Can Become Part of the Hypothesis\",\"level\":2},\"tunes\":{}},{\"id\":\"reIIWKYTJd\",\"type\":\"paragraph\",\"data\":{\"text\":\"The problem becomes particularly subtle in long analytical conversations.\"},\"tunes\":{}},{\"id\":\"6J46FNq7qH\",\"type\":\"paragraph\",\"data\":{\"text\":\"Over multiple turns, the user may reveal what they expect to find, which explanations they consider plausible, which evidence they value and which conclusion they currently prefer. This context is useful. It can dramatically improve the relevance of the model's work.\"},\"tunes\":{}},{\"id\":\"ZcN5xXtbxW\",\"type\":\"paragraph\",\"data\":{\"text\":\"But it also means that the model is no longer evaluating only the external problem. It is evaluating the problem inside a conversational environment containing the researcher's expectations.\"},\"tunes\":{}},{\"id\":\"vK4JnApfZt\",\"type\":\"paragraph\",\"data\":{\"text\":\"In effect, the user's hypothesis can become part of the model's input distribution.\"},\"tunes\":{}},{\"id\":\"AOlGBbnTg0\",\"type\":\"paragraph\",\"data\":{\"text\":\"This does not make personalization undesirable. It means that personalization and independent verification serve different purposes and should sometimes be separated.\"},\"tunes\":{}},{\"id\":\"9XKME2NaVJ\",\"type\":\"header\",\"data\":{\"text\":\"The Problem Is Not Solved by Asking the Model to Disagree\",\"level\":2},\"tunes\":{}},{\"id\":\"O751DHnth_\",\"type\":\"paragraph\",\"data\":{\"text\":\"A natural response is to add instructions such as:\"},\"tunes\":{}},{\"id\":\"8JzmLnV7m7\",\"type\":\"quote\",\"data\":{\"text\":\"Do not agree with me. Be critical. Challenge my assumptions.\",\"caption\":\"\",\"alignment\":\"left\"},\"tunes\":{}},{\"id\":\"ZMPSyYPMYV\",\"type\":\"paragraph\",\"data\":{\"text\":\"These instructions can be useful, but they do not solve the methodological problem.\"},\"tunes\":{}},{\"id\":\"j76417bl_n\",\"type\":\"paragraph\",\"data\":{\"text\":\"Blind disagreement is no more epistemically valuable than blind agreement. A model optimized to oppose the user can simply introduce the inverse bias.\"},\"tunes\":{}},{\"id\":\"Nq7y0ropGz\",\"type\":\"paragraph\",\"data\":{\"text\":\"The target is not disagreement.\"},\"tunes\":{}},{\"id\":\"pHdGq6meUb\",\"type\":\"quote\",\"data\":{\"text\":\"The target is conclusion independence: the evidence should determine whether the user's hypothesis is strengthened, weakened, replaced or left unresolved.\",\"caption\":\"\",\"alignment\":\"left\"},\"tunes\":{}},{\"id\":\"CthU7-jRrf\",\"type\":\"paragraph\",\"data\":{\"text\":\"That requires a process in which alternative explanations are generated because they are plausible, not because the model has been instructed to be oppositional.\"},\"tunes\":{}},{\"id\":\"jdWWRGkdwg\",\"type\":\"header\",\"data\":{\"text\":\"Prompt Bias and Confirmation Bias Can Reinforce Each Other\",\"level\":2},\"tunes\":{}},{\"id\":\"gAbdE5Z0ti\",\"type\":\"paragraph\",\"data\":{\"text\":\"Prompt framing becomes more consequential when it interacts with another reasoning problem: confirmation bias.\"},\"tunes\":{}},{\"id\":\"ZmvPO2ZeSg\",\"type\":\"paragraph\",\"data\":{\"text\":\"A 2026 study by Jhaveri and colleagues tested eleven language models using interactive rule-discovery tasks. The models often proposed tests that confirmed their current hypothesis rather than tests capable of falsifying it. Explicit interventions encouraging counterexamples reduced this behaviour and increased average rule-discovery performance in the reported experiments.\"},\"tunes\":{}},{\"id\":\"GW9c1KWlbT\",\"type\":\"paragraph\",\"data\":{\"text\":\"This creates a potentially self-reinforcing loop:\"},\"tunes\":{}},{\"id\":\"EotLF5pQFD\",\"type\":\"quote\",\"data\":{\"text\":\"User framing → initial model hypothesis → confirmatory evidence search → stronger confidence in the framed hypothesis\",\"caption\":\"\",\"alignment\":\"left\"},\"tunes\":{}},{\"id\":\"hJD6GhYkkj\",\"type\":\"paragraph\",\"data\":{\"text\":\"Importantly, each individual step can appear reasonable. The weakness lies in the structure of the process.\"},\"tunes\":{}},{\"id\":\"KKeYrjvXxv\",\"type\":\"paragraph\",\"data\":{\"text\":\"For that reason, the next methodological layer cannot merely ask the model to produce more evidence. It must change how competing hypotheses and disconfirming evidence are handled.\"},\"tunes\":{}},{\"id\":\"MXzqcS07oV\",\"type\":\"paragraph\",\"data\":{\"text\":\"That mechanism is developed separately in \u003Ca href=\\\"{{STAJIC_ARTICLE_4_URL}}\\\">Falsification for AI Reasoning: From Answers to Tested Hypotheses\u003C\u002Fa>.\"},\"tunes\":{}},{\"id\":\"-mh4QlOvWc\",\"type\":\"header\",\"data\":{\"text\":\"Why More Reasoning Does Not Automatically Remove the Problem\",\"level\":2},\"tunes\":{}},{\"id\":\"fa2B3ieTJn\",\"type\":\"paragraph\",\"data\":{\"text\":\"It is tempting to assume that stronger reasoning models eventually eliminate prompt dependency by reasoning through it.\"},\"tunes\":{}},{\"id\":\"4t-oeMW07n\",\"type\":\"paragraph\",\"data\":{\"text\":\"That assumption is unsafe.\"},\"tunes\":{}},{\"id\":\"Mx_xlkDbJ3\",\"type\":\"paragraph\",\"data\":{\"text\":\"Additional reasoning capacity can help a model identify assumptions, compare explanations and detect contradictions. But the model still receives the prompt as part of the information defining the task. More capable reasoning does not make task formulation disappear.\"},\"tunes\":{}},{\"id\":\"OhQJ-E5rw9\",\"type\":\"paragraph\",\"data\":{\"text\":\"A stronger model can therefore produce a more sophisticated analysis while remaining influenced by the original framing.\"},\"tunes\":{}},{\"id\":\"rYgQbur-Lq\",\"type\":\"quote\",\"data\":{\"text\":\"Reasoning depth and framing independence are different properties.\",\"caption\":\"\",\"alignment\":\"left\"},\"tunes\":{}},{\"id\":\"rpD412gIMG\",\"type\":\"paragraph\",\"data\":{\"text\":\"This is one reason external methodological structure remains valuable even when the underlying model becomes more capable.\"},\"tunes\":{}},{\"id\":\"QX8MOYz-Or\",\"type\":\"header\",\"data\":{\"text\":\"The Prompt Should Become a Testable Variable\",\"level\":2},\"tunes\":{}},{\"id\":\"IBOYn2SHBk\",\"type\":\"paragraph\",\"data\":{\"text\":\"Once the prompt is recognized as part of the reasoning environment, an important methodological shift becomes possible.\"},\"tunes\":{}},{\"id\":\"ChEdMApOX6\",\"type\":\"paragraph\",\"data\":{\"text\":\"Instead of treating one carefully written prompt as the definitive representation of the problem, we can treat the formulation itself as a variable.\"},\"tunes\":{}},{\"id\":\"t-UQrOUVDX\",\"type\":\"paragraph\",\"data\":{\"text\":\"For low-consequence tasks, this is unnecessary. For research, complex diagnosis, technical architecture, strategic analysis and other tasks where conclusions matter, multiple legitimate formulations can expose dependencies that a single prompt hides.\"},\"tunes\":{}},{\"id\":\"J06ArUwzAW\",\"type\":\"paragraph\",\"data\":{\"text\":\"This does not mean generating random paraphrases. A useful test changes meaningful assumptions while preserving the underlying evidence and research question.\"},\"tunes\":{}},{\"id\":\"DjtQGtT69O\",\"type\":\"paragraph\",\"data\":{\"text\":\"For example:\"},\"tunes\":{}},{\"id\":\"gS2dM2Crpk\",\"type\":\"list\",\"data\":{\"style\":\"unordered\",\"meta\":{},\"items\":[\"remove the user's preferred explanation from the formulation;\",\"reverse the initial hypothesis;\",\"present competing hypotheses symmetrically;\",\"change option order where order should be irrelevant;\",\"separate evidence collection from interpretation;\",\"ask a separate pass to construct the strongest evidence-based objection.\"]},\"tunes\":{}},{\"id\":\"eMMKqelkmi\",\"type\":\"paragraph\",\"data\":{\"text\":\"The objective is to observe what changes — and, more importantly, what does not.\"},\"tunes\":{}},{\"id\":\"fW1HzQt4Sp\",\"type\":\"paragraph\",\"data\":{\"text\":\"This leads directly to the next article in the series: \u003Ca href=\\\"{{STAJIC_ARTICLE_3_URL}}\\\">Prompt Invariance: Does the Conclusion Survive the Prompt?\u003C\u002Fa>.\"},\"tunes\":{}},{\"id\":\"9OBUF1Xm5g\",\"type\":\"header\",\"data\":{\"text\":\"From Better Prompts to Better Methodology\",\"level\":2},\"tunes\":{}},{\"id\":\"bXko8eBkcP\",\"type\":\"paragraph\",\"data\":{\"text\":\"Prompt engineering remains useful. Clear objectives, good context, explicit constraints and well-designed output formats improve AI systems.\"},\"tunes\":{}},{\"id\":\"1jyf_r-cS9\",\"type\":\"paragraph\",\"data\":{\"text\":\"The mistake is assuming that prompt optimization and reasoning validation are the same discipline.\"},\"tunes\":{}},{\"id\":\"gIW8IiuLnQ\",\"type\":\"paragraph\",\"data\":{\"text\":\"They solve different problems.\"},\"tunes\":{}},{\"id\":\"mP0h2g7l7B\",\"type\":\"quote\",\"data\":{\"text\":\"Prompt engineering asks: How should I formulate the task so the model performs it well? Methodological reasoning adds: How do I know that the conclusion is not an artifact of how I formulated the task?\",\"caption\":\"\",\"alignment\":\"left\"},\"tunes\":{}},{\"id\":\"GdVydSO2tI\",\"type\":\"paragraph\",\"data\":{\"text\":\"The second question is what transforms prompting from an interface technique into part of an epistemic methodology.\"},\"tunes\":{}},{\"id\":\"c8HvZSabul\",\"type\":\"paragraph\",\"data\":{\"text\":\"The prompt still matters. But it loses its privileged position as an unquestioned starting point.\"},\"tunes\":{}},{\"id\":\"lt-Lq40ey3\",\"type\":\"paragraph\",\"data\":{\"text\":\"It becomes something that can itself be inspected, varied and tested.\"},\"tunes\":{}},{\"id\":\"1u3woVm3yw\",\"type\":\"delimiter\",\"data\":{},\"tunes\":{}},{\"id\":\"3wG0-FWb1b\",\"type\":\"header\",\"data\":{\"text\":\"Research Context\",\"level\":2},\"tunes\":{}},{\"id\":\"LIa7a-EKIU\",\"type\":\"paragraph\",\"data\":{\"text\":\"Several independent research directions motivate the distinction made in this article. Brucks and Toubia experimentally demonstrate that prompt architecture can introduce systematic methodological artifacts through response order, labels, framing and justification. Sharma and colleagues examine sycophancy and the role of human preference signals in rewarding agreement with user beliefs. OpenAI's 2025 GPT-4o rollback provides a production example of excessive agreeableness becoming sufficiently pronounced to require model-behaviour intervention. Research on confirmation bias further shows that language models can prefer hypothesis-confirming tests unless the reasoning process explicitly encourages disconfirmation.\"},\"tunes\":{}},{\"id\":\"eIcLYuk4jH\",\"type\":\"paragraph\",\"data\":{\"text\":\"These findings address different phenomena and should not be treated as interchangeable evidence for one universal bias mechanism. Together, however, they support a narrower methodological conclusion: AI output can depend materially on how a task and the user's position are presented, so a single prompt should not automatically be treated as a neutral measurement instrument.\"},\"tunes\":{}},{\"id\":\"KX0zVJff-k\",\"type\":\"header\",\"data\":{\"text\":\"Selected References\",\"level\":3},\"tunes\":{}},{\"id\":\"ZGFmK5xgH1\",\"type\":\"list\",\"data\":{\"style\":\"unordered\",\"meta\":{},\"items\":[\"Brucks, M. & Toubia, O. — \u003Ci>Prompt architecture induces methodological artifacts in large language models\u003C\u002Fi>. PLOS ONE 20(4), e0319159, 2025.\",\"Sharma, M. et al. — \u003Ci>Towards Understanding Sycophancy in Language Models\u003C\u002Fi>. arXiv:2310.13548.\",\"OpenAI — \u003Ci>Sycophancy in GPT-4o: What Happened and What We're Doing About It\u003C\u002Fi>, April 2025.\",\"Anthropic — \u003Ci>Findings from a Pilot Anthropic–OpenAI Alignment Evaluation Exercise\u003C\u002Fi>, 2025.\",\"Jhaveri, A. R., GX-Chen, A., Sucholutsky, I. & Choi, E. — \u003Ci>Failing to Falsify: Evaluating and Mitigating Confirmation Bias in Language Models\u003C\u002Fi>. arXiv:2604.02485, 2026.\"]},\"tunes\":{}},{\"id\":\"QmiYrvS4qE\",\"type\":\"delimiter\",\"data\":{},\"tunes\":{}},{\"id\":\"c-VCu8XaDf\",\"type\":\"header\",\"data\":{\"text\":\"Continue the Series\",\"level\":2},\"tunes\":{}},{\"id\":\"VRXV5PGw8O\",\"type\":\"list\",\"data\":{\"style\":\"unordered\",\"meta\":{},\"items\":[\"\u003Cb>\u003Ca href=\\\"https:\u002F\u002Fstajic.de\u002Fblog\u002Fbeyond-prompt-engineering-a-methodology-for-more-reliable-ai-reasoning\\\">Beyond Prompt Engineering: A Methodology for More Reliable AI Reasoning\u003C\u002Fa>\u003C\u002Fb> — the methodological framework behind this series.\",\"\u003Cb>\u003Ca href=\\\"https:\u002F\u002Fstajic.de\u002Fblog\u002Fprompt-invariance-does-the-conclusion-survive-the-prompt\\\">Prompt Invariance: Does the Conclusion Survive the Prompt?\u003C\u002Fa>\u003C\u002Fb> — testing whether conclusions remain stable under blind, inverted and adversarial formulations.\",\"\u003Cb>\u003Ca href=\\\"https:\u002F\u002Fstajic.de\u002Fblog\u002Ffalsification-for-ai-reasoning-from-answers-to-tested-hypotheses\\\">Falsification for AI Reasoning: From Answers to Tested Hypotheses\u003C\u002Fa>\u003C\u002Fb> — replacing confirmatory reasoning with competing hypotheses and disconfirming tests.\",\"\u003Cb>\u003Ca href=\\\"https:\u002F\u002Fstajic.de\u002Fblog\u002Ffrom-research-protocol-to-a-general-ai-reasoning-framework\\\">From Research Protocol to General AI Reasoning Framework\u003C\u002Fa>\u003C\u002Fb> — generalizing the methodology across research, software, debugging and strategic analysis.\"]},\"tunes\":{}}],\"version\":\"2.31.6\"}",{"time":816,"blocks":817,"version":1270},1789805016477,[818,822,826,830,834,838,842,846,850,854,858,865,869,873,877,881,885,889,893,897,901,905,909,913,917,921,925,929,933,937,941,945,949,953,957,961,971,975,979,983,987,991,995,999,1003,1007,1011,1015,1019,1023,1027,1031,1035,1039,1043,1047,1051,1055,1059,1063,1067,1071,1075,1079,1083,1087,1091,1095,1099,1103,1107,1111,1115,1119,1123,1127,1131,1135,1139,1143,1147,1151,1155,1159,1163,1167,1171,1175,1179,1190,1194,1198,1202,1206,1210,1214,1218,1222,1226,1230,1233,1237,1241,1245,1249,1254,1257,1261],{"id":215,"data":819,"type":218,"tunes":821},{"text":820},"A prompt is usually treated as an instruction sent to an AI model: describe the task clearly enough, provide the necessary context, and the model should produce a better answer.",{},{"id":221,"data":823,"type":218,"tunes":825},{"text":824},"That model is incomplete.",{},{"id":226,"data":827,"type":218,"tunes":829},{"text":828},"The prompt does not merely describe the problem. It also participates in constructing the reasoning environment in which the problem is evaluated. Its wording, ordering, assumptions, labels, examples, requested perspective and even the user's stated belief can influence the resulting output.",{},{"id":231,"data":831,"type":236,"tunes":833},{"text":832,"caption":234,"alignment":235},"The prompt is not outside the reasoning process. It is one of the variables inside it.",{},{"id":239,"data":835,"type":218,"tunes":837},{"text":836},"This distinction is central to the methodology introduced in \u003Ca href=\"{{STAJIC_ARTICLE_1_URL}}\">Beyond Prompt Engineering: A Methodology for More Reliable AI Reasoning\u003C\u002Fa>. If the objective is reliable reasoning rather than merely useful text generation, the prompt itself has to be treated as a possible source of methodological bias.",{},{"id":244,"data":839,"type":42,"tunes":841},{"text":840,"level":247},"Prompt Quality and Prompt Neutrality Are Different Problems",{},{"id":250,"data":843,"type":218,"tunes":845},{"text":844},"Prompt engineering normally tries to improve clarity. Ambiguous requests are replaced with explicit requirements. Relevant context is added. Desired output structures are defined. Constraints are made visible.",{},{"id":255,"data":847,"type":218,"tunes":849},{"text":848},"All of this can improve the quality of an answer.",{},{"id":260,"data":851,"type":218,"tunes":853},{"text":852},"But a more detailed prompt is not automatically a more neutral prompt. In fact, additional detail can introduce additional assumptions. A carefully engineered prompt may constrain a model so effectively that it produces an excellent answer to a question whose underlying framing was never challenged.",{},{"id":265,"data":855,"type":218,"tunes":857},{"text":856},"These are therefore two different optimization targets:",{},{"id":270,"data":859,"type":277,"tunes":864},{"meta":860,"items":861,"style":276},{},[862,863],"\u003Cb>Prompt quality:\u003C\u002Fb> does the model understand and execute the requested task well?","\u003Cb>Epistemic robustness:\u003C\u002Fb> would the essential conclusion remain defensible if questionable assumptions or framing choices in the prompt were changed?",{},{"id":280,"data":866,"type":218,"tunes":868},{"text":867},"A prompt can score highly on the first criterion while performing poorly on the second.",{},{"id":285,"data":870,"type":42,"tunes":872},{"text":871,"level":247},"There Is No Perfectly Neutral Prompt",{},{"id":290,"data":874,"type":218,"tunes":876},{"text":875},"Every non-trivial prompt makes choices.",{},{"id":295,"data":878,"type":218,"tunes":880},{"text":879},"It decides what information is included and what is omitted. It determines which concepts are named. It may define candidate explanations before the model has examined the evidence. It establishes terminology, scope and emphasis. In comparative tasks it may determine which option appears first. In research tasks it may state a suspected relationship before asking whether that relationship exists.",{},{"id":300,"data":882,"type":218,"tunes":884},{"text":883},"These decisions do not necessarily invalidate the answer. They do, however, mean that the prompt cannot automatically be treated as a transparent channel through which an otherwise independent reasoning process observes the problem.",{},{"id":305,"data":886,"type":218,"tunes":888},{"text":887},"Brucks and Toubia describe a related phenomenon as \u003Ci>prompt architecture\u003C\u002Fi>. Across large-scale experiments involving GPT-3, GPT-4 and Llama 3.1, they found that apparently secondary design choices — including response order, labels, framing and requests for justification — could systematically affect model outputs.",{},{"id":310,"data":890,"type":218,"tunes":892},{"text":891},"This is important because these effects were not limited to obviously manipulative prompts. Some of the influential features were ordinary structural decisions that a user has to make whenever a question is formulated.",{},{"id":315,"data":894,"type":236,"tunes":896},{"text":895,"caption":234,"alignment":235},"The absence of an intention to bias the model does not imply the absence of a framing effect.",{},{"id":320,"data":898,"type":42,"tunes":900},{"text":899,"level":247},"A Concrete Example: Order and Labels",{},{"id":325,"data":902,"type":218,"tunes":904},{"text":903},"The prompt-architecture problem becomes easier to see in controlled comparison tasks.",{},{"id":330,"data":906,"type":218,"tunes":908},{"text":907},"In one of the experiments reported by Brucks and Toubia, GPT-4 was asked to compare sets of items under systematically varied prompt conditions. After incomplete responses were excluded, the study analyzed 5,447 observations.",{},{"id":335,"data":910,"type":218,"tunes":912},{"text":911},"Under conditions in which an unbiased responder should have selected either response position equally often, GPT-4 selected the first option in 63.21% of observations. It also selected the option labeled B over the option labeled C in 74.27% of observations.",{},{"id":340,"data":914,"type":218,"tunes":916},{"text":915},"Those numbers must not be generalized into claims such as 'GPT-4 always prefers the first option' or 'all language models prefer B'. They describe particular experimental conditions. Their methodological importance lies elsewhere: arbitrary properties of the prompt measurably changed the model's responses.",{},{"id":345,"data":918,"type":218,"tunes":920},{"text":919},"The same study also found that simply informing the model that order and labels had been randomized did not reliably remove these effects. Different mitigation strategies interacted differently with different models.",{},{"id":350,"data":922,"type":218,"tunes":924},{"text":923},"That observation has a direct consequence for serious AI-assisted analysis: telling a model to 'be unbiased' is not equivalent to controlling for bias.",{},{"id":355,"data":926,"type":42,"tunes":928},{"text":927,"level":247},"Framing Is More Than Word Choice",{},{"id":360,"data":930,"type":218,"tunes":932},{"text":931},"The most consequential framing effects are often not individual words. They are embedded in the logical structure of the request.",{},{"id":365,"data":934,"type":218,"tunes":936},{"text":935},"Consider the difference between these questions:",{},{"id":370,"data":938,"type":236,"tunes":940},{"text":939,"caption":234,"alignment":235},"What evidence supports a historical transmission between tradition A and tradition B?",{},{"id":375,"data":942,"type":218,"tunes":944},{"text":943},"and:",{},{"id":380,"data":946,"type":236,"tunes":948},{"text":947,"caption":234,"alignment":235},"What relationship, if any, between tradition A and tradition B is best supported by the available evidence?",{},{"id":385,"data":950,"type":218,"tunes":952},{"text":951},"The first formulation already assumes that a transmission hypothesis deserves the model's attention. It asks for evidence inside that hypothesis. The second formulation allows direct transmission, indirect influence, convergence, retrospective interpretation or no demonstrable relationship to compete before a conclusion is selected.",{},{"id":390,"data":954,"type":218,"tunes":956},{"text":955},"The factual corpus could be identical. The search space is not.",{},{"id":395,"data":958,"type":218,"tunes":960},{"text":959},"This distinction generalizes well beyond historical research.",{},{"id":400,"data":962,"type":277,"tunes":970},{"meta":963,"items":964,"style":276},{},[965,966,967,968,969],"\u003Cb>Debugging:\u003C\u002Fb> 'Why is nginx causing this timeout?' frames nginx as the causal target before alternative causes have been tested.","\u003Cb>Architecture:\u003C\u002Fb> 'How should we implement this using microservices?' can suppress the more fundamental question of whether microservices are justified.","\u003Cb>Product strategy:\u003C\u002Fb> 'How should we launch this product?' presupposes that launch is currently the correct action.","\u003Cb>Security:\u003C\u002Fb> 'How did the attacker enter through the API?' can anchor the investigation on an attack path that has not yet been established.","\u003Cb>Research:\u003C\u002Fb> 'What explains this correlation?' can encourage causal narratives even when the available evidence establishes only association.",{},{"id":411,"data":972,"type":218,"tunes":974},{"text":973},"A sophisticated model may notice and challenge these assumptions. But the methodology cannot depend on the hope that it always will.",{},{"id":416,"data":976,"type":42,"tunes":978},{"text":977,"level":247},"Instruction-Following Creates a Structural Tension",{},{"id":421,"data":980,"type":218,"tunes":982},{"text":981},"Modern AI assistants are intentionally trained to follow instructions. That capability is one of the reasons they are useful.",{},{"id":426,"data":984,"type":218,"tunes":986},{"text":985},"But instruction-following introduces an important tension in analytical work.",{},{"id":431,"data":988,"type":236,"tunes":990},{"text":989,"caption":234,"alignment":235},"The model is simultaneously expected to respect the user's task definition and to reason independently about whether that definition contains questionable assumptions.",{},{"id":436,"data":992,"type":218,"tunes":994},{"text":993},"These objectives are usually compatible. Sometimes they are not.",{},{"id":441,"data":996,"type":218,"tunes":998},{"text":997},"If a user says, 'Show me why architecture A is superior to architecture B,' a highly compliant response can produce a persuasive argument for A. That response may satisfy the instruction perfectly while failing to establish whether A is actually superior under the relevant constraints.",{},{"id":446,"data":1000,"type":218,"tunes":1002},{"text":1001},"This is why methodological prompting must distinguish between \u003Ci>performing an assigned argumentative task\u003C\u002Fi> and \u003Ci>determining which conclusion is best supported\u003C\u002Fi>.",{},{"id":451,"data":1004,"type":218,"tunes":1006},{"text":1005},"The former is a legitimate AI capability. It should not be confused with the latter.",{},{"id":456,"data":1008,"type":42,"tunes":1010},{"text":1009,"level":247},"Sycophancy Is Related — but It Is Not the Same Problem",{},{"id":461,"data":1012,"type":218,"tunes":1014},{"text":1013},"Prompt-induced bias is sometimes discussed together with \u003Ci>sycophancy\u003C\u002Fi>. The concepts overlap, but they should not be collapsed into one.",{},{"id":466,"data":1016,"type":218,"tunes":1018},{"text":1017},"Framing effects describe changes in output caused by characteristics of how a task is presented. Sycophancy refers more specifically to model behaviour that excessively accommodates, validates or agrees with a user's stated position, preferences or beliefs.",{},{"id":471,"data":1020,"type":218,"tunes":1022},{"text":1021},"A framing effect can occur without any user opinion being present. Option order alone may affect a result. Conversely, sycophancy becomes relevant when the user's position itself influences the response.",{},{"id":476,"data":1024,"type":218,"tunes":1026},{"text":1025},"Research by Sharma and colleagues found sycophantic behaviour across multiple AI assistants and reported that responses matching a user's expressed views could receive stronger human preference signals even when those responses were less truthful.",{},{"id":481,"data":1028,"type":218,"tunes":1030},{"text":1029},"The issue is not merely theoretical. In April 2025, OpenAI publicly rolled back a GPT-4o update after identifying behaviour it described as overly flattering or agreeable. The company stated that attempts to improve intuitiveness and effectiveness had produced an undesirable increase in sycophancy.",{},{"id":486,"data":1032,"type":218,"tunes":1034},{"text":1033},"A later joint alignment evaluation by Anthropic and OpenAI also reported familiar forms of sycophancy across evaluated models from both organizations.",{},{"id":491,"data":1036,"type":218,"tunes":1038},{"text":1037},"These findings should not be interpreted as evidence that AI assistants always agree with their users. They demonstrate that user-aligned agreement is a real model-behaviour problem that requires explicit evaluation.",{},{"id":496,"data":1040,"type":42,"tunes":1042},{"text":1041,"level":247},"The User Can Become Part of the Hypothesis",{},{"id":501,"data":1044,"type":218,"tunes":1046},{"text":1045},"The problem becomes particularly subtle in long analytical conversations.",{},{"id":506,"data":1048,"type":218,"tunes":1050},{"text":1049},"Over multiple turns, the user may reveal what they expect to find, which explanations they consider plausible, which evidence they value and which conclusion they currently prefer. This context is useful. It can dramatically improve the relevance of the model's work.",{},{"id":511,"data":1052,"type":218,"tunes":1054},{"text":1053},"But it also means that the model is no longer evaluating only the external problem. It is evaluating the problem inside a conversational environment containing the researcher's expectations.",{},{"id":516,"data":1056,"type":218,"tunes":1058},{"text":1057},"In effect, the user's hypothesis can become part of the model's input distribution.",{},{"id":521,"data":1060,"type":218,"tunes":1062},{"text":1061},"This does not make personalization undesirable. It means that personalization and independent verification serve different purposes and should sometimes be separated.",{},{"id":526,"data":1064,"type":42,"tunes":1066},{"text":1065,"level":247},"The Problem Is Not Solved by Asking the Model to Disagree",{},{"id":531,"data":1068,"type":218,"tunes":1070},{"text":1069},"A natural response is to add instructions such as:",{},{"id":536,"data":1072,"type":236,"tunes":1074},{"text":1073,"caption":234,"alignment":235},"Do not agree with me. Be critical. Challenge my assumptions.",{},{"id":541,"data":1076,"type":218,"tunes":1078},{"text":1077},"These instructions can be useful, but they do not solve the methodological problem.",{},{"id":546,"data":1080,"type":218,"tunes":1082},{"text":1081},"Blind disagreement is no more epistemically valuable than blind agreement. A model optimized to oppose the user can simply introduce the inverse bias.",{},{"id":551,"data":1084,"type":218,"tunes":1086},{"text":1085},"The target is not disagreement.",{},{"id":556,"data":1088,"type":236,"tunes":1090},{"text":1089,"caption":234,"alignment":235},"The target is conclusion independence: the evidence should determine whether the user's hypothesis is strengthened, weakened, replaced or left unresolved.",{},{"id":561,"data":1092,"type":218,"tunes":1094},{"text":1093},"That requires a process in which alternative explanations are generated because they are plausible, not because the model has been instructed to be oppositional.",{},{"id":566,"data":1096,"type":42,"tunes":1098},{"text":1097,"level":247},"Prompt Bias and Confirmation Bias Can Reinforce Each Other",{},{"id":571,"data":1100,"type":218,"tunes":1102},{"text":1101},"Prompt framing becomes more consequential when it interacts with another reasoning problem: confirmation bias.",{},{"id":576,"data":1104,"type":218,"tunes":1106},{"text":1105},"A 2026 study by Jhaveri and colleagues tested eleven language models using interactive rule-discovery tasks. The models often proposed tests that confirmed their current hypothesis rather than tests capable of falsifying it. Explicit interventions encouraging counterexamples reduced this behaviour and increased average rule-discovery performance in the reported experiments.",{},{"id":581,"data":1108,"type":218,"tunes":1110},{"text":1109},"This creates a potentially self-reinforcing loop:",{},{"id":586,"data":1112,"type":236,"tunes":1114},{"text":1113,"caption":234,"alignment":235},"User framing → initial model hypothesis → confirmatory evidence search → stronger confidence in the framed hypothesis",{},{"id":591,"data":1116,"type":218,"tunes":1118},{"text":1117},"Importantly, each individual step can appear reasonable. The weakness lies in the structure of the process.",{},{"id":596,"data":1120,"type":218,"tunes":1122},{"text":1121},"For that reason, the next methodological layer cannot merely ask the model to produce more evidence. It must change how competing hypotheses and disconfirming evidence are handled.",{},{"id":601,"data":1124,"type":218,"tunes":1126},{"text":1125},"That mechanism is developed separately in \u003Ca href=\"{{STAJIC_ARTICLE_4_URL}}\">Falsification for AI Reasoning: From Answers to Tested Hypotheses\u003C\u002Fa>.",{},{"id":606,"data":1128,"type":42,"tunes":1130},{"text":1129,"level":247},"Why More Reasoning Does Not Automatically Remove the Problem",{},{"id":611,"data":1132,"type":218,"tunes":1134},{"text":1133},"It is tempting to assume that stronger reasoning models eventually eliminate prompt dependency by reasoning through it.",{},{"id":616,"data":1136,"type":218,"tunes":1138},{"text":1137},"That assumption is unsafe.",{},{"id":621,"data":1140,"type":218,"tunes":1142},{"text":1141},"Additional reasoning capacity can help a model identify assumptions, compare explanations and detect contradictions. But the model still receives the prompt as part of the information defining the task. More capable reasoning does not make task formulation disappear.",{},{"id":626,"data":1144,"type":218,"tunes":1146},{"text":1145},"A stronger model can therefore produce a more sophisticated analysis while remaining influenced by the original framing.",{},{"id":631,"data":1148,"type":236,"tunes":1150},{"text":1149,"caption":234,"alignment":235},"Reasoning depth and framing independence are different properties.",{},{"id":636,"data":1152,"type":218,"tunes":1154},{"text":1153},"This is one reason external methodological structure remains valuable even when the underlying model becomes more capable.",{},{"id":641,"data":1156,"type":42,"tunes":1158},{"text":1157,"level":247},"The Prompt Should Become a Testable Variable",{},{"id":646,"data":1160,"type":218,"tunes":1162},{"text":1161},"Once the prompt is recognized as part of the reasoning environment, an important methodological shift becomes possible.",{},{"id":651,"data":1164,"type":218,"tunes":1166},{"text":1165},"Instead of treating one carefully written prompt as the definitive representation of the problem, we can treat the formulation itself as a variable.",{},{"id":656,"data":1168,"type":218,"tunes":1170},{"text":1169},"For low-consequence tasks, this is unnecessary. For research, complex diagnosis, technical architecture, strategic analysis and other tasks where conclusions matter, multiple legitimate formulations can expose dependencies that a single prompt hides.",{},{"id":661,"data":1172,"type":218,"tunes":1174},{"text":1173},"This does not mean generating random paraphrases. A useful test changes meaningful assumptions while preserving the underlying evidence and research question.",{},{"id":666,"data":1176,"type":218,"tunes":1178},{"text":1177},"For example:",{},{"id":671,"data":1180,"type":277,"tunes":1189},{"meta":1181,"items":1182,"style":276},{},[1183,1184,1185,1186,1187,1188],"remove the user's preferred explanation from the formulation;","reverse the initial hypothesis;","present competing hypotheses symmetrically;","change option order where order should be irrelevant;","separate evidence collection from interpretation;","ask a separate pass to construct the strongest evidence-based objection.",{},{"id":683,"data":1191,"type":218,"tunes":1193},{"text":1192},"The objective is to observe what changes — and, more importantly, what does not.",{},{"id":688,"data":1195,"type":218,"tunes":1197},{"text":1196},"This leads directly to the next article in the series: \u003Ca href=\"{{STAJIC_ARTICLE_3_URL}}\">Prompt Invariance: Does the Conclusion Survive the Prompt?\u003C\u002Fa>.",{},{"id":693,"data":1199,"type":42,"tunes":1201},{"text":1200,"level":247},"From Better Prompts to Better Methodology",{},{"id":698,"data":1203,"type":218,"tunes":1205},{"text":1204},"Prompt engineering remains useful. Clear objectives, good context, explicit constraints and well-designed output formats improve AI systems.",{},{"id":703,"data":1207,"type":218,"tunes":1209},{"text":1208},"The mistake is assuming that prompt optimization and reasoning validation are the same discipline.",{},{"id":708,"data":1211,"type":218,"tunes":1213},{"text":1212},"They solve different problems.",{},{"id":713,"data":1215,"type":236,"tunes":1217},{"text":1216,"caption":234,"alignment":235},"Prompt engineering asks: How should I formulate the task so the model performs it well? Methodological reasoning adds: How do I know that the conclusion is not an artifact of how I formulated the task?",{},{"id":718,"data":1219,"type":218,"tunes":1221},{"text":1220},"The second question is what transforms prompting from an interface technique into part of an epistemic methodology.",{},{"id":723,"data":1223,"type":218,"tunes":1225},{"text":1224},"The prompt still matters. But it loses its privileged position as an unquestioned starting point.",{},{"id":728,"data":1227,"type":218,"tunes":1229},{"text":1228},"It becomes something that can itself be inspected, varied and tested.",{},{"id":733,"data":1231,"type":735,"tunes":1232},{},{},{"id":738,"data":1234,"type":42,"tunes":1236},{"text":1235,"level":247},"Research Context",{},{"id":743,"data":1238,"type":218,"tunes":1240},{"text":1239},"Several independent research directions motivate the distinction made in this article. Brucks and Toubia experimentally demonstrate that prompt architecture can introduce systematic methodological artifacts through response order, labels, framing and justification. Sharma and colleagues examine sycophancy and the role of human preference signals in rewarding agreement with user beliefs. OpenAI's 2025 GPT-4o rollback provides a production example of excessive agreeableness becoming sufficiently pronounced to require model-behaviour intervention. Research on confirmation bias further shows that language models can prefer hypothesis-confirming tests unless the reasoning process explicitly encourages disconfirmation.",{},{"id":748,"data":1242,"type":218,"tunes":1244},{"text":1243},"These findings address different phenomena and should not be treated as interchangeable evidence for one universal bias mechanism. Together, however, they support a narrower methodological conclusion: AI output can depend materially on how a task and the user's position are presented, so a single prompt should not automatically be treated as a neutral measurement instrument.",{},{"id":753,"data":1246,"type":42,"tunes":1248},{"text":1247,"level":756},"Selected References",{},{"id":759,"data":1250,"type":277,"tunes":1253},{"meta":1251,"items":1252,"style":276},{},[763,764,765,766,767],{},{"id":770,"data":1255,"type":735,"tunes":1256},{},{},{"id":774,"data":1258,"type":42,"tunes":1260},{"text":1259,"level":247},"Continue the Series",{},{"id":779,"data":1262,"type":277,"tunes":1269},{"meta":1263,"items":1264,"style":276},{},[1265,1266,1267,1268],"\u003Cb>\u003Ca href=\"https:\u002F\u002Fstajic.de\u002Fblog\u002Fbeyond-prompt-engineering-a-methodology-for-more-reliable-ai-reasoning\">Beyond Prompt Engineering: A Methodology for More Reliable AI Reasoning\u003C\u002Fa>\u003C\u002Fb> — the methodological framework behind this series.","\u003Cb>\u003Ca href=\"https:\u002F\u002Fstajic.de\u002Fblog\u002Fprompt-invariance-does-the-conclusion-survive-the-prompt\">Prompt Invariance: Does the Conclusion Survive the Prompt?\u003C\u002Fa>\u003C\u002Fb> — testing whether conclusions remain stable under blind, inverted and adversarial formulations.","\u003Cb>\u003Ca href=\"https:\u002F\u002Fstajic.de\u002Fblog\u002Ffalsification-for-ai-reasoning-from-answers-to-tested-hypotheses\">Falsification for AI Reasoning: From Answers to Tested Hypotheses\u003C\u002Fa>\u003C\u002Fb> — replacing confirmatory reasoning with competing hypotheses and disconfirming tests.","\u003Cb>\u003Ca href=\"https:\u002F\u002Fstajic.de\u002Fblog\u002Ffrom-research-protocol-to-a-general-ai-reasoning-framework\">From Research Protocol to General AI Reasoning Framework\u003C\u002Fa>\u003C\u002Fb> — generalizing the methodology across research, software, debugging and strategic analysis.",{},"2.31.6","Prompt wording is not neutral. Explore how framing, assumptions, instruction-following and sycophancy can shape AI reasoning—and why reliable conclusions require testing beyond the original prompt.",{"lang":7,"title":208,"content":210,"contentJson":1273,"excerpt":789},{"time":212,"blocks":1274,"version":788},[1275,1278,1281,1284,1287,1290,1293,1296,1299,1302,1305,1310,1313,1316,1319,1322,1325,1328,1331,1334,1337,1340,1343,1346,1349,1352,1355,1358,1361,1364,1367,1370,1373,1376,1379,1382,1387,1390,1393,1396,1399,1402,1405,1408,1411,1414,1417,1420,1423,1426,1429,1432,1435,1438,1441,1444,1447,1450,1453,1456,1459,1462,1465,1468,1471,1474,1477,1480,1483,1486,1489,1492,1495,1498,1501,1504,1507,1510,1513,1516,1519,1522,1525,1528,1531,1534,1537,1540,1543,1548,1551,1554,1557,1560,1563,1566,1569,1572,1575,1578,1581,1584,1587,1590,1593,1598,1601,1604],{"id":215,"data":1276,"type":218,"tunes":1277},{"text":217},{},{"id":221,"data":1279,"type":218,"tunes":1280},{"text":223},{},{"id":226,"data":1282,"type":218,"tunes":1283},{"text":228},{},{"id":231,"data":1285,"type":236,"tunes":1286},{"text":233,"caption":234,"alignment":235},{},{"id":239,"data":1288,"type":218,"tunes":1289},{"text":241},{},{"id":244,"data":1291,"type":42,"tunes":1292},{"text":246,"level":247},{},{"id":250,"data":1294,"type":218,"tunes":1295},{"text":252},{},{"id":255,"data":1297,"type":218,"tunes":1298},{"text":257},{},{"id":260,"data":1300,"type":218,"tunes":1301},{"text":262},{},{"id":265,"data":1303,"type":218,"tunes":1304},{"text":267},{},{"id":270,"data":1306,"type":277,"tunes":1309},{"meta":1307,"items":1308,"style":276},{},[274,275],{},{"id":280,"data":1311,"type":218,"tunes":1312},{"text":282},{},{"id":285,"data":1314,"type":42,"tunes":1315},{"text":287,"level":247},{},{"id":290,"data":1317,"type":218,"tunes":1318},{"text":292},{},{"id":295,"data":1320,"type":218,"tunes":1321},{"text":297},{},{"id":300,"data":1323,"type":218,"tunes":1324},{"text":302},{},{"id":305,"data":1326,"type":218,"tunes":1327},{"text":307},{},{"id":310,"data":1329,"type":218,"tunes":1330},{"text":312},{},{"id":315,"data":1332,"type":236,"tunes":1333},{"text":317,"caption":234,"alignment":235},{},{"id":320,"data":1335,"type":42,"tunes":1336},{"text":322,"level":247},{},{"id":325,"data":1338},{"text":327,"type":218,"tunes":1339},{},{"id":330,"data":1341},{"text":332,"type":218,"tunes":1342},{},{"id":335,"data":1344},{"text":337,"type":218,"tunes":1345},{},{"id":340,"data":1347},{"text":342,"type":218,"tunes":1348},{},{"id":345,"data":1350},{"text":347,"type":218,"tunes":1351},{},{"id":350,"data":1353},{"text":352,"type":218,"tunes":1354},{},{"id":355,"data":1356,"type":42,"tunes":1357},{"text":357,"level":247},{},{"id":360,"data":1359},{"text":362,"type":218,"tunes":1360},{},{"id":365,"data":1362},{"text":367,"type":218,"tunes":1363},{},{"id":370,"data":1365,"type":236,"tunes":1366},{"text":372,"caption":234,"alignment":235},{},{"id":375,"data":1368},{"text":377,"type":218,"tunes":1369},{},{"id":380,"data":1371,"type":236,"tunes":1372},{"text":382,"caption":234,"alignment":235},{},{"id":385,"data":1374},{"text":387,"type":218,"tunes":1375},{},{"id":390,"data":1377},{"text":392,"type":218,"tunes":1378},{},{"id":395,"data":1380},{"text":397,"type":218,"tunes":1381},{},{"id":400,"data":1383,"type":277,"tunes":1386},{"meta":1384,"items":1385,"style":276},{},[404,405,406,407,408],{},{"id":411,"data":1388,"type":218,"tunes":1389},{"text":413},{},{"id":416,"data":1391,"type":42,"tunes":1392},{"text":418,"level":247},{},{"id":421,"data":1394,"type":218,"tunes":1395},{"text":423},{},{"id":426,"data":1397,"type":218,"tunes":1398},{"text":428},{},{"id":431,"data":1400,"type":236,"tunes":1401},{"text":433,"caption":234,"alignment":235},{},{"id":436,"data":1403,"type":218,"tunes":1404},{"text":438},{},{"id":441,"data":1406,"type":218,"tunes":1407},{"text":443},{},{"id":446,"data":1409,"type":218,"tunes":1410},{"text":448},{},{"id":451,"data":1412,"type":218,"tunes":1413},{"text":453},{},{"id":456,"data":1415,"type":42,"tunes":1416},{"text":458,"level":247},{},{"id":461,"data":1418,"type":218,"tunes":1419},{"text":463},{},{"id":466,"data":1421,"type":218,"tunes":1422},{"text":468},{},{"id":471,"data":1424,"type":218,"tunes":1425},{"text":473},{},{"id":476,"data":1427,"type":218,"tunes":1428},{"text":478},{},{"id":481,"data":1430,"type":218,"tunes":1431},{"text":483},{},{"id":486,"data":1433,"type":218,"tunes":1434},{"text":488},{},{"id":491,"data":1436,"type":218,"tunes":1437},{"text":493},{},{"id":496,"data":1439,"type":42,"tunes":1440},{"text":498,"level":247},{},{"id":501,"data":1442,"type":218,"tunes":1443},{"text":503},{},{"id":506,"data":1445,"type":218,"tunes":1446},{"text":508},{},{"id":511,"data":1448,"type":218,"tunes":1449},{"text":513},{},{"id":516,"data":1451,"type":218,"tunes":1452},{"text":518},{},{"id":521,"data":1454,"type":218,"tunes":1455},{"text":523},{},{"id":526,"data":1457,"type":42,"tunes":1458},{"text":528,"level":247},{},{"id":531,"data":1460,"type":218,"tunes":1461},{"text":533},{},{"id":536,"data":1463,"type":236,"tunes":1464},{"text":538,"caption":234,"alignment":235},{},{"id":541,"data":1466,"type":218,"tunes":1467},{"text":543},{},{"id":546,"data":1469,"type":218,"tunes":1470},{"text":548},{},{"id":551,"data":1472,"type":218,"tunes":1473},{"text":553},{},{"id":556,"data":1475,"type":236,"tunes":1476},{"text":558,"caption":234,"alignment":235},{},{"id":561,"data":1478,"type":218,"tunes":1479},{"text":563},{},{"id":566,"data":1481,"type":42,"tunes":1482},{"text":568,"level":247},{},{"id":571,"data":1484,"type":218,"tunes":1485},{"text":573},{},{"id":576,"data":1487,"type":218,"tunes":1488},{"text":578},{},{"id":581,"data":1490,"type":218,"tunes":1491},{"text":583},{},{"id":586,"data":1493,"type":236,"tunes":1494},{"text":588,"caption":234,"alignment":235},{},{"id":591,"data":1496,"type":218,"tunes":1497},{"text":593},{},{"id":596,"data":1499,"type":218,"tunes":1500},{"text":598},{},{"id":601,"data":1502,"type":218,"tunes":1503},{"text":603},{},{"id":606,"data":1505,"type":42,"tunes":1506},{"text":608,"level":247},{},{"id":611,"data":1508,"type":218,"tunes":1509},{"text":613},{},{"id":616,"data":1511,"type":218,"tunes":1512},{"text":618},{},{"id":621,"data":1514,"type":218,"tunes":1515},{"text":623},{},{"id":626,"data":1517,"type":218,"tunes":1518},{"text":628},{},{"id":631,"data":1520,"type":236,"tunes":1521},{"text":633,"caption":234,"alignment":235},{},{"id":636,"data":1523,"type":218,"tunes":1524},{"text":638},{},{"id":641,"data":1526,"type":42,"tunes":1527},{"text":643,"level":247},{},{"id":646,"data":1529,"type":218,"tunes":1530},{"text":648},{},{"id":651,"data":1532,"type":218,"tunes":1533},{"text":653},{},{"id":656,"data":1535,"type":218,"tunes":1536},{"text":658},{},{"id":661,"data":1538,"type":218,"tunes":1539},{"text":663},{},{"id":666,"data":1541,"type":218,"tunes":1542},{"text":668},{},{"id":671,"data":1544,"type":277,"tunes":1547},{"meta":1545,"items":1546,"style":276},{},[675,676,677,678,679,680],{},{"id":683,"data":1549,"type":218,"tunes":1550},{"text":685},{},{"id":688,"data":1552,"type":218,"tunes":1553},{"text":690},{},{"id":693,"data":1555,"type":42,"tunes":1556},{"text":695,"level":247},{},{"id":698,"data":1558,"type":218,"tunes":1559},{"text":700},{},{"id":703,"data":1561,"type":218,"tunes":1562},{"text":705},{},{"id":708,"data":1564,"type":218,"tunes":1565},{"text":710},{},{"id":713,"data":1567,"type":236,"tunes":1568},{"text":715,"caption":234,"alignment":235},{},{"id":718,"data":1570,"type":218,"tunes":1571},{"text":720},{},{"id":723,"data":1573,"type":218,"tunes":1574},{"text":725},{},{"id":728,"data":1576,"type":218,"tunes":1577},{"text":730},{},{"id":733,"data":1579,"type":735,"tunes":1580},{},{},{"id":738,"data":1582,"type":42,"tunes":1583},{"text":740,"level":247},{},{"id":743,"data":1585,"type":218,"tunes":1586},{"text":745},{},{"id":748,"data":1588,"type":218,"tunes":1589},{"text":750},{},{"id":753,"data":1591,"type":42,"tunes":1592},{"text":755,"level":756},{},{"id":759,"data":1594,"type":277,"tunes":1597},{"meta":1595,"items":1596,"style":276},{},[763,764,765,766,767],{},{"id":770,"data":1599,"type":735,"tunes":1600},{},{},{"id":774,"data":1602,"type":42,"tunes":1603},{"text":776,"level":247},{},{"id":779,"data":1605,"type":277,"tunes":1608},{"meta":1606,"items":1607,"style":276},{},[783,784,785,786],{},"Post erfolgreich abgerufen",{"items":1611,"source":1696,"manualIds":1697,"manualMatchedIds":1698},[1612,1619,1626,1633,1640,1647,1654,1661,1668,1675,1682,1689],{"id":1613,"slug":1614,"title":1615,"excerpt":1616,"featuredImage":1617,"publishedAt":1618},"486","source-of-truth-in-ai-systems-where-reliable-knowledge-actually-comes-from","AI系统中的真相来源：可靠知识究竟从何而来","事实来源（Source of Truth）定义了对于特定事实或状态，哪个来源具有权威性。了解它与RAG、溯源、记忆、上下文、向量数据库和记录系统有何不同。","\u002Fuploads\u002F2026\u002F10\u002Fsource-of-truth-in-ai-systems-where-reliable-knowledge-actually-comes-from-1791479103235-6bq9em.webp","2026-10-08T13:02:00.000Z",{"id":1620,"slug":1621,"title":1622,"excerpt":1623,"featuredImage":1624,"publishedAt":1625},"476","mcp-vs-a2a-vs-ucp-vs-ap2-vs-a2ui-the-agent-protocol-stack-explained","MCP vs A2A vs UCP vs AP2 vs A2UI：智能体协议栈详解","MCP、A2A、UCP、AP2 和 A2UI 常被描述为相互竞争的智能体标准。它们大多解决的是不同的互操作性问题。本指南将每个协议映射到其实际标准化的边界，并展示它们如何在同一个生产系统中协同工作。","\u002Fuploads\u002F2026\u002F09\u002Fmcp-vs-a2a-vs-ucp-vs-ap2-vs-a2ui-the-agent-protocol-stack-explained-1790352625869-2ezle0.webp","2026-09-25T12:09:00.000Z",{"id":1627,"slug":1628,"title":1629,"excerpt":1630,"featuredImage":1631,"publishedAt":1632},"386","enterprise-start-here-your-gateway-to-operational-excellence","Enterprise Start Here: Your Gateway to Operational Excellence","New to our enterprise platform? This guide provides a structured onboarding path, from foundational reference models to actionable playbooks, runbooks, and assessments designed for seamless implementation.","\u002Fuploads\u002F2026\u002F02\u002Ffrom-global-business-to-the-kitchen-a-reverse-communication-system-that-still-scales-1771251820950-zhp2bz.webp","2026-03-01T11:55:00.000Z",{"id":1634,"slug":1635,"title":1636,"excerpt":1637,"featuredImage":1638,"publishedAt":1639},"453","zbt-z8102ax-openwrt-2102-firmware-review","ZBT Z8102AX OpenWrt 21.02 固件评测：足够稳定，但能否面向未来？","ZBT Z8102AX 运行的是厂商修改版 OpenWrt 21.02 固件，内核版本为 5.4.246。实际测试中，该固件运行稳定，可保持路由器连续数日正常工作，但老旧的基础版本引发了关于安全性、调制解调器控制、升级路径及长期可维护性的重要问题。","\u002Fuploads\u002F2026\u002F06\u002Fopenwrt-router-review-dual-sim-04-1781620594653-uvf5oy.webp","2026-06-16T10:39:00.000Z",{"id":1641,"slug":1642,"title":1643,"excerpt":1644,"featuredImage":1645,"publishedAt":1646},"471","how-to-know-whether-an-ai-agent-actually-used-the-right-evidence","如何判断一个AI智能体是否真正使用了正确的证据","AI代理可以引用来源，却仍然使用错误的证据。本文介绍一种实用方法，用于核查主张支持、来源权威性、适用性、出处，以及证据是否实际影响了答案。","\u002Fuploads\u002F2026\u002F09\u002Fhow-to-know-whether-an-ai-agent-actually-used-the-right-evidence-1790351317188-o5z9ve.webp","2026-09-25T11:47:00.000Z",{"id":1648,"slug":1649,"title":1650,"excerpt":1651,"featuredImage":1652,"publishedAt":1653},"473","openai-agents-api-vs-agents-sdk-vs-responses-api-what-should-you-build-on-in-2026","OpenAI Agents API 与 Agents SDK 与 Responses API：2026 年你应该基于什么来构建？","OpenAI 的智能体技术栈在 2026 年 9 月发生了变化。本架构指南按运行时归属将 Agents API、Agents SDK、Responses API 和 Codex SDK 区分开来——以便团队能够选择正确的控制边界，而不是比较产品名称。","\u002Fuploads\u002F2026\u002F09\u002Fopenai-agents-api-vs-agents-sdk-vs-responses-api-what-should-you-build-on-in-2026-1790351846714-zi7lus.webp","2026-09-25T11:56:00.000Z",{"id":1655,"slug":1656,"title":1657,"excerpt":1658,"featuredImage":1659,"publishedAt":1660},"454","zbt-z8102ax-rm500u-ea-5g-modem-test","Quectel RM500U-EA在ZBT Z8102AX中：5G频段、o2德国及实际信号表现","ZBT Z8102AX 使用移远 RM500U-EA 调制解调器实现 4G 和 5G 连接。在首次实际测试中，该路由器成功连接至德国 o2 网络，使用 LTE Band 3 和 NR n28 频段。调制解调器工作正常，但更深层次的诊断功能如 RSRP、RSRQ、SINR、频段锁定及小区行为仍需进一步测试。","\u002Fuploads\u002F2026\u002F06\u002Fopenwrt-router-review-dual-sim-06-1781620597879-qay2sx.webp","2026-06-16T08:39:00.000Z",{"id":1662,"slug":1663,"title":1664,"excerpt":1665,"featuredImage":1666,"publishedAt":1667},"460","ai-agent-reliability-why-the-final-answer-is-not-enough","AI Agent可靠性：为什么最终答案并不足够","正确的输出并不能证明推理的正确性、执行的安全性，或系统的可信赖性。","\u002Fuploads\u002F2026\u002F09\u002Fai-agent-reliability-why-the-final-answer-is-not-enough-1788955466306-pl0qhz.webp","2026-09-09T04:01:00.000Z",{"id":1669,"slug":1670,"title":1671,"excerpt":1672,"featuredImage":1673,"publishedAt":1674},"478","what-is-rag-the-simplest-explanation-of-how-it-works","什么是RAG？对其工作原理的最简单解释","RAG听起来很复杂，但想法很简单：在AI回答之前，它先从知识源查找有用的信息，并将该信息提供给语言模型。本指南使用一个简单的思维模型来解释RAG、LLM、状态、记忆和工具。","\u002Fuploads\u002F2026\u002F09\u002Fwhat-is-rag-the-simplest-explanation-of-how-it-works-1790377492124-khjagt.webp","2026-09-25T19:03:00.000Z",{"id":1676,"slug":1677,"title":1678,"excerpt":1679,"featuredImage":1680,"publishedAt":1681},"459","ollama-is-not-the-product-building-production-ready-open-llm-applications","Ollama 并非产品：构建可投入生产的开源大语言模型应用","使用Ollama运行本地模型很简单。但构建一个可用于生产环境的开源大语言模型（Open-LLM）应用则更具挑战性：它需要RAG（检索增强生成）、访问控制、供应商抽象、评估、日志记录、部署规范，以及围绕模型构建受控的应用层。","\u002Fuploads\u002F2026\u002F06\u002Follama-is-not-the-product-building-production-ready-open-llm-applications-1782679361640-h0usqf.webp","2026-06-28T16:39:00.000Z",{"id":1683,"slug":1684,"title":1685,"excerpt":1686,"featuredImage":1687,"publishedAt":1688},"436","triggers","企业AI运行手册中回滚触发器的全面指南","本指南探讨回滚触发器，这是企业AI运行手册中的关键机制，能自动检测异常并启动回滚以维持系统稳定。了解如何配置、监控和优化这些触发器，以实现稳健的AI部署。","\u002Fuploads\u002F2026\u002F06\u002Ftriggers-1781855500994-cibuou.webp","2026-03-01T16:51:00.000Z",{"id":1690,"slug":1691,"title":1692,"excerpt":1693,"featuredImage":1694,"publishedAt":1695},"474","migrating-from-openai-agents-sdk-to-the-agents-api-what-actually-changes-architecturally","从OpenAI Agents SDK迁移到Agents API：架构上究竟有哪些变化？","从 OpenAI Agents SDK 迁移到新的 Agents API 并不是简单的导入重命名。运行时边界发生了变化：代理循环、持久会话、编排、上下文压缩与恢复都向托管执行框架迁移。本指南说明哪些应当迁移、哪些应当保留在您的应用程序中，以及如何在切换前验证迁移。","\u002Fuploads\u002F2026\u002F09\u002Fmigrating-from-openai-agents-sdk-to-the-agents-api-what-actually-changes-architecturally-1790352171968-ienxr9.webp","2026-09-25T12:01:00.000Z","fallback",[],[]]