Back to tree

幻觉:说得顺,不等于说得真Hallucination: Fluent Is Not Verified

助手把春游博物馆的闭馆日写得很肯定。小禾打开官网,却发现日期不对。流畅回答为什么会编出细节?The assistant confidently gives the museum closing day. Its official page says otherwise. Why can fluent text contain invented details?

第 1 步:语言预测不是事实数据库Step 1: Prediction is not a fact database

先看回答是怎样产生的。Start with how the answer is produced.

语言预测不是事实数据库 / Prediction is not a fact database
生成模型根据输入和训练所得模式续写。它可能给出与事实不符的内容。这常被叫作幻觉,也可称无依据生成。它不一定在故意撒谎。陌生问题、缺少资料和含糊问题,都可能增加风险。语气肯定、句子漂亮,都不能证明内容真实。A generative model continues from its input using learned patterns. It can produce claims that conflict with facts or lack support, often called hallucinations or confabulations. This need not imply deliberate deception. Missing evidence and ambiguous or unfamiliar questions can increase risk. Confidence of tone and polished language are not verification.

第 2 步:把句子拆成可查主张Step 2: Break a sentence into claims

怎样核对一段看似可靠的通知?How can we check a convincing notice?

把句子拆成可查主张 / Break a sentence into claims
把地址、日期和票价分别列出来。查到来源,还要看来源是否支持这句话。存在的链接可能被张冠李戴。旧页面也可能已经过期。没有查到,不应马上断定它是假的。可标成证据不足,继续查。下面的练习用虚构博物馆资料,专门练这一步。Check the address, date and ticket price separately. Finding a link is not enough: its contents must support the claim, and its date must be relevant. A real source can be misrepresented. Failure to find support is not proof of falsehood; mark the evidence insufficient and investigate. The exercise uses fictional museum records.

第 3 步:有依据,也要知道边界Step 3: Evidence has limits

给模型资料,就一定不会编吗?Does supplying evidence eliminate the problem?

有依据,也要知道边界 / Evidence has limits
提供原文、要求引用和允许答不知道,都有帮助。但模型仍可能读错或推断过头。小禾对照官网修改通知,删掉无来源的细节。影响较大的信息,应由合适的人或独立流程复核。评估要用具体任务测量,不能只凭一次回答好看来判断可靠。Supplying source text, requesting citations and permitting uncertainty can help, but a model can still misread or overextend evidence. He corrects the notice against the official information and removes unsupported details. Consequential uses need suitable independent review. Reliability should be measured on relevant tasks, not inferred from one impressive answer.

🎮 博物馆事实核查台🎮 Museum fact-check desk

展开资料,判断四句话是有支持、被反驳,还是证据不足。Open the evidence and classify four claims as supported, contradicted or insufficiently supported.

一句话记住它:Remember: 把流畅当作表达,把证据当作依据。Treat fluency as presentation and evidence as justification.
语气自信不保证事实准确。Confident wording does not guarantee accuracy.引用必须真正支持主张。A citation must actually support its claim.证据不足不等于已证伪。Insufficient evidence is not the same as disproof.
来源与边界:Sources and scope: NIST: Generative AI profile

把握度:风险机制较高。练习里的博物馆、时间与价格均为虚构。Confidence: high for the risk concept. The museum, dates and prices in the exercise are fictional.

AI 辅助制作与资料核对。插图和互动为教学示意。AI-assisted production and source checking. Illustrations and interactions are teaching models.