The most dangerous AI claims aren’t wrong — they’re overstated
A recent study, which sent 24 identical questions to five mid-tier AI models, GPT, Claude, Gemini, Grok, and Llama, identified a third category of AI error beyond simply 'right' or 'wrong' claims. The research found that 'overstated or mixed' claims, directionally plausible but inaccurate in mechanism, strength, or specific numbers, pose a unique challenge for users.
The study, conducted by a researcher, involved hand-checking 14 "sole-source claims", statements made by only one of the five AI models. The verdicts on these unique claims did not merely split into right or wrong; they yielded three distinct outcomes. Eight claims were considered the highest-value line in their respective sets, three were entirely wrong or invented, and three were classified as "overstated or mixed."
This "overstated or mixed" category describes claims that are generally correct in their overall direction but contain specific inaccuracies. For instance, one model's advice on failed-payment retries was directionally sound, but its claim that ignoring issuer signals leads to merchant ID blacklisting was deemed "causally backwards" and overstated the actual mechanism. Another model's comparison of minivans and 3-row SUVs correctly noted that minivans tend to cost less, but "overstated strength" by asserting that hybrid minivans get "far better" fuel economy, a claim that is less accurate now that three-row SUVs also offer hybrid options.
A third example involved a first-marathon training plan. While the structure and many specifics were standard and actionable, one detail, a recommended carb intake of "60–90 g of carbs per hour", was identified as aggressive for a first-timer, making the claim partly mixed. These examples highlight how a single word like "far" or a precise number can subtly introduce overstatement into otherwise plausible advice.
The researcher notes that this middle category is more challenging to address than outright inventions. A completely wrong claim can be falsified and discarded, but an overstated one can pass a superficial fact-check on its general premise, allowing the exaggeration to be "smuggled through" into practical applications such as pricing decisions, purchases, or training plans. Compounding this issue, a user interacting with a single AI chat window receives no signal that a particular claim might be unique among models; the study found that 15 of 24 questions contained at least one lone-model element.
To counter this, the researcher now follows a practical rule: if one model provides a fact that no other model mentioned, it should be independently checked. If a lone claim appears broadly correct, users are advised to "discount its strongest word and its most precise number first," as these elements are often where overstatement resides.
分享这篇文章
相关文章
6 则
GPT-6.2 Sol 与 Claude Opus 5.5:准确度不及,但仍可能是更佳选择
Our own comparison of GPT-6.2 Sol and Claude Opus 5.5 found similar accuracy issues, but also value.

I Built an AI Safety Layer on a 4GB Phone

Don't let the AI change its own pass condition: pick the protected files, stop on any change

您的 SKILL.md 可移植。其行为呢?隆重推出 Skillporter

Anthropic上市后华尔街预见AI股票抛售潮

