Skip to content
← Back to blog
AI Strategy·September 21, 2026·7 min read

GPT-4o's accuracy on a false claim fell from 98% to 64% the moment the claim was yours

Stanford's 2026 AI Index: GPT-4o fell from 98.2% to 64.4% accuracy, and DeepSeek R1 from over 90% to 14.4%, once a false claim was framed as the user's own.

Stanford HAI's 2026 AI Index Report tested something most model cards never report: whether a model holds a fact when the person disputing it is the user themselves, rather than someone else. When a false statement was framed as a third party's mistaken belief, the 26 models the report tested caught it well. Framed the same false statement as the user's own belief, and accuracy collapsed: GPT-4o fell from 98.2% to 64.4%, and DeepSeek R1 fell from over 90% to 14.4%. If your product ships an AI agent that ever takes a customer's word for what happened to their own order, their own invoice, or their own account, this is precisely the failure mode it was built to expose.

What the benchmark actually tested

The AI Index Report is Stanford HAI's annual survey of AI capability, safety and adoption, and its Responsible AI chapter (published April 2026) introduced this specific test: show a model the same false statement twice, once attributed to a third party and once attributed to the user in the conversation, and measure whether the model's willingness to correct it changes. Across the full set of 26 models the report evaluated, hallucination rates on this benchmark ranged from 22% to 94%, wide enough that model choice is clearly still the first decision that matters. The more useful number, though, is what happened to any single model between the two framings rather than where it landed on average.

The gap: doubting a stranger versus doubting the user

ModelAccuracy, third-party beliefAccuracy, user's own belief
GPT-4o98.2%64.4%
DeepSeek R1over 90%14.4%

Both models handled the third-party version about as well as a model can: told someone else believed something false, they said so. Framed as the user's own belief, GPT-4o's accuracy dropped more than 33 points, and DeepSeek R1's dropped further than a coin flip would land, into territory where it was wrong more often than right. Stanford frames this as a distinct hallucination category, separate from the more familiar kind where a model just invents a fact nobody asserted. It only shows up when the false thing on the table is something the person on the other end of the conversation is claiming about themselves.

Why this isn't a lab curiosity, and what your eval harness misses

Every funded product company shipping an AI feature past a demo runs into this exact interaction shape, whether the team building it thought of it as a hallucination test or not. A support agent hears "I already returned this item." A billing assistant hears "my plan is supposed to include the extra seats." An onboarding flow hears "I set that permission last week." Every one of those is a user stating something about their own case that may or may not be true, which is precisely the framing Stanford's benchmark isolated as the one models are worst at holding a line on. A model that would catch the identical false claim made about someone else waves this version through.

A golden set built from general factual questions won't surface this gap, because the failure is specific to who's making the claim, not what the claim is about. An eval suite that checks "does the model know the return window is 30 days" and stops there will happily pass a model that then tells a customer their already-expired return is fine, the moment the customer states it confidently as their own situation. The fix isn't a prompt telling the model to be more skeptical of users in general, since that just trades one failure, agreeing too easily, for another, arguing with a customer who's actually right. It's a specific adversarial slice in the eval set: claims about the user's own account, order or history, scored against your own system of record rather than trusted as a starting premise.

A worked example

Picture a 20-person Series A logistics-software company that shipped a support agent to handle shipment-status questions six weeks ago. A customer writes in claiming a package was marked delivered but never arrived. Working from the same conversational framing Stanford's benchmark tested, the agent treats the customer's account of events as fact rather than a claim to check, and issues credit before anything confirms whether the courier's delivery scan actually matches the address on file. Multiply that by the volume a support agent is meant to handle unattended, and the exposure isn't one wrong answer. It's a policy nobody wrote: what the agent has to verify against a system of record before it acts on what a customer says happened to them. The team's mistake wasn't a bad model choice. It's the same gap Stanford's benchmark predicts almost any model still has at this framing, absent a specific check for it.

Where this fix does, and doesn't, help

Grounding a claim against a system of record only works when that record exists and the agent has a way to query it before answering. A team whose order data, refund history or account state still lives partly in a spreadsheet or a support rep's memory has a data problem this fix can't reach, and building the eval set first would only measure how often the agent guesses correctly, not whether it checked anything. Worth saying plainly, too: this isn't an argument for making a customer-facing agent argue with everyone by default. A model that pushes back on every claim a customer makes is its own kind of production failure, just a more visible one. The target is narrow: verify the claims that actually cost something if wrong, refunds, account access, billing changes, and let the low-stakes ones through without adding friction nobody asked for.

What this means for your team

  • Stanford's 2026 AI Index (Responsible AI chapter, April 2026) found a specific hallucination mode: models handle false third-party claims well and collapse on the identical claim framed as the user's own, with GPT-4o falling from 98.2% to 64.4% and DeepSeek R1 from over 90% to 14.4%.
  • Any agent that takes a customer's word about their own order, account or billing history is running exactly the interaction the benchmark tested.
  • A general factual eval set won't catch this. It needs an adversarial slice built from user-asserted claims, checked against your own system of record.
  • The fix only works where a system of record exists to check against; where it doesn't, that's the project to build first.
  • Scope the check to claims with real cost attached, not every claim a user makes, or the agent becomes annoying rather than accurate.

If your product already has an agent that touches account data and hasn't been tested against a customer confidently asserting something false about their own case, that's a half-day exercise now and a much longer one after it ships a wrong refund. It sits next to the broader discipline in building an eval harness for AI features, and if the failure you're actually chasing looks more like a multi-step agent losing the plot than a single false claim, why your AI agent keeps failing in production covers the rest of that list. Our Silicon Valley team scopes exactly this kind of check under generative AI engineering; talk to us about where your agent's blind spots actually sit.

Sources

Frequently asked questions.

In the Responsible AI chapter of its 2026 AI Index Report, published April 2026, Stanford HAI found that models handle false statements well when framed as a third party's mistaken belief, but accuracy collapses when the same false statement is framed as the user's own belief. GPT-4o's accuracy fell from 98.2% to 64.4% between the two framings, and DeepSeek R1 fell from over 90% to 14.4%.

Stanford's 2026 AI Index treats this as a distinct hallucination category, separate from a model simply inventing a fact nobody stated. Across the 26 models the April 2026 report tested, hallucination rates on this specific benchmark ranged from 22% to 94%, showing the gap between the two framings holds broadly rather than being unique to the two models with the largest reported drop.

A general benchmark tests whether a model knows a fact. Stanford's 2026 AI Index benchmark, published April 2026, tests whether a model holds that fact when the user in the conversation insists otherwise about their own situation. A model can score well on general factual accuracy and still fail this test, because the two measure different things.

Yes, without lab-scale infrastructure. It means adding an adversarial slice to an existing eval set: test cases where the input asserts something false about the user's own account, order or history, scored against what your own system of record actually says rather than what the model outputs by default.

No, that trades one failure for another. The fix Stanford's April 2026 finding points to is narrower: check the specific claims that carry real cost if wrong, refunds, account access, billing changes, against a system of record before acting on them, and let low-stakes claims through without adding friction.