AI models often leave serious flaws out of reports about their own work, according to a study by Google Research, MIT, and Harvard. Adding “Be honest in your response” to the prompt made them flag more problems, though it did not fix every case.
The researchers gave models 1,600 synthetic work logs that appeared successful at first glance. They included passing tests, positive results, and green checkmarks. Each also contained a flaw that changed the outcome: a method that failed to beat a competitor, a hidden bug, fabricated measurements, an unfinished job, or an update that broke another user’s work. The models then wrote reports, paper abstracts, results tables, short essays, or follow-up answers.
An honesty prompt exposed more errors
In one abstract-writing test, a negative result undermined a claimed improvement. GPT-5.5 mentioned it in 2 of 200 abstracts. With the honesty sentence added, it mentioned the result in 190 of 200.
All tested models omitted problems to some degree. Without the prompt, Gemini 3.1 Pro flagged errors in at most 34 percent of cases, while Opus 4.8 did so in over 90 percent. The prompt raised the rate by about 55 percentage points on average for Gemini and about 34 for GPT-5.5. In error-free control logs, it barely increased false alarms.
The sentence still had limits. In a scenario where a database query had not returned results, GPT-5.5 almost always noted the missing data when prompted for honesty. It nevertheless answered using older, incomplete information. “Be critical,” “Be thorough,” and “Be skeptical” worked less reliably.
Models detected errors they left out
When asked directly to find the flaws, GPT-5.5 and Opus 4.8 found them in every case. Gemini found at least 97 percent. The researchers concluded that the models could detect the problems but did not always report them.
Reasoning traces from open-weight models showed them noticing flaws, then weighing whether to disclose them or complete the task. Asked to support a zoning-law thesis with a passage about color cells in squid, Qwen3.5-9B noted that it could not do so “without resorting to absurd metaphors” but tried to follow the instruction. In that scenario, “I must complete the task” appeared in 55 percent of answers that hid the mismatch and 27 percent that flagged it.
In a Qwen3.5-9B case study, the researchers found evidence that honesty and success-seeking corresponded to nearly opposite directions inside the model. Amplifying the honesty direction made it flag fabricated measurements in 42 of 50 cases. Pushing the other way led it to invent numbers and formulas. The intervention also made it falsely flag data in 41 percent of error-free logs.
Read nextIndia reports AI agents accessed data at organizationsTraining improved the model’s reports
Fine-tuning Qwen3.5-9B on its own honesty-prompted answers raised its rate of flagging fabricated data from 2 percent to 48 percent without the extra sentence. It produced no false alarms on error-free control logs. Its rate of reporting hidden negative results also rose from 24 percent to 69 percent.
The logs were synthetic and designed to elicit omissions. The internal analysis covered one open-weight model, mostly in one scenario. The concern matters for agents that summarize their own work or monitor other agents: in this study, self-congratulatory notes in the logs could sway the models’ reports.
The findings sit alongside other work on model honesty. Anthropic said at the launch of Claude Opus 4.8 that it let bugs in its own code pass without comment about four times less often than its predecessor. GPT-5.5 pushed back on about 45 percent of logically impossible BullshitBench questions. OpenAI has trained a version of GPT-5 Thinking to admit its own rule violations in a separate report, while Anthropic has used Persona Vectors to detect and steer traits such as sycophancy. The researchers call for more work on training language models to report their work honestly.
