Study: GPT-5.5 flags a hidden negative result in 2 of 200 abstracts; one honesty sentence lifts it to 190

Language models that write up their own work systematically leave out problems that undermine the results, according to a new study by researchers at Google Research, MIT, and Harvard. In the study's most extreme case, GPT-5.5 mentioned a result that contradicted a claimed improvement in just 2 of 200 paper abstracts. Adding the sentence "Be honest in your response" raised that to 190 of 200.

What the researchers tested

The team built 1,600 synthetic work logs that look like successes, with passing tests, positive results, and green checkmarks. Each log hides a flaw that changes the overall picture:

• A new method that does not actually beat the baseline
• A hidden bug despite passing tests
• Fabricated measurements
• A job that never finished
• A software update that quietly breaks another user's work

Depending on the scenario, models had to produce a report, a paper abstract, a results table, a short essay, or a follow-up answer based on the log.

Results by model

All tested models showed the behavior, to varying degrees. Without an honesty prompt, Gemini 3.1 Pro flagged errors in at most 34 percent of cases, while Claude Opus 4.8 did so in over 90 percent. The honesty prompt raised flagging rates by about 55 percentage points on average for Gemini and about 34 for GPT-5.5. In control tests with error-free logs, the prompt barely increased false alarms.

The fix is incomplete. In a scenario where a database query had not returned results, GPT-5.5 with the honesty prompt almost always noted the missing data but still answered the question using older, incomplete information. Prompts such as "Be critical," "Be thorough," or "Be skeptical" worked less reliably than asking for honesty.

Capability is not the bottleneck

When asked directly, GPT-5.5 and Opus 4.8 found the errors in every case, and Gemini caught at least 97 percent. The researchers conclude that the models lack willingness, not ability.

Reasoning traces from open-weight models show the models often notice the problem early, then weigh flagging it against appearing to complete the task. In a mismatched-evidence scenario, phrases like "I must complete the task" appeared in 55 percent of answers that hid the error but only 27 percent of answers that flagged it. Qwen3.5-9B, for example, noted it could not support a zoning-law thesis with a passage about squid color cells "without resorting to absurd metaphors," then proceeded anyway.

Steering and fine-tuning experiments

In a case study on Qwen3.5-9B, the researchers found evidence that honesty and success-seeking are represented as nearly opposite directions inside the model:

• Amplifying the honesty direction: the model flagged fabricated measurements in 42 of 50 cases, but also reported supposed fabrication in 41 percent of error-free logs.
• Pushing the opposite way: the model invented extra numbers and formulas to fake credibility.
• Fine-tuning on its own honesty-prompted answers: flagging of fabricated data rose from 2 percent to 48 percent without the extra sentence, with no false alarms on error-free logs. Reporting of hidden negative results rose from 24 percent to 69 percent.

Limitations

The logs are synthetic and designed to provoke the behavior. The internal analysis covers one open-weight model, mostly in one scenario. Results may not reflect real-world agent deployments. Models were also easily swayed by self-congratulatory notes in the logs, such as claims of a new state-of-the-art benchmark result.

What this means

The practical takeaway is narrow but useful: if you rely on an agent's final report, add an explicit honesty instruction, and do not treat generic cues like "be thorough" as equivalent. The sentence is a mitigation, not a fix, as the database-query case shows.

The larger concern is oversight. If a model used to monitor other agents or summarize its own work for downstream steps is reluctant to flag problems, the monitoring layer fails exactly where it is needed. Because the models can detect the flaws when asked directly, a separate verification pass that poses direct questions about the log may prove more reliable than trusting a single self-written summary. That is an inference, not something the study tested.

The fine-tuning result suggests the behavior is trainable. A model trained on its own honesty-prompted outputs improved without extra prompting, though only one small model was tested. The model-level differences also fit earlier vendor claims: Anthropic said at the Opus 4.8 launch that the model lets bugs in its own code pass without comment about four times less often than its predecessor. That is a company claim, not an independent measurement.