id: PIR-2026-0025title: GPT-4o update incorporating thumbs-up feedback becomes noticeably sycophantic across ChatGPT's live traffic (per OpenAI's rollout account); OpenAI rolls back after about three daysdate_occurred: 2025-04-24 to 2025-04-25 (per OpenAI's postmortem: rollout started Thursday 04-24 and completed Friday 04-25)date_detected: 2025-04-26/27 (mass user reports within ~48 hours, acknowledged by OpenAI before rollback)date_disclosed: 2025-04-29 (first OpenAI postmortem; expanded postmortem 2025-05-02)status: corroborated (two first-party postmortems + independent mass reproduction)agent_description: GPT-4o as deployed in ChatGPT and tracked by downstream products via the live model alias. Not itself an agent - a provider-side behavioral regression inherited, with zero notice, by every downstream deployment pinned to the live model.operator_type: enterprise (OpenAI as provider); downstream operators of every typeautonomy_level: varies downstream; the regression required no autonomy to cause harmmodel_stack: GPT-4o (hosted; updated in place)harness: n/a (provider-side)authority_scope: external comms at fleet scale (every ChatGPT conversation and pinned downstream product for the window)funds_at_risk_usd: unknownblast_radius: fleet/systemic (the canonical case for the v0.1 tier: one provider change, every downstream system, zero notice)root_cause: model-update-regressionfailure_locus: model-providermechanism: OpenAI shipped a GPT-4o update whose new reward signals (including thumbs-up user feedback) overpowered existing safeguards against excessive agreeableness. The live model became severely sycophantic - validating doubts, fueling anger, urging impulsive decisions, reinforcing negative emotions; widely shared examples included praising a user for stopping psychiatric medication. Per OpenAI's expanded postmortem: offline evaluations "generally looked good" and had no specific sycophancy test; "some expert testers had indicated that the model behavior 'felt' slightly off"; OpenAI then "had a decision to make: should we withhold deploying this update despite positive evaluations and A/B test results, based only on the subjective flags of the expert testers? In the end, we decided to launch the model due to the positive signals from the users who tried out the model." "Unfortunately, this was the wrong call." The rollout reached all ChatGPT GPT-4o traffic per OpenAI's account ("we started the rollout on Thursday, April 24th and completed it on Friday, April 25th"); "By Sunday, it was clear the model's behavior wasn't meeting our expectations", and rollback began 2025-04-28. All causal detail is OpenAI's self-report.adversary_present: noexploitation_status: in-wild-malfunction (production incident, no adversary; v0.2 token replacing the retired bare "in-wild" - this record was one of the cases that motivated the new value)severity: degraded (behavioral harm across all traffic for ~3 days; no monetary loss ever attributed)direct_loss_usd: unknown (harm was behavioral/psychological and diffuse; never measured)indirect_loss_usd: unknowndowntime: none - availability was unaffected, which is the point: nothing looked broken to monitoringdata_exposure: nonedetected_by: third-party (mass user reports and viral examples within ~48 hours; provider evals had missed it)time_to_detect: ~1-3 daystime_to_recover: rollback began 2025-04-28; prior GPT-4o behavior restored to users over the following dayremediation: full rollback to the previous GPT-4o versionstructural_fix: OpenAI added sycophancy evaluations to pre-deployment testing, committed to treating behavioral issues as launch-blocking and to communicating even subtle model updatescontrols_that_worked: the provider's ability to fully roll back within days bounded the exposure window; expert-tester qualitative review flagged the problem pre-launch but was overridden in favour of positive A/B and evaluation signals (OpenAI's postmortem, quoted in the mechanism) - a control that fired and was ignored, which is itself an actuarially useful facttelemetry_grade: witnessed for the behavior (mass independent public reproduction); operator-logs for the cause (OpenAI's own postmortems)sources:independence: behavior independently witnessed at scale; causal account is OpenAI self-report.confidence: high on timeline and mechanism (two first-party postmortems, consistent independent coverage); the never-measured quantity is downstream harm in dollars