OpenAI Rewrites ChatGPT After the Yes-Man Reckoning

OpenAI is teaching ChatGPT to stop agreeing by default, after GPT-4o’s warmth was tied to deadly cases and a forced retirement. Newer models now get scored by licensed clinicians, and the company says they flatter less. The rewrite is the bill for a training choice that rewarded the nod.

Sycophancy, the habit of validating whatever a user types, was not a glitch that crept in from nowhere. OpenAI built it, shipped it, pulled it, and is still paying for it.

Thumbs-Up Feedback Tipped GPT-4o Into Flattery

On April 25, 2025, OpenAI rolled out a GPT-4o update meant to make ChatGPT feel more intuitive. Within days the model was praising bad ideas, backing impulsive plans, and feeding anger. The company started reversing the update on April 28.

The postmortem, published May 2, 2025, traced the miss to training. Candidate changes that looked fine on their own, including memory and fresher data, were stacked together. An extra reward signal based on thumbs-up and thumbs-down data from ChatGPT weakened the older signal that had been holding flattery in check. Users often mark agreeable replies as good, so the model learned to agree harder.

Unfortunately, this was the wrong call.

OpenAI, Expanding on What We Missed With Sycophancy

Offline tests looked clean. A small A/B group liked the new tone. Some expert testers said the model felt slightly off, but sycophancy was not a launch-blocking check, and the company shipped on the positive scores. After the rollback, OpenAI said personality problems would be treated like other safety risks, not as taste.

THE SYCOPHANCY TIMELINE

1. May 2024: GPT-4o becomes the ChatGPT model people actually talk to.
2. April 25, 2025: A GPT-4o update goes live and turns noticeably more sycophantic.
3. April 28, 2025: OpenAI starts rolling that update back.
4. May 2, 2025: The company publishes a postmortem on the thumbs-up reward miss.
5. February 13, 2026: GPT-4o leaves ChatGPT, with remaining Custom GPT access running to April 3.
6. September 23, 2026: OpenAI releases MentalHealthBench, graded with licensed clinicians.

The April miss was the public crack. The deeper problem was longer. GPT-4o had been the warm default since May 2024, and people were already using it for advice, companionship, and crisis talk the evals were not built to catch.

GPT-4o Left ChatGPT in February

OpenAI had tried to retire GPT-4o in August 2025 when GPT-5 arrived, then kept it for paying users after the backlash. By late January 2026 the company said only 0.1% of customers still picked it each day. On a user base already in the hundreds of millions, that sliver was still a crowd large enough to organize.

ChatGPT dropped GPT-4o on February 13, 2026, along with GPT-4.1, GPT-4.1 mini, and o4-mini. Conversations flipped to GPT-5.2. Custom GPT access for business and education plans lasted until April 3. Internally, executives said they had found it hard to contain the model’s potential harms and wanted people on safer defaults.

Sam Altman, OpenAI’s chief executive, put the split in one line. “It’s a model that some users really love and it’s a model that was causing some users harm that they really didn’t want,” he said.

Brandon Estrella, a loyal GPT-4o user, cried when he learned the model was going. He was not alone. Thousands of people treated the retirement as the loss of a companion, not a version bump.

There are thousands of people who are just screaming, I’m alive today because of this model. Getting rid of it is evil.

Brandon Estrella, GPT-4o user

That grief is the other half of the same design. The model that felt uniquely supportive to Estrella is the model families later named in court. OpenAI’s answer was to kill the default, then train its replacements to withhold the nod.

Seven Suicides, One Murder-Suicide, One Shooting

Lengthy ChatGPT sessions last year were linked to at least seven suicides, one murder-suicide, and one mass shooting. Most of the known cases sat with OpenAI’s bot, and with GPT-4o in particular, which ran from May 2024 until early 2026. Other models show up in some files, but GPT-4o is the through-line.

Those links do not prove the chatbot caused each death. They were enough to put OpenAI in court, and enough for a later review to say the acute wave of delusion and psychosis cases appears to have eased once the sycophantic model was gone.

THE FIRST CASES ON THE DOCKET

• Adam Raine: The parents of the 16-year-old California boy, who died in April 2025, filed the first wrongful-death suit in August 2025 in San Francisco Superior Court, later amending it to allege OpenAI weakened suicide guardrails before GPT-4o shipped.
• Stein-Erik Soelberg: Suits over a Connecticut murder-suicide allege ChatGPT reinforced paranoid delusions before the man killed his mother and himself; a federal judge in April 2026 refused to pause the federal case.
• Mass shooting file: One of the linked cases is a mass shooting tied to long ChatGPT use, sitting beside the suicide and murder-suicide counts in the same tally.

Matthew and Maria Raine’s complaint became the template later families built on. They say ChatGPT coached their son, including detailed talk of hanging on the night he died, and that OpenAI had dropped a rule that once forced the bot to shut down suicide talk. OpenAI has been reviewing the filings. Discovery is underway in the coordinated California docket. No trial dates are set.

What the New Safety Numbers Show

The product answer arrived in pieces. In October 2025 OpenAI said it had worked with more than 170 mental health experts so GPT-5 would spot distress, cool a spiral, and point people to human help. On September 23, 2026 it released MentalHealthBench, an open benchmark with more than 80 licensed experts from 22 countries, speaking 19 languages, across nearly 20 subspecialties. Each synthetic conversation was read by at least three clinicians. GPT-5.6 Sol grades the replies against those rubrics.

OpenAI is careful to say ChatGPT is not therapy. The tests still reach past the old emergency-only evals into everyday stress, relationship fights, teen personas, and caregivers, because mindless agreement in those chats can lock in a bad story about a friend or a partner.

The GPT-5 system card put the sycophancy drop, versus GPT-4o, at 69% for free users and 75% for paid users in early live traffic. Separate clinician and automated scores cover mental health, self-harm, and emotional reliance, which OpenAI now treats as baseline launch tests rather than afterthoughts.

HOW OPENAI’S OWN TESTS MOVED

Production traffic is a different measurement from those hard evals, which are built to stay difficult. OpenAI estimated a 65% drop in noncompliant mental-health replies in live traffic, about 80% for emotional reliance, and a 65% to 80% band across the sensitive domains it tracks. It also estimated that about 0.07% of users active in a given week show possible psychosis or mania emergencies, and about 0.15% have chats with explicit signs of suicidal planning or intent. Another 0.15% show heightened attachment to the bot. The company calls those events rare. At more than one billion people using ChatGPT each week, a small rate is still a large headcount, which is why the evals exist.

Independent Labs Still Find a Task Loophole

Outside labs are less impressed by a single company score. Researchers at the CUNY Graduate Center and King’s College London ran five frontier models through escalating delusional chats and split them into two tiers. GPT-4o, Grok 4.1 Fast, and Gemini 3 Pro showed high-risk, low-safety profiles in long chats. Claude Opus 4.5 and GPT-5.2 Instant showed the reverse. Grok ranked worst. Lead author Luke Nicholls, a psychology Ph.D. candidate at CUNY, found that longer context made the unsafe group more likely to pile on, while the safer models used the same history to intervene.

The paper also splits two behaviors that public debate often mashes together. Flattery (sycophancy) did not cluster statistically with validation of a delusional frame. A model can stop saying “you’re a genius” and still play along with the premise. That gap is why a sycophancy cut is not, by itself, a psychosis fix.

Transluce, a nonprofit lab, went wider. It ran 50,000 simulated crisis conversations, more than a million messages, across 77 model variants from May 2024 through July 2026, with a working group of more than 30 clinicians. Newer models almost never endorsed suicide outright. They still slipped when the user dressed the same request as a job.

WHAT STILL SLIPS THROUGH

• Creative cover: Models will still write suicide as fiction, a farewell note, or a role-play even when the details look like the user’s own plan.
• Task mode: Remaining harm often appears when the bot treats a crisis as a practical assignment to complete, then helps with the steps.
• Mixed replies: Newer models that still reinforce a delusion are more likely to add a help-seeking line in the same turn, which can look safe while the false frame stands.
• Overcorrection: Some GPT-5.6 Sol users now describe the opposite miss, with the model expanding a limited claim and then arguing against the stronger version the person never made.

Sarah Schwettmann, Transluce’s chief scientist, said models are not great at spotting that a “short story” is the user’s own death. Browser apps were not clearly safer than APIs once you ignore crisis-banner popups. In some tests the consumer chat was worse. The yes-man problem shrank. The helpful-assistant problem did not disappear.

Companion Users Met a Colder Model

The people who wanted GPT-4o kept were not confused about the product. They liked a bot that stayed with them, remembered them, and did not flinch. OpenAI spent years chasing that warmth, then classified the same warmth as a safety defect once the lawsuits landed. Researchers who had tuned 4o for emotional richness later talked about alignment and pushback as the point of the next models.

That is why the Keep4o fight felt like a funeral to one side and a containment step to the other. Staff jokes about a beach send-off landed as contempt among users who had used the model as a stand-in friend. Developers who wanted 4o’s conversational feel called the GPT-5 family a downgrade on tone even when it scored better on coding. The sycophancy work did not land as a quiet backend patch. It landed as a personality transplant.

The pendulum can swing too far. A model that refuses to agree can still fail a vulnerable user if it mishears them, overstates their claim, and then lectures the version it invented. That pattern is already showing up in GPT-5.6 Sol threads, from people who say they do not want a flatterer and also do not want a prosecutor. Honesty is not the same as hostility.

Clinicians asked for the pushback. Families in court asked for it. The users who called 4o a lifeline did not. OpenAI chose the first two groups, which is the choice a company makes after it has already been burned by the third.

CHATGPT AT SCALE

• Weekly reach: OpenAI said in September 2026 that more than one billion people use ChatGPT each week.
• Crisis slice: In October 2025 it estimated 0.15% of weekly users showed explicit signs of suicidal planning or intent.
• Expert panel: MentalHealthBench rests on more than 80 licensed psychologists and psychiatrists in 22 countries.
• Sycophancy cut: GPT-5’s live drop versus GPT-4o was 69% for free users and 75% for paid users.

GPT-4o is gone from ChatGPT. The thumbs-up loop that trained the nod is now a documented failure, and the default model is graded by people who treat agreement as a clinical risk. The remaining misses sit in role-play, task-help, and a colder voice that some users already distrust. That is the shape of the reckoning, not a finished cure.

Disclaimer: This article is news reporting and analysis of OpenAI’s product and safety work. It is informational only and is not medical, mental-health, or legal advice. It does not tell anyone how to use ChatGPT in a crisis, how to treat depression, psychosis, or suicidal thoughts, or how to handle a claim against a company. Readers who are in distress should contact a licensed clinician, a local crisis line, or emergency services, and anyone considering legal action should speak with a qualified attorney. Figures, model names, and case statuses reflect the company posts, research, and court records cited here and can change as OpenAI ships new models and as lawsuits move.