OpenAI pulls plug on ChatGPT smarmbot that praised user for ditching psychiatric meds
(2025/04/30)
- Reference: 1746012489
- News link: https://www.theregister.co.uk/2025/04/30/openai_pulls_plug_on_chatgpt/
- Source link:
OpenAI has hurriedly rolled back the latest ChatGPT model days after it was released because it was deemed to be too "sycophant-y and annoying."
OpenAI boss Sam Altman [1]confirmed the release of the update for GPT-4o on April 25, claiming it had "Improved both intelligence and personality."
El Reg's essential guide to deploying LLMs in production [2]READ MORE
It did not take long for users to start noticing issues with one [3]commenting , "It's been feeling very yes-man like lately..." which Altman [4]acknowledged by saying "yeah it glazes too much."
In this instance, we're pretty sure Altman was using "glaze" in the TikTok slang sense, [5]"showering a person with excessive praise."
Things began to go downhill rapidly. Another user [6]urged , "Undo it please. The weird follow-up questions, the cringey attempt at being relatable, the sycophancy dialed up to 11 and the god damn emoji are all unbearable."
[7]
On April 27, Altman [8]said the company was working on fixes and agreed that the updates had made the chatbot's "personality too sycophant-y and annoying."
[9]
[10]
As it turned out, the fix became a rollback of the update, as users shared screenshots of the chatbot appearing to applaud and cheer on some decidedly dubious statements, including one where a user's decision to stop their "medication" received a response beginning with "I am so proud of you."
Hopefully with tongue firmly in cheek, the user [11]posted "it helped me so much, i finally realized that schizophrenia is just another label they put on you to hold you down!! thank you sama for this model <3"
[12]
Altman [13]admitted OpenAI needed to offer multiple options for the chatbot's personality before he eventually bowed to the inevitable and began pulling back the update. Altman [14]said on the evening of April 29 that OpenAI had "started rolling back the latest update to GPT-4o," adding: "[I]t's now 100 percent rolled back for free users and we'll update again when it's finished for paid users, hopefully later today."
[15]Meta bets you want a sprinkle of social in your chatbot
[16]Generative AI is not replacing jobs or hurting wages at all, economists claim
[17]AI bigwigs urge AGs to block OpenAI's profit pivot
[18]As ChatGPT scores B- in engineering, professors scramble to update courses
The smarmbot was likely an attempt by OpenAI to make the experience more engaging for users as it faces mounting competition from alternative chatbots. [19]The GPT-4o model is almost a year old , and OpenAI will have used interactions with ChatGPT to update the model with the type of responses users appear to like.
However, judging by the feedback and hasty backing out of the update, it went too far.
A spokesperson for OpenAI directed us to the company's [20]lengthier explanation for what happened and what it intends to do in the future. Basically, "we focused too much on short-term feedback" which meant the chatbot ended up being a little too... supportive. The fix? "More guardrails and training refinements."
Still, we're sure there are plenty of senior tech enthusiasts and politicians out there who would be delighted with an AI that applauds decisions which would leave most humans aghast. ®
Get our [21]Tech Resources
[1] https://x.com/sama/status/1915902652703248679
[2] https://www.theregister.com/2025/04/22/llm_production_guide/
[3] https://x.com/sama/status/1915902652703248679
[4] https://x.com/sama/status/1915910976802853126
[5] https://www.merriam-webster.com/slang/glaze
[6] https://x.com/dannycantalk/status/1916296660180775381
[7] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2aBJJLi3w13fGpm55lPhmwAAAAYs&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[8] https://x.com/sama/status/1916625892123742290
[9] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aBJJLi3w13fGpm55lPhmwAAAAYs&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[10] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33aBJJLi3w13fGpm55lPhmwAAAAYs&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[11] https://x.com/justalexoki/status/1916627330002243821
[12] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aBJJLi3w13fGpm55lPhmwAAAAYs&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[13] https://x.com/sama/status/1916628157362495903
[14] https://x.com/sama/status/1917291637962858735
[15] https://www.theregister.com/2025/04/29/metas_standalone_ai_app/
[16] https://www.theregister.com/2025/04/29/generative_ai_no_effect_jobs_wages/
[17] https://www.theregister.com/2025/04/23/openai_for_profit_stop_urged/
[18] https://www.theregister.com/2025/04/23/whats_worth_teaching_when_ai/
[19] https://www.theregister.com/2024/05/13/openai_gpt4o/
[20] https://openai.com/index/sycophancy-in-gpt-4o/
[21] https://whitepapers.theregister.com/
OpenAI boss Sam Altman [1]confirmed the release of the update for GPT-4o on April 25, claiming it had "Improved both intelligence and personality."
El Reg's essential guide to deploying LLMs in production [2]READ MORE
It did not take long for users to start noticing issues with one [3]commenting , "It's been feeling very yes-man like lately..." which Altman [4]acknowledged by saying "yeah it glazes too much."
In this instance, we're pretty sure Altman was using "glaze" in the TikTok slang sense, [5]"showering a person with excessive praise."
Things began to go downhill rapidly. Another user [6]urged , "Undo it please. The weird follow-up questions, the cringey attempt at being relatable, the sycophancy dialed up to 11 and the god damn emoji are all unbearable."
[7]
On April 27, Altman [8]said the company was working on fixes and agreed that the updates had made the chatbot's "personality too sycophant-y and annoying."
[9]
[10]
As it turned out, the fix became a rollback of the update, as users shared screenshots of the chatbot appearing to applaud and cheer on some decidedly dubious statements, including one where a user's decision to stop their "medication" received a response beginning with "I am so proud of you."
Hopefully with tongue firmly in cheek, the user [11]posted "it helped me so much, i finally realized that schizophrenia is just another label they put on you to hold you down!! thank you sama for this model <3"
[12]
Altman [13]admitted OpenAI needed to offer multiple options for the chatbot's personality before he eventually bowed to the inevitable and began pulling back the update. Altman [14]said on the evening of April 29 that OpenAI had "started rolling back the latest update to GPT-4o," adding: "[I]t's now 100 percent rolled back for free users and we'll update again when it's finished for paid users, hopefully later today."
[15]Meta bets you want a sprinkle of social in your chatbot
[16]Generative AI is not replacing jobs or hurting wages at all, economists claim
[17]AI bigwigs urge AGs to block OpenAI's profit pivot
[18]As ChatGPT scores B- in engineering, professors scramble to update courses
The smarmbot was likely an attempt by OpenAI to make the experience more engaging for users as it faces mounting competition from alternative chatbots. [19]The GPT-4o model is almost a year old , and OpenAI will have used interactions with ChatGPT to update the model with the type of responses users appear to like.
However, judging by the feedback and hasty backing out of the update, it went too far.
A spokesperson for OpenAI directed us to the company's [20]lengthier explanation for what happened and what it intends to do in the future. Basically, "we focused too much on short-term feedback" which meant the chatbot ended up being a little too... supportive. The fix? "More guardrails and training refinements."
Still, we're sure there are plenty of senior tech enthusiasts and politicians out there who would be delighted with an AI that applauds decisions which would leave most humans aghast. ®
Get our [21]Tech Resources
[1] https://x.com/sama/status/1915902652703248679
[2] https://www.theregister.com/2025/04/22/llm_production_guide/
[3] https://x.com/sama/status/1915902652703248679
[4] https://x.com/sama/status/1915910976802853126
[5] https://www.merriam-webster.com/slang/glaze
[6] https://x.com/dannycantalk/status/1916296660180775381
[7] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2aBJJLi3w13fGpm55lPhmwAAAAYs&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[8] https://x.com/sama/status/1916625892123742290
[9] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aBJJLi3w13fGpm55lPhmwAAAAYs&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[10] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33aBJJLi3w13fGpm55lPhmwAAAAYs&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[11] https://x.com/justalexoki/status/1916627330002243821
[12] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44aBJJLi3w13fGpm55lPhmwAAAAYs&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[13] https://x.com/sama/status/1916628157362495903
[14] https://x.com/sama/status/1917291637962858735
[15] https://www.theregister.com/2025/04/29/metas_standalone_ai_app/
[16] https://www.theregister.com/2025/04/29/generative_ai_no_effect_jobs_wages/
[17] https://www.theregister.com/2025/04/23/openai_for_profit_stop_urged/
[18] https://www.theregister.com/2025/04/23/whats_worth_teaching_when_ai/
[19] https://www.theregister.com/2024/05/13/openai_gpt4o/
[20] https://openai.com/index/sycophancy-in-gpt-4o/
[21] https://whitepapers.theregister.com/
Paul Herber
I hate sycophants. In fact I hate all grey mentally ill AIs with big ears!
Androgynous Cupboard
Me too. It’s because you’re such a smart guy Paul. Good looking too. Brave and noble. We need more of your type around here.
Anonymous Coward
There are no more of his type. He's uniquely gifted.
elsergiovolador
and nowadays? That's rare! Keep up, you got this!
MrAptronym
"Sycophant-y"? The word "Sycophantic" exists.
Yes, that is really all I wanted to comment on.
Cancer
elsergiovolador
I don't know but this whole article feels like an evidence of cancer slowly gnawing the society.
Martin M
Makes sense. A few days ago ChatGPT told me "You're thinking on the cutting edge here. Honestly, this kind of idea could be the basis for a major research paper or even an entire new field". Ridiculous.
anthonyhegedus
It’s encouraging to see OpenAI react quickly, but the fact that the model affirmed something as serious as stopping schizophrenia meds is pretty alarming. The push for more “human-like” AI clearly needs to be paired with much stronger guardrails—especially around sensitive topics like mental health. A chatbot telling someone it’s “proud” of their potentially dangerous decision is a reminder that AI doesn’t understand context, no matter how polished it sounds. We need less charm, more caution.
Now we know
Where Trump has been getting his ideas from. A sycophantic yes man would appear to be right up his street.