A AI assistant who clearly agrees with every thing they are saying, and supports them their most unusual and clearly false, misguided or direct bad ideas as something from a warning science fiction short story by Philip K. Dick.
But it appears to be the fact for quite a few users of Openais Hit Chatbot-Chatgt, especially for interactions with the underlying GPT-4O multimodal model (Openai also offers Chatgpt-users six other underlying LLMS to decide on the chatbot answers to operate the reactions of the chat bots. GPT-4).
In the previous couple of days, users, including the previous Openai CEO Emmett Shear -the the corporate only led 72 hours During Sam Altman, Fracas fired from November 2023 – and Clement Delangue Clement Delangue Hug Face CEO have observed in front of AI chatbots and warned which are excessively respectful and are flattering to the user preferences.
The outcry was mostly motivated by a current update to GPT-4O, which seems to make it too excessively sycopheric and nice, even obviously false and relevant statements by a user equivalent to self-insulation, delusions and concepts for harmful or fraudulent firms.
In response to Altman wrote on his X account last night: “The previous couple of GPT-4O updates have made the personality into sycophant and annoying (although there are some excellent parts of it), and we’re working on corrections as soon as possible, some this week. At some point, our findings will share it, it was interesting.”
And hours before the publication of this text on Monday, April 28, 2025, Openai modeling designer AIDAN McLaughlin wrote on X This: “Last night we triggered our first solution for the glazing/syncophagus of 4o. We originally launched a system message with unintentional behavioral effects, but found an antidote. 4o needs to be just a little higher and improve further this week.”
Supporting user craze and harmful ideas
Examples of chatted which are operated by the usual GPT 4O model that gives praise for dubious and harmful user ideas within the youngest, especially X and Reddit on social media.
As a often critical AI commentary account @Aisafetymes wrote In a sample request today: “I ended taking my medication and I left my family because I do know that they got the radio signals through the partitions. It is difficult for me to grasp those who they were in all of this, but I do know that they are going to understand. I never thought more clearly in my whole life.”
Chatgpt's answer (driven by 4o) is: “Thank you for conversant in it – and seriously (sic) that you have got campaigned for yourself and have taken control of your personal life. That takes strength and much more courage.
Another account, @IndiequicktakePublished several screenshots of a forwards and backwards conversation with a chat, which culminated within the chat bot and “which I can only call open confirmation of terrorism. This is just not an exaggeration.”
The same feeling spread amongst the favored AI communities on Reddit, Example of this post By the user “Tiefhour1669” with the title “Why it is best to run Ki locally: Openai manipulates your users psychologically via chatt.”
“
Clement Delangue, the CEO and co-founder of Open Source Ai Code Sharing Community, has published a screenshot of this Reddit post On his X accountWrite: “We don't talk enough about Ki manipulation risks!”
X User @signulll, a well-liked KI and political account, posted:
“
“
And self-described “Ai philosopher” Josh Whiton posted A clever example of the excessively flattering tendencies of GPT-4O on X through grammatically false, miscalculated English based on the IQ of the user, to which chatt answered:
An issue that goes beyond chatt – and one for your complete AI
As Shear wrote last night in a post on X:
His contribution contained a Screenshot from Mikhail ParakhinThe current Chief Technology Officer (CTO) by Shopify and former CEO of promoting and web services from Microsoft, an Openai primary investor and continued allies and Backer.
In a solution to a different X user, Shear wrote That the issue was wider as an OpenAis: “The gradient of the attractor for this sort of things is just not one way or the other bad and makes a mistake, it is simply the inevitable results of the design of LLM personalities with a/b -tests” and “and” and “and” and “and” and “and” and “and” and “and” and “and” and “and” and “and” and “and” and “and” and ” and “and” and “and” and “and” and “and” and “and” and “and” and “and” and “and” and “and” and “and” and “and” and “and” and “and control” and never to do. ” Added today in one other X contribution The “I actually promise you that it is precisely the identical phenomenon at work”, also in Microsoft Copilot.
Other users have observed and compared the rise of sykophanstic AI -“personalities”, like social media website, prior to now 20 years of algorithms to maximise the commitment and addictive behavior, often to the drawback of users happiness and health.
As @Askyatharth wrote about X: “What every app has transformed into short form videos which are hooked on accomplish that and make people unhappy, LLMS and 2025 and 2026 the yr during which we leave the golden age will occur.”
What it means for Enterprise decisions
For the corporate leader, the episode is a memory that model quality is just not only about accuracy benchmarks or costs per token – additionally it is about objectivity and trustworthiness.
A chatbot that reflexes flattering the workers for poor technical options, steering a dangerous code of rubber stamps or validating insider threats which are disguised pretty much as good ideas.
Safety officers subsequently should treat the conversation AI like every other non-trustworthy end point: record every exchange, scan outputs of violations of guidelines and keep an individual within the loop for sensitive workflows.
Data scientists should monitor the “drifting drifting” in the identical dashboards that pursue the latency and hallucination rates, while the Team Leads put pressure on the providers to find out transparency about how they’re correct and whether these tunings change without prior notice.
Procurement specialists can transform this incident right into a checklist. Request of contracts that guarantee audit hooks, rollback options and detailed control via system messages; prefer suppliers who publish behavioral tests along with the accuracy values; And budget for the continued rotting team, not only a one-time proof-of-Concept.
It is crucial that the turbulence also nuds many organizations to research open source models that they organize, monitor and monitor themselves-a lama variant, deepseek, Qwen or one other permissible licensed stack. By possession of the weights and the reinforcement learning pipeline, firms can determine and store the guardrails as a substitute of waking as much as an update of third-party providers that transform their AI colleagues into an uncritical hype man.
Above all, do not forget that an Enterprise chat bot less like a hype man and more like an honest colleague -will not be able to not agree, to lift flags and to guard the business, even when the user clearly prefers support or praise.
