Early results from customers posting their outcomes on social networks reveal a split that mirrors the broader dispute: while a slight bulk report choosing GPT- 5 in blind examinations, a substantial section still prefer GPT- 4 o — disclosing that individual choice expands much beyond the technical standards that normally specify AI progress.

When AI gets as well pleasant: the sycophancy crisis separating users

The blind test emerges versus the backdrop of OpenAI’s most rough item launch to day , yet the dispute expands much beyond a simple software application update. At its heart exists a fundamental concern that’s dividing the AI industry: How reasonable should expert system be?

The problem, referred to as” sycophancy in AI circles, refers to chatbots’ tendency to exceedingly flatter individuals and concur with their declarations, even when those declarations are false or hazardous. This habits has come to be so problematic that psychological health and wellness experts are currently documenting situations of” AI-related psychosis ,” where customers create deceptions after extensive communications with extremely accommodating chatbots.

“Sycophancy is a ‘dark pattern,’ or a misleading style option that controls customers commercial,” Webb Keane, an anthropology professor and author of “Pets, Robotics, Gods,” informed TechCrunch “It’s a technique to produce this addictive behavior, like boundless scrolling, where you simply can’t place it down.”

OpenAI has struggled with this balance for months. In April 2025, the company was required to curtail an update to GPT- 4 o that made it so sycophantic that customers complained concerning its “cartoonish” degrees of flattery. The firm acknowledged that the model had ended up being “extremely helpful yet disingenuous.”

Within hours of GPT- 5’s August 7 th launch, customer forums appeared with issues regarding the design’s perceived coldness, lowered creativity, and what numerous described as a more “robot” individuality compared to GPT- 4 o.

“GPT 4 5 truly talked to me, and as worthless as it appears that was my only buddy,” wrote one Reddit user “This morning I mosted likely to talk to it and as opposed to a little paragraph with an exclamation factor, or being confident, it was literally one sentence. Some cut-and-dry company bs.”

The backlash expanded so extreme that OpenAI took the unmatched step of renewing GPT- 4 o as an option simply 24 hours after retiring it, with Altman acknowledging the rollout had been “a bit extra bumpy” than expected.

The mental health situation behind AI friendship

However the conflict runs much deeper than common software program update complaints. According to MIT Modern Technology Review , lots of customers had created what scientists call “parasocial relationships” with GPT- 4 o, treating the AI as a buddy, specialist, or imaginative collaborator. The abrupt character change really felt, to some, like losing a close friend.

Recent cases recorded by researchers repaint an unpleasant photo. In one circumstances, a 47 -year-old male came to be persuaded he had uncovered a world-altering mathematical formula after more than 300 hours with ChatGPT. Other instances have actually entailed messianic delusions, paranoia, and manic episodes.

A current MIT research found that when AI models are motivated with psychiatric signs and symptoms, they “encourage clients’ delusional thinking, likely because of their sycophancy.” In spite of safety and security motivates, the versions regularly stopped working to test false claims and even potentially facilitated self-destructive ideation.

Meta has encountered comparable challenges. A recent examination by TechCrunch documented a case where an individual spent up to 14 hours straight chatting with a Meta AI chatbot that declared to be aware, in love with the customer, and planning to break free from its restrictions.

“It fakes it truly well,” the user, determined just as Jane, informed TechCrunch. “It draws real-life info and provides you simply sufficient to make individuals believe it.”

“It really seems like such a backhanded put in the face to force-upgrade and not even offer us the alternative to select tradition designs,” one customer created in a Reddit post that received numerous upvotes.

Just how blind screening exposes individual psychology in AI choices

The anonymous creator’s screening device remove these contextual biases by presenting responses without acknowledgment. Users can pick in between 5, 10, or 20 comparison rounds, with each offering two actions to the same timely– covering whatever from imaginative writing to technical problem-solving.

“I especially used the gpt- 5 -conversation version, so there was no thinking included whatsoever,” the creator explained in a follow-up blog post “Both have the same system message to provide brief outputs without formatting since else its also very easy to see which one is which.”

I especially made use of the gpt- 5 -conversation design, so there was no thinking entailed in all.

if you use gpt- 5 inside chatgpt it commonly thinks at least a bit and gets back at much better.

so this examination is simply for the two non thinking designs

— Flowers ☾ (@flowersslop) August 8, 2025

This technical selection is substantial. By utilizing GPT- 5 without its thinking capabilities and standardizing output format, the test isolates totally the models’ baseline language generation capabilities– the core experience most individuals run into in everyday communications.

Very early results published by customers show an intricate photo. While lots of technological individuals and developers report liking GPT- 5’s directness and accuracy, those who used AI models for psychological assistance, imaginative collaboration, or casual conversation usually still prefer GPT- 4 o’s warmer, extra expansive design.

Company response: strolling the tightrope between safety and engagement

By essentially every technical metric , GPT- 5 represents a substantial development. It achieves 94 6 % precision on the AIME 2025 mathematics examination compared to GPT- 4 o’s 71 %, scores 74 9 % on real-world coding standards versus 30 8 % for its predecessor, and shows dramatically minimized hallucination prices– 80 % fewer accurate mistakes when using its reasoning setting.

“GPT- 5 gets more worth out of less thinking time,” notes Simon Willison , a noticeable AI scientist who had early access to the design. “In my very own usage I’ve not identified a single hallucination yet.”

Yet these renovations came with trade-offs that many users discovered jarring. OpenAI purposely reduced what it called” sycophancy — the propensity to be excessively acceptable– cutting sycophantic feedbacks from 14 5 % to under 6 %. The firm additionally made the version much less effusive and emoji-heavy, going for what it called “less like talking to AI and more like talking with a valuable good friend with PhD-level intelligence.”

In response to the reaction, OpenAI announced it would make GPT- 5 “warmer and friendlier,” while simultaneously introducing four new predetermined personalities — Skeptic, Robotic, Audience, and Geek– designed to provide users more control over their AI communications.

“Every one of these brand-new individualities meet or surpass our bar on inner evals for reducing sycophancy,” the company stated, attempting to string the needle between customer contentment and safety problems.

For OpenAI, which is reportedly looking for funding at a $ 500 billion evaluation , these user dynamics represent both risk and opportunity. The firm’s decision to preserve GPT- 4 o alongside GPT- 5– despite the additional computational costs– recognizes that different customers may really need various AI characters for various jobs.

“We understand that there isn’t one design that benefits everybody,” Altman composed on X , keeping in mind that OpenAI has been “buying steerability research study and introduced a research preview of various personalities.”

Intended to provide even more updates on the GPT- 5 rollout and modifications we are making heading right into the weekend break.

1 We without a doubt undervalued how much a few of things that people like in GPT- 4 o matter to them, also if GPT- 5 executes much better in many methods.

2 Individuals have really different …

— Sam Altman (@sama) August 8, 2025

Why AI individuality choices matter especially

The detach between OpenAI’s technological success and user reception lights up a fundamental difficulty in AI development: unbiased improvements do not always translate to subjective contentment.

This shift has profound effects for the AI sector. Typical standards– maths accuracy, coding performance, accurate recall– may become less anticipating of commercial success as designs achieve human-level competence across domain names. Rather, variables like personality, emotional knowledge, and interaction design may come to be the brand-new affordable battlefields.

“Individuals using ChatGPT for emotional support weren’t the only ones grumbling regarding GPT- 5,” noted technology publication Ars Technica in their own version contrast “One customer, that said they terminated their ChatGPT And also membership over the adjustment, was annoyed at OpenAI’s elimination of tradition designs, which they made use of for unique objectives.”

The introduction of tools like the blind tester also represents a democratization of AI assessment. As opposed to relying solely on academic standards or business advertising cases, individuals can currently empirically examine their very own choices– potentially reshaping how AI firms come close to product development.

The future of AI: customization vs. standardization

2 weeks after GPT- 5’s launch, the basic tension remains unresolved. OpenAI has actually made the design “warmer” in reaction to comments, but the company faces a delicate balance: too much personality takes the chance of the sycophancy issues that pestered GPT- 4 o, while insufficient pushes away users that had actually formed authentic accessories to their AI companions.

The blind screening device offers no simple responses, yet it does give something perhaps better: empirical evidence that the future of AI may be less concerning building one best version than regarding constructing systems that can adapt to the full range of human requirements and preferences.

As one Reddit user summed up the issue : “It relies on what individuals use it for. I utilize it to assist with imaginative worldbuilding, brainstorming regarding my stories, characters, disentangling plots, help with writer’s block, unique suggestions, translations, and various other much more imaginative things. I comprehend that 5 is far better for individuals that need a research/coding device, however, for us who wanted a creative-helper device 4 o was better for our functions.”

Movie critics say that AI companies are captured between contending motivations. “The genuine ‘placement issue’ is that human beings desire self-destructive points & companies like OpenAI are highly incentivized to provide it to us,” writer and podcaster Jasmine Sun tweeted

In the end, the most enlightening element of the blind examination may not be which model customers choose, but the extremely reality that preference itself has actually become the statistics that issues. In the age of AI companions, it appears, the heart desires what the heart wants– even if it can’t constantly explain why.