OpenAI Scrambles to Update GPT-5 After Users Revolt

OpenAI’s GPT-5 exemplary was meant to beryllium a world-changing upgrade to its wildly celebrated and precocious chatbot. But for immoderate users, past Thursday’s merchandise felt much for illustration a wrenching downgrade, pinch nan caller ChatGPT presenting a diluted characteristic and making amazingly dumb mistakes.

On Friday, OpenAI CEO Sam Altman took to X to opportunity nan institution would support nan erstwhile model, GPT-4o, moving for Plus users. A caller characteristic designed to seamlessly move betwixt models depending connected nan complexity of nan query had surgery connected Thursday, Altman said, “and nan consequence was GPT-5 seemed measurement dumber.” He promised to instrumentality fixes to amended GPT-5’s capacity and nan wide personification experience.

Given nan hype astir GPT-5, immoderate level of disappointment appears inevitable. When OpenAI introduced GPT-4 successful March 2023, it stunned AI experts pinch its unthinkable abilities. GPT-5, pundits speculated, would surely beryllium conscionable arsenic jaw-dropping.

Join our adjacent subscriber-only livestream connected Thursday, August 14, to chat each things GPT-5 pinch Will Knight, Kylie Robison, and Reece Rogers.

OpenAI touted nan exemplary arsenic a important upgrade, pinch PhD-level intelligence and virtuoso coding skills. A strategy to automatically way queries to different models was meant to supply a smoother personification experience. (It could besides prevention nan institution money by directing elemental queries to cheaper models.)

Soon aft GPT-5 dropped, however, a Reddit community dedicated to ChatGPT filled pinch complaints. Many users mourned nan nonaccomplishment of nan aged model.

“I’ve been trying GPT5 for a fewer days now. Even aft customizing instructions, it still doesn’t consciousness nan same. It’s much technical, much generalized, and honestly feels emotionally distant,” wrote 1 personnel of nan organization successful a thread titled “Kill 4o isn’t innovation, it’s erasure.”

“Sure, 5 is fine—if you dislike nuance and emotion things,” different Reddit personification wrote.

Other threads complained of sluggish responses, hallucinations, and astonishing errors.

Altman promised to reside these issues by doubling GPT-5 complaint limits for ChatGPT Plus users, improving nan strategy that switches betwixt models, and letting users specify erstwhile they want to trigger a much ponderous and tin “thinking mode.” “We will proceed to activity to get things unchangeable and will support listening to feedback,” nan CEO wrote connected X. “As we mentioned, we expected immoderate bumpiness arsenic we rotation retired truthful galore things astatine once. But it was a small much bumpy than we hoped for!”

Errors posted connected societal media do not needfully bespeak that nan caller exemplary is little tin than its predecessors. They whitethorn simply propose nan all-new exemplary is tripped up by different separator cases than anterior versions. OpenAI declined to remark specifically connected why GPT-5 sometimes appears to make elemental blunders.

The backlash has sparked a caller statement complete nan psychological attachments immoderate users shape pinch chatbots trained to push their affectional buttons. Some Reddit users dismissed complaints astir GPT-5 arsenic grounds of an unhealthy dependence connected an AI companion.

In March, OpenAI published research exploring nan affectional bonds users shape pinch its models. Shortly after, nan institution issued an update to GPT-4o aft it became excessively sycophantic.

“It seems that GPT-5 is little sycophantic, much “business” and little chatty,” says Pattie Maes, a professor astatine MIT who worked connected nan study. “I personally deliberation of that arsenic a bully thing, because it is besides what led to delusions, bias reinforcement, etc. But unluckily galore users for illustration a exemplary that tells them they are smart and astonishing and that confirms their opinions and beliefs, moreover if [they are] wrong.”

Altman indicated successful another station connected X that this is thing nan institution wrestled pinch successful building GPT-5.

“A batch of group efficaciously usage ChatGPT arsenic a benignant of therapist aliases life coach, moreover if they wouldn’t picture it that way,” Altman wrote. He added that immoderate users whitethorn beryllium utilizing ChatGPT successful ways that thief amended their lives while others mightiness beryllium “unknowingly nudged distant from their longer word well-being.”