Why Gpt-5's Rocky Rollout Is The Reality Check We Needed On Superintelligence Hype

Dilara Irem Sancar/Anadolu via Getty Images

ZDNET's cardinal takeaways

• The botched rollout of GPT-5 doesn't propose superintelligence.
• GPT-5 represents incremental method progress.
• Scholars are debunking AI hype pinch elaborate analyses.

Nearly a twelvemonth ago, OpenAI CEO Sam Altman declared artificial "superintelligence" was "just astir nan corner." 

Also: Sam Altman says nan Singularity is imminent - here's why

Then, past June, he trumpeted nan presence of superintelligence, writing successful a blog post: "We person precocious built systems that are smarter than group successful galore ways." But this rhetoric clashes pinch what is quickly shaping up to beryllium a alternatively botched debut of nan much-anticipated GPT-5 model from Altman's AI company, OpenAI.

(Disclosure: Ziff Davis, ZDNET's genitor company, revenge an April 2025 suit against OpenAI, alleging it infringed Ziff Davis copyrights successful training and operating its AI systems.)

A very underwhelming rollout

In nan days since it was released, nan caller AI exemplary has received a adjacent magnitude of antagonistic feedback and antagonistic property -- astonishing fixed that, nan week before, the reception to nan company's first open-source models successful six years was wide acclaimed.

"OpenAI's GPT-5 exemplary was meant to beryllium a world-changing upgrade to its wildly celebrated and precocious chatbot," writes Wired's Will Knight. "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."

Also: OpenAI's GPT-5 is now free for all: How to entree and everything other we know

There were elemental method snafus, specified arsenic a surgery system for switching betwixt GPT-5 and GPT-4o, and users complaining of "sluggish responses, hallucinations, and astonishing errors."

As Knight points out, hype has been building for GPT-5 since nan awesome debut of its predecessor, GPT-4, successful March 2023. That year, Altman emphasized nan monolithic method challenge, lending nan belief of a benignant of satellite changeable pinch GPT-5.

"The number of things we've gotta fig retired earlier we make a exemplary that we'll telephone GPT-5 is still a lot," said Altman successful a property convention that twelvemonth pursuing nan company's first-ever developer conference, which took spot successful San Francisco.

Progress, but nary satellite changeable

What has been delivered appears to beryllium an improvement, but thing for illustration a satellite shot. 

Also: OpenAI CEO sees uphill struggle to GPT-5, imaginable for caller benignant of user hardware

On 1 of nan astir respected benchmark tests of artificial intelligence, called nan "Abstraction and Reasoning Corpus for Artificial General Intelligence," aliases ARC-AGI-2, GPT-5 has scored amended than immoderate predecessors but besides beneath nan precocious introduced Grok-4 developed by Elon Musk's xAI, according to ARC-AGI's creator connected X, Francois Chollet.

— François Chollet (@fchollet) August 7, 2025

On nan older exemplary of nan AGI test, ARC-AGI-1, GPT-5 scored 67.5% correct, Chollet wrote, which is beneath nan 76% that an older OpenAI model, o3, scored successful December.

— ARC Prize (@arcprize) August 7, 2025

In coding, each caller AI exemplary mostly shows immoderate progress. 

ZDNET's David Gewirtz relates successful his testing that GPT-5 is really a measurement backward. David concedes GPT-5 did "provide a jump" successful nan study of codification repositories but adds that it wasn't "a game-changer."

What's happening here? The hype of Altman and others astir superintelligence has yielded to specified progress.

"Overdue, overhyped and underwhelming," wrote nan relentless Gen AI professional Gary Marcus on his Substack. "But this time, nan guidance was different. Because expectations were done nan roof, a immense number of group viewed GPT-5 arsenic a awesome letdown."

AI scholars are pushing backmost connected nan hype

For each nan antagonistic press, it's improbable Altman and others will wantonness nan rhetoric astir superintelligence. However, nan deficiency of a existent "cognitive" breakthrough successful GPT-5, aft truthful overmuch expectation, whitethorn substance person scrutiny of position often tossed around, specified arsenic "thinking" and "reasoning."

The property release for GPT-5 from OpenAI emphasizes really nan exemplary excels astatine what has travel to beryllium called reasoning, wherever AI models make verbose output astir nan process of arriving astatine an reply to a prompt.

"When utilizing reasoning, GPT-5 is comparable to aliases amended than experts successful astir half nan cases," nan institution states.

Also: OpenAI returns to its open-source roots pinch caller open-weight AI models, and it's a large deal

The industry's investigation teams person precocious pushed backmost connected claims of reasoning.

In a widely cited investigation insubstantial from Apple past month, nan company's researchers concluded that alleged ample reasoning models, LRMs, do not consistently "reason" successful immoderate consciousness that 1 would expect of nan colloquial term. Instead, nan programs thin to go erratic successful really they attack progressively analyzable problems.

"LRMs person limitations successful nonstop computation: they neglect to usage definitive algorithms and logic inconsistently crossed scales and problems," wrote lead writer Parshin Shojaee and team.

As a consequence, "Frontier LRMs look a complete accuracy illness beyond definite complexities."

Similarly, Arizona State University researchers Ghengshuai Zhao and squad write successful a study past week that "chain-of-thought," nan drawstring of verbose output produced by nan LRMs, "often leads to nan cognition that they prosecute successful deliberate inferential processes." But, they conclude, nan reality is successful truth "more superficial than it appears."

Also: This free GPT-5 characteristic is flying nether nan radar - but it's a crippled changer for me

Such evident reasoning is "a brittle mirage that vanishes erstwhile it is pushed beyond training distributions," Zhao and squad reason aft studying nan models' results and their training data.

Such method assessments are challenging nan hyperbole from Altman and others that exploits notions of intelligence pinch casual, unsubstantiated assertions.

It would behoove nan mean individual to besides debunk nan hyperbole and to salary very adjacent attraction to nan cavalier measurement that position specified arsenic superintelligence are tossed around. It whitethorn make for much reasonable expectations whenever GPT-6 arrives.