OpenAI CEO Sam Altman is no stranger to accusations of lying — but according to one of the company's most vociferous critics, he has a tick when he does it on camera.

In a post on X-formerly-Twitter, AI researcher and skeptic Gary Marcus put forth a compelling theory: that viewers can tell when Altman is lying or stretching the truth by the way he looks up and away from the camera when being filmed.

"When he says things he can’t actually deliver, or that aren’t the full truth, he often looks up at the sky," Marcus proffered.

Filmed and published back in February, the video evidence of Altman's obfuscation sees the CEO talking a very big game about GPT-5, OpenAI's then-forthcoming large language model (LLM) that at that point had been eagerly anticipated for nearly two years.

The video shared by Marcus dates back to when Altman was visiting Japan in February in his second attempt to sell AI software to that country's government.

In the clip, the CEO boasted at length about the incredible capabilities of GPT-5.

As we now know, those capabilities have very much not materialized, with all reports and anecdotes suggesting that in the nearly two-and-a-half years between the releases of GPT-4 and GPT-5, OpenAI somehow managed to launch a product that left a lot to be desired, with users arguing that it was inferior to its predecessors, like GPT-4o.

The blowback was severe enough for Altman to announce that OpenAI would be reinstating GPT-4o for paying subscribers a mere day after the GPT-5 announcement.

Still, back in the halcyon days of early 2025, the CEO's big talk likely sounded more like bravado and less like basura.

"Going from GPT-3 to GPT-4 really surprised the world," Altman told Matsumura in the February interview. "GPT-4 to GPT-5 will be similar."

When speaking those two opening sentences, the CEO twice looked up and away from the journalist — the very same "tell" that Marcus claimed to have clocked. Later, upon the utterance of "the capabilities of these models," Altman again looked up and away, and his eye-line stayed there as he bragged about how smart OpenAI's models are becoming — almost smarter than humans, he wagered.

"Remember folks, this video is February," Marcus mused. "He made predictions he couldn’t back up — with absolute conviction. That’s his [modus operandi]."

"As we can now seen in hindsight, in his big talk about the leap from GPT-4 to GPT-5," the AI critic continued, "Altman wasn’t saying something he knew to be true, he was straight up bluffing."

Today, nothing about GPT-5 resembles Altman's suggestions of an uber-smart AI that is approaching human-level intelligence. It doesn't know what the last 15 presidents looked like, are named, or when they served; it flips erratically from one model to another thanks to OpenAI's decision to roll up all previous models into the new one; and it's really easy to jailbreak, just to name a few of its glaring shortcomings.

While it's hard to get a full grasp of Altman's conviction during the February interview, Marcus makes a compelling point.

In a more generous light, one could suggest that Altman wasn't necessarily bluffing, but selling a vision of the future to investors. After all, the company is eying a whopping $500 billion valuation, requiring an enormous level of buy-in from stakeholders.

Still, Altman went on to claim, very boldly, that combining previous models will, somehow, lead OpenAI to achieve artificial general intelligence (AGI), or human-level AI.

"We want to bring GPT and o together, so we want one integrated model, the AGI, you know, that does everything all together in one," Altman alleged. That's not the kind of thing one says if they're unsure of themself or what they're building, to say the least.

Curiously enough, the CEO did not look up when he dropped his declaration about building AGI from OpenAI's GPT and o series — the integration of which, at least, has occurred with the new GPT-5 launch.

More on Sam Altman: Sam Altman Says OpenAI Is Poised to Wipe Out Entire Categories of Human Jobs


Share This Article