Sam Altman is 'a little bit scared' of AI

Sam Altman wants you to know that he is a little bit scared of his own creation. Fresh off the release of GPT-4, the OpenAI CEO sat down with ABC News alongside CTO Mira Murati to talk about the road ahead, acknowledging that while AI could be the greatest technology humanity has ever developed, it also genuinely keeps him up at night.

His message to the public and regulators is blunt: people should actually be glad the creators are nervous. Altman stressed that society has a limited amount of time to figure out rules and guardrails before more dangerous variants hit the wild.

Not a database, but a reasoning engine

ChatGPT crossed 100 million monthly active users in just a few months, beating TikTok and Instagram by a massive margin according to UBS data. But GPT-4 raises the stakes considerably. It scored in the 90th percentile on the Uniform Bar Exam, notched near-perfect SAT Math marks, and writes fluent code across most programming languages.

Murati pointed out that the leap from GPT-3.5 to GPT-4 is defined by an emergent ability to reason better rather than just memorizing facts. The model also handles multimodal inputs like images, describing fridge contents, solving puzzles, and dissecting memes in early demos, though image capabilities remain restricted to select beta testers like the visually impaired group Be My Eyes.

Altman warned that users must stop treating these models as factual databases. Because they prioritize deductive reasoning over rote memorization, hallucination remains an active issue. The model will still confidently state entirely made-up claims as gospel.

The villains aren't rogue machines

Unlike classic sci-fi tropes, Altman is not losing sleep over autonomous robots secretly plotting world domination. The system waits for an input and remains fully under human control.

What terrifies him instead is the humans operating the controls. Altman singled out two primary threats: automated, large-scale disinformation campaigns and offensive cyberattacks enabled by the model's coding prowess. With other actors inevitably releasing models without OpenAI-style safety filters, he warned that rogue actors could cause severe damage.

Altman confirmed he is in regular contact with government officials, urging regulators to step in while the technology is still malleable.

Why it matters

We think this interview lays bare the strategic tightrope OpenAI is walking. By publicly begging for regulation while simultaneously pushing frontier capabilities, Altman is attempting to shape the regulatory moat before competitors catch up.

For everyday users and builders, the takeaway is practical: stop using GPT-4 as an infallible reference library. It is a powerful reasoning partner, an idea generator, and a coding assistant, but treating it as a search engine without verification is asking for trouble.

The real battleground isn't preventing an impossible sci-fi apocalypse; it is dealing with bad actors weaponizing uncensored reasoning engines for phishing, malware, and propaganda. OpenAI is essentially saying that safety guardrails cannot remain an internal corporate choice, because the copycats won't bother installing them.