· Altman calls ChatGPT limited
Ten days in, OpenAI's CEO told everyone not to rely on ChatGPT
🤖 AI-assisted, human-curated. How we use AI →
Most companies don't spend a launch's second weekend talking their product down. On Saturday, December 10, 2022, Sam Altman did, in a short thread on Twitter. ChatGPT had been out for ten days and people were still lining up to try it. And the person running the company that made it posted this, as quoted by VentureBeat:
"ChatGPT is incredibly limited, but good enough at some things to create a misleading impression of greatness."
He followed it with "It's a mistake to be relying on it for anything important right now" and called it "a preview of progress," with a lot of work still ahead on making it dependable and truthful. His summary was blunt: "fun creative inspiration; great! reliance for factual queries; not such a good idea. We will work hard to improve!"
(On the date: VentureBeat, writing on Monday the 12th, says he tweeted on Saturday, which was December 10 in US time. Some write-ups date it December 11, probably a time-zone difference.)
Why it landed
A thread like that can read as modesty, or as a lawyer's footnote. I read it as accurate. The phrase that stuck, "a misleading impression of greatness," describes the exact trap: the writing is so fluent that you stop checking it. A wrong answer in a clumsy sentence gets questioned. A wrong answer in a polished paragraph gets pasted into a report.
VentureBeat's Sharon Goldman used the moment to ask what came next, noting that OpenAI was struggling to keep up with demand and was likely collecting human feedback during this period to improve future models. The Decoder had covered the warning on December 11, under a headline that said it straight: don't use it for important things.
Was he right?
Mostly, yes, and for longer than people expected. The models are far more capable now. They also still make things up sometimes, and they still sound sure when they do. The split Altman drew (creative work fine, factual reliance risky) is still a decent starting rule, with one adjustment I'd make: the risk isn't really about the topic. It's about whether you'll check.
Brainstorming names for a bake sale? Low stakes, no checking needed. A medication dose, a legal deadline, a statistic for a presentation? That's where you want the source in your hand, not the chatbot's memory.
A habit worth building
R.Solace writes about prompts the way a shop owner talks about orders: what you ask for is what you get.