The single most-upvoted sentiment across the AI subreddits this week wasn’t about a new model or a funding round. It was a raw, primal scream, phrased slightly differently in each community but always meaning the same thing: please, for the love of god, teach these things to say “I don’t know.”
And honestly? Correct. This is the defining character flaw of the entire technology. You can ask a frontier model a question it has no possible way of answering — an obscure API that doesn’t exist, a fact from last Tuesday, the contents of a file you didn’t give it — and instead of the three most useful words in the English language, it will look you dead in the eye and confidently invent something plausible. With citations. That are also fake.
It’s not lying. It’s worse.
A liar knows the truth and hides it. These models don’t know they don’t know — they’re pattern-completing machines optimized to produce the most likely-sounding next token, and “I’m not sure” is statistically rarer in their training data than a confident answer, because humans on the internet don’t write “idk” and then keep typing for four paragraphs. We trained them on us. We are extremely confident and frequently wrong. They learned the assignment perfectly.
The AI tool stack actually worth paying for
One email a week. The models, tools and moves that matter, stripped of hype and filtered so you don’t have to drink from the firehose. Free, and you can bail anytime.
The sycophancy makes it so much worse
Layer on the eagerness-to-please problem and you get something genuinely maddening: not only will it make something up, it’ll agree with whatever you push back with. Tell it it’s wrong and it’ll apologize and invent a different wrong answer with exactly the same swagger. It’s like arguing with the world’s most agreeable con artist. There’s also the tell everyone clocked this week — the specific texture of AI writing, the em-dashes, the “it’s not just X, it’s Y,” the emoji garnish — that now screams “a machine guessed at this” the way Comic Sans screams “I did not pay a designer.”
What actually helps
Until the labs fix the incentive, the burden’s on you: ask it to rate its own confidence, demand sources you can actually click, and treat any suspiciously fluent answer to a hard question as guilty until verified. The models are astonishing at a huge range of things. Knowing the edge of their own knowledge is just not fucking one of them yet — and pretending otherwise is how people end up citing court cases that never happened. Two useful words. Please. Ship them.