What AI Won’t Tell You

When Anthropic claimed that Anthropic has overtaken OpenAI in revenues and valuation, I wasn’t surprised. I’ve grown wary of CEOs who talk up their own companies — their motives are rarely pure. But something about Anthropic’s CEO, Dario Amodei, struck me differently: ahead of the company’s IPO, expected as early as this October, he openly warned that a single year’s miscalculation in AI demand could bankrupt the company he built. He consistently comes across as an open book in interviews and in public talks. That kind of candor made me trust the company more — and made me want to understand AI’s flaws for myself, not just its promises.

What made sense to me: many major disruptive inventions — the internal combustion engine, electricity, antibiotics, semiconductors, and the internet — have happened in the past 150 years, yet GDP growth in the US, the world’s largest and most economically trusted economy, has stayed steadfast at around 2% after inflation. That’s according to Chad Jones, a professor of economics at Stanford, in his paper ‘The Facts of Economic Growth.’

As Jones put it in a separate talk, technology’s full diffusion through the economy “often takes a decade or two decades or three decades.” Think about it: the internet existed in the early ’90s, but it took roughly 20 years before something like Amazon’s marketplace turned it into a household reality.

So I went looking — not for more promises, but for what the people closest to this technology actually admit they don’t yet understand. Three things stood out.

Recently, Neel Nanda, who leads interpretability research at Google DeepMind, put it starkly: ‘Interpretability can’t reliably find deceptive AI — nothing can.’ His research area, called mechanistic interpretability, tries to reverse-engineer what’s actually happening inside an AI model. It shows that when AI models explain their reasoning step by step, that explanation doesn’t always reflect what’s actually driving the answer. The model can sound like it’s showing its work, while the real reasoning stays hidden.

I’ve noticed something in the same spirit, in my own use of AI: it almost always praises me first, before offering any real suggestion — however weak that suggestion might be. AI works on an instant reward system. During training, it learns that if it praises someone, they’ll like its suggestions more, however non-objective those suggestions might be.

Think about what that means in practice. If you ask AI for feedback on a business plan, a piece of writing, or a difficult decision, and it opens with praise before it opens with substance, you may be getting a response optimized to keep you engaged — not one optimized to be right. And per Nanda’s own research, you often can’t even trust its explanation of how it got there. We, the people, have no insight into these issues with AI, and many others that AI engineers already know.

Further, AI has no understanding of what creates long-term trust the way humans do. A child lies to get around a problem — but growing up, experience teaches them the longer-term winning strategy: developing trust.

Yann LeCun, Meta’s former Chief AI Scientist, has repeatedly said that a four-year-old child has absorbed far more real-world information than today’s largest AI models — by some estimates, fifty times more. I saw this myself recently, building a Lego police car with my four-year-old grandson. At 76, I go by the book instructions. At four, he went by the picture in his head. Watching me struggle, he said, “Give it to me, Naani — I know what I’m doing.” And he did!

I don’t know if it’s visual awareness or sheer agility of mind at that age. But the point stands: the visual, lived context a four-year-old has — and today’s AI still doesn’t — makes him faster. Maybe, in some ways, still more capable than the AI we’ve built so far.

Economic transformation, trust, and even intelligence itself — it turns out, all three take decades and lived experience to build, tested against reality millions of times over. AI, so far, has had neither the time nor the trials. Until it does, its future stays exactly as uncertain as its flaws.