Opinion / Artificial Intelligence
Everyone’s obsessed with what AI can do. I’m more interested in what it can’t.
There’s a pattern I’ve noticed in how people talk about AI, and it’s starting to bother me.
Every week, a new benchmark gets shattered. Every month, a capability that was “years away” arrives early. The tone of AI coverage has settled into a kind of breathless incrementalism — always zooming in on what just became possible, never pausing to look at the shape of what remains impossible.
I think that’s a mistake. Not just intellectually, but practically. If you want to understand where AI is actually going — and where your own edge might lie — the most useful question isn’t “what can it do now?” It’s “what does it still genuinely struggle with, and why?”
So here are five things that, as of early 2026, AI still can’t do well. I’m not saying “never.” I’m saying “not yet, and the reasons are more interesting than you might think.”
1. Care About the Outcome
AI systems are extraordinarily good at optimising for a target. What they can’t do is genuinely care whether the target was the right one. This sounds abstract, but it plays out in very concrete ways.
A lawyer who cares about their client doesn’t just answer the question asked — they notice the question behind the question. A doctor who cares about their patient doesn’t just process symptoms — they notice the patient’s face. AI, for all its power, is still fundamentally reactive. It responds to what you give it. It doesn’t have skin in the game.
This is not a small gap. Most of the truly valuable work humans do — parenting, leadership, mentorship, great medicine, great teaching — depends not just on capability, but on genuine investment in the outcome. That’s still ours.
2. Navigate a Room
Embodied, real-time social intelligence — reading a room, sensing when someone has checked out of a meeting, knowing that now is not the moment to push — remains stubbornly difficult for AI systems. Language models are trained on text. Text is what people decided to write down. It is, almost by definition, not the full picture.
“Text is what people decided to write down. It is, almost by definition, not the full picture.”
The unspoken negotiation that happens in any human interaction — the pauses, the micro-expressions, the sense that something is off — is still largely invisible to AI. Multimodal models are making inroads, but real-time social navigation at a human level remains a frontier, not a solved problem.
3. Be Consistently Wrong in a Useful Way
This one sounds strange, so let me explain. When a human expert is wrong, their errors are often diagnostic. They reveal assumptions, blind spots, theoretical commitments. You can learn something from the shape of a human’s mistake.
AI errors are different. They’re often confident, plausible, and structurally unrelated to the truth. A hallucination doesn’t reveal a belief — it reveals a statistical pattern gone sideways. There’s no “why” to interrogate in the way there is with human error.
For fields where learning from failure is central — scientific research, strategy, engineering — this matters more than it might seem. Human mistakes are data. AI hallucinations are mostly just noise.
4. Have a Reputation to Protect
Trust, in human society, is built and maintained through accountability. A professional who gives bad advice suffers consequences — reputation damage, lost clients, in some cases legal liability. Those consequences shape behaviour, and that shaping is part of what makes advice trustworthy.
AI has no reputation to protect. It cannot be embarrassed. It won’t lose sleep over a bad call. This is not just a philosophical observation — it has structural implications for how much we should trust AI outputs in high-stakes domains, and how we should design the systems around them.
The humans who sit between AI and its outputs — the ones who sign their name at the bottom, who take the call, who stand in the room — are not redundant. They are the accountability layer. For now, that layer matters enormously.
5. Want Something New
AI can recombine. It can synthesise. It can extrapolate patterns in directions you might not have thought to look. What it cannot do — at least not yet, and perhaps not in any meaningful sense — is want something that isn’t implicit in its training.
The truly disruptive ideas in human history didn’t come from pattern-matching against the existing data. They came from people who were obsessed with a problem nobody else thought was worth solving, or who saw a connection that the prevailing paradigm had made invisible. That kind of directed, motivated originality — the kind that bends the curve of history — still has a distinctly human signature.
Why This Is Actually Good News
I want to be clear: I’m not writing this as a comfort blanket. I’m not trying to reassure you that humans are fine and nothing will change. A lot will change. A lot already has.
But there’s something genuinely useful in mapping the gaps. Because if you know what AI can’t do — really can’t, structurally, not just “not yet” — then you know where to invest. You know which skills are appreciating in value rather than depreciating. You know where to put yourself.
Care. Social intelligence. Accountability. Reputation. Genuine curiosity. These aren’t soft skills. In an AI-saturated economy, they are increasingly the hard ones.
The race isn’t to outrun the machine. It’s to become more deeply, irreducibly human.
Tags: Artificial Intelligence • Opinion • Future of Work • Technology & Society