What AI still can't do in 2026 — a working list from someone who tries constantly
The can-do lists are everywhere. After a year of pushing these tools at the edges of my actual work, here's the list of failures that keeps repeating.
The capability lists write themselves — every launch post is one. I'm more interested in the other list: the things I keep trying because they should work, and keep failing in the same ways. This is that list, from a year of using these tools for real work, not demos. It's shorter than it was in 2024. It's not empty.
One: knowing when it doesn't know. My email AI answered a client's compliance question this year with total confidence and total wrongness — plausible structure, invented regulation number. The tell was only in the citation. Models are better calibrated than they were, but 'confidently wrong on facts I can't easily verify' is still the failure mode that makes me keep every tool on a leash. Until a model says 'I'm not sure' at the rate humans are uncertain, verification stays a job.
Two: holding taste across a long project. Ask for a landing page draft and it's good. Ask it to keep the voice consistent across forty pages of a site, and drift creeps in — page thirty has the vocabulary of page three but not the ear. I fix this with style guides and examples, and it works about 85% of the way. The last 15% is the part I'd call taste, and it's still mine to enforce.
Three: stopping. Not in the science-fiction sense — in the deadline sense. Agents that run my errands do not distinguish between 'the task took 20 minutes' and 'the task deserved 20 minutes.' I've watched one spend forty-five minutes polishing a summary nobody was going to read again. Judgment about effort allocation — what's worth doing well — remains stubbornly human, and it's arguably the most valuable judgment there is.
Four, the quiet one: caring. Every tool now drafts, summarizes, generates, automates. None of them noticed that a client's 'just a quick question' email this spring was actually the first message of them losing a major account. I read it that way because I know them, because their tone changed six months ago, because it matters to me. The list of things AI can't do will keep shrinking. That one — the noticing, the stakes, the why-bother — is the one worth watching, because it's the one that makes the rest of the work worth doing.