Two things are true about AI right now, and they refuse to sit in the same room.
In one, AI is astonishing. It passes the bar exam, writes working software in an afternoon, reads a radiology scan, drafts a legal brief. The demos are real and the benchmarks climb every month. In the other, almost none of it is showing up: 95% of enterprise AI projects return no measurable value, and study after study finds little effect on earnings or hours worked.1
Most people assume one of these stories must be a lie—that the skeptics are in denial, or the boosters are selling something. Neither is. Both are true at once. They don't reconcile because the two camps are answering different questions, and once you see which question each is asking, the confusing picture snaps into focus.
Two questions, not one
Arvind Narayanan and Sayash Kapoor, the researchers behind AI as Normal Technology, draw the distinction the rest of the debate keeps missing: capability is not reliability.2
Capability asks: can the system ever do this? Run it a thousand times, and if it succeeds impressively even once, it's capable. That's the right question if you're worried about danger—a model that can write malware or breach a network only has to succeed once to matter, so "a modest chance it works" is exactly the threshold that should alarm you.
Reliability asks the opposite: will it do this almost every time? That's what you care about if you want to hand the system real work. A coworker who is brilliant on Monday and wrong on Tuesday, on the same task, is not someone you can build a process around. For anything approaching autonomy you need success rates near 100%—and, as Narayanan and Kapoor's reliability work shows, that is a far higher bar, one the field is only beginning to measure by borrowing standards from aviation and nuclear safety.3
Here is the punchline: two years of breathtaking capability gains have produced only modest gains in reliability.3 The systems got dramatically better at sometimes doing impressive things, and only a little better at doing them dependably. We flatten both into a single phrase—"how good is AI?"—and then wonder why the demo and the deployment tell different stories.
Why the demo fools us
A demo is an existence proof. "AI can do X" almost always means "AI did X once, impressively, while someone was watching." We hear that as a promise of reliability because that's how normal tools behave: once a calculator adds correctly, it adds correctly every time. AI broke that intuition. It can be brilliant and unreliable in the same breath, and the distance between the two isn't a rounding error—it's the whole point. Getting from 80% to 90% to 99% and on to the last few nines isn't a little more work; each step costs far more than the last, which is why so many pilots that dazzle in the demo quietly stall before production.
There's a reason the capability story is so loud, and it isn't dishonesty. The single most-cited measure of AI progress—METR's chart showing the length of tasks AI can handle roughly doubling every few months—tracks software tasks, at a 50% success rate.4 Software is where the signal is brightest, and that's no accident: code is the one domain with a built-in answer key, where you can test an output, catch the failures, and iterate all the way to reliability—something I've written about as the measurable-output test. The people generating the loudest capability claims live in the one field where the reliability problem is closest to solved. It's understandable that they'd generalize from it. It's also wrong. Most work has no unit test.
What this explains
Once you separate the two questions, the contradictions dissolve. The breathless "AI will replace whole professions" runs on capability—look what it can do. The stubborn "95% get nothing" runs on reliability—look what it dependably does. Both readings are honest; they're measuring different things.
It also explains why one number can mean opposite things. You would never hire a security engineer who caught only half the vulnerabilities—50% is nowhere near good enough to trust. But an AI that introduces a vulnerability even 10% of the time is genuinely dangerous, because that failure only has to slip through once. Same technology, same success rate, opposite verdict—depending on whether you're measuring reliability or risk.
And it's visible wherever companies tried to skip the distinction. Ford has been rehiring veteran "grey beard" engineers because its automated quality-control systems weren't good enough to trust.5 The capability was there. The reliability wasn't.
The useful question
None of this means AI is overhyped junk. Where a mistake is cheap, or caught before it matters—drafting, brainstorming, generating code you'll test anyway—capability is plenty, and the gains are real. The gap only bites when you need the system to run on its own, or to be right nearly every time.
So the question to ask isn't the one everyone asks. "Can AI do this?" almost always resolves to a misleading yes. The better question—the one that tells you where to deploy and where to wait—is: how often does this need to be right, and what does it cost me the time it's wrong? Answer that honestly, task by task, and the fog around AI's "capabilities" clears.
Because the real state of the technology isn't that it's powerful, or that it's disappointing. It's both at once: capable enough to be dangerous, and not yet reliable enough to be useful in most of the places we keep insisting it belongs.5 The demo is not the deliverable. Before you reorganize around what AI can do, ask how often it does it—and what happens the time it doesn't.
Footnotes
-
MIT, "The State of AI in Business 2025" — widely cited for the finding that roughly 95% of enterprise generative-AI pilots produced no measurable P&L impact. ↩
-
Arvind Narayanan and Sayash Kapoor, "AI as Normal Technology," Knight First Amendment Institute (2025): https://knightcolumbia.org/content/ai-as-normal-technology ↩
-
Stephan Rabanser, Sayash Kapoor, Arvind Narayanan, et al., "Towards a Science of AI Agent Reliability" (2026) — decomposes reliability into consistency, robustness, calibration, and safety, and finds capability gains far outpacing reliability gains: https://www.normaltech.ai/p/new-paper-towards-a-science-of-ai (paper: https://arxiv.org/abs/2602.16666) ↩ ↩2
-
METR, "Measuring AI Ability to Complete Long Tasks" (2025) — the 50%-success "time horizon," measured on software tasks: https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/ ↩
-
Financial Times, "The AI Shift: The risk vs. reliability problem" (2026), which frames AI's two questions and reports Ford rehiring experienced engineers after automated quality control fell short. ↩ ↩2