Why is AI good?
AI's strongest case is that it makes expertise cheap for people who had none — translation, tutoring, medical triage, legal help. Its verified wins are real but narrower than the marketing: protein structure prediction dropped from years to hours. Its measured business returns are thin — MIT found ~95% of 300+ enterprise AI pilots showed no profit impact.
Why — the first-principles explanation
AI is good at a specific shape of problem, and knowing the shape tells you where the benefit is real and where it's oversold. The shape: tasks where patterns exist in large amounts of data, where being right most of the time is genuinely useful, and where a human can check the output. Translation fits. Drafting fits. Spotting a tumor in a scan fits — a radiologist reviews it. Deciding whether to fire someone doesn't fit, because there's no verifiable pattern and no honest check.
The economic core is that AI drives the cost of cognitive work down sharply. And the key insight most "AI is good" articles miss: cheapness helps the unserved far more than the served. A corporate lawyer with three paralegals gets a marginal efficiency bump from AI. A person facing eviction with no lawyer at all goes from zero to something. Roughly half the world lacks reliable access to essential health services — an AI meaningfully worse than a good doctor is still infinitely better than no doctor. The comparison that matters isn't AI versus the best human. It's AI versus nothing, which is what most of the world actually has. That's where the honest upside concentrates.
The verified wins deserve to be named separately from the promised ones, because the gap is large. Protein structure prediction genuinely collapsed from years per protein to hours — that happened, it's checkable, it accelerated real biology. Machine translation genuinely lets people read across languages they'll never learn. Speech recognition genuinely gave captions to deaf users. These aren't forecasts.
Now the discipline that makes this page worth reading. Measured returns are not showing up yet. MIT studied 300+ enterprise AI initiatives and found roughly 95% delivered zero measurable P&L impact on $30–40 billion of spending. That's not proof AI doesn't work — electricity took decades to appear in productivity statistics, and general-purpose technologies reliably do this. But it does mean the honest form of "AI is good" is "AI is genuinely capable, its distribution of benefit is plausibly enormous, and its delivered returns so far are unproven." Anyone stating it more confidently than that is ahead of the evidence — in either direction.
An example that makes it click
Think about a pocket calculator, and who it actually helped.
When calculators got cheap in the 1970s, they didn't help the mathematics professor much — he could already do arithmetic, and faster than you'd think. They transformed the shopkeeper, the nurse checking a dosage, the kid who was bright but kept losing track in long division. The gift wasn't to people who had the skill. It was to everyone who didn't, and who had been quietly locked out of anything requiring it.
Calculators also made some people worse at mental arithmetic. That's real, and it happened. But the trade — a small loss for the already-skilled, a huge gain for everyone else — is why nobody argues about calculators anymore. That's the bet with AI, at a much larger scale, with the same structure: the ones who benefit most are the ones who had nothing.
Key facts
- An MIT study of 300+ enterprise AI initiatives (July 2025) found roughly 95% delivered zero measurable P&L impact on $30-40 billion of enterprise generative AI spending in 2024-25.
- AI-driven protein structure prediction reduced a problem that previously took years per protein to hours — a verified, checkable acceleration in biology rather than a forecast.
- The IEA projects data centre electricity roughly doubling from ~415 TWh (2024) to ~945 TWh by 2030, with CO2 emissions of 300-500 Mt by 2035 — under 1.5% of energy sector emissions.
- The BIS Annual Economic Report (June 2026) notes hyperscaler AI capex exceeds $1 trillion across 2025-2026, outpacing earnings and free cash flow.
- General-purpose technologies including electricity historically took roughly 20-40 years from introduction to measurable aggregate productivity gains.
- AI performs best on tasks with abundant pattern data, tolerance for being right most of the time, and a human able to verify output — and worst where no verifiable pattern exists.
▶ The 60-second explainer (script)
Why is AI good? Here's the strongest honest case — and where the evidence actually stops. AI is good at a specific shape of problem: tasks where patterns exist in lots of data, where being right most of the time is useful, and where a human can check the answer. Translation fits. Drafting fits. Spotting a tumor in a scan fits, because a radiologist reviews it. Deciding who to fire doesn't fit — no verifiable pattern, no honest check. The economic core is that AI drives the cost of cognitive work down. And here's what most articles miss: cheapness helps the unserved far more than the served. A corporate lawyer with three paralegals gets a small efficiency bump. Someone facing eviction with no lawyer goes from zero to something. That's not an improvement, that's a category change. Half the world lacks reliable access to essential health services. An AI that's worse than a good doctor is still infinitely better than no doctor. The comparison isn't AI versus the best human. It's AI versus nothing — which is what most of the world has. Now, the verified wins, separate from the promised ones. Protein structure prediction went from years per protein to hours. That happened. It's checkable. Machine translation lets people read across languages they'll never learn. Speech recognition gave deaf users captions. Not forecasts. But here's the discipline. MIT studied over three hundred enterprise AI projects — about ninety-five percent showed no measurable profit impact. That doesn't prove AI fails. Electricity took decades to show up in productivity numbers. But it means the honest sentence is: AI is genuinely capable, its potential distribution of benefit is enormous, and its delivered returns so far are unproven. Anyone more confident than that — in either direction — is ahead of the evidence.
What authoritative sources say
People also ask
What is AI's most clearly proven benefit?
Protein structure prediction — a problem that took years per protein now takes hours. That's verified and checkable, not a projection. Machine translation and speech recognition captions are similarly real, delivered benefits.
Who benefits most from AI?
People who currently have no access to expertise. Someone with a lawyer already gets a marginal gain; someone with no lawyer goes from zero to something. The upside concentrates among the unserved, not the well-served.
If AI is so good, why aren't companies seeing returns?
MIT found ~95% of 300+ enterprise initiatives showed no measurable profit impact. That's genuinely unresolved. General-purpose technologies like electricity took decades to appear in productivity statistics, so it's weak evidence either way — but it's also not something to wave away.
What is AI genuinely bad at?
Anything without a verifiable pattern or a way to check the answer. It's confidently wrong in ways that look identical to being right, which makes it dangerous exactly where you can't evaluate the output yourself.
Is the energy cost worth the benefit?
The IEA projects data centres at 300-500 Mt CO2 by 2035 — under 1.5% of energy-sector emissions. Real but not dominant. Whether it's worth it depends on benefits that, as of 2026-07, remain largely unmeasured at the economy level.
The same question, asked other ways
- Is AI good?880/mo