What will AI do for humanity?
Most likely: make expertise cheap. AI's clearest benefit is putting a competent-if-imperfect doctor, tutor, lawyer, and translator within reach of billions who currently have none. The honest caveat is that measured gains so far are thin — an MIT study of 300+ enterprise AI projects found ~95% delivered no measurable profit impact. Real capability, unproven delivery.
Why — the first-principles explanation
Strip away the drama and AI does one economic thing: it drives the cost of cognitive work toward zero. Not to zero, and not for every kind of work — but sharply down. Everything good or bad follows from that one move, so it's worth asking what it means.
When something expensive becomes cheap, the beneficiaries are the people who couldn't afford it before, not the people who already had it. A partner at a law firm already has research assistants; a cheap AI lawyer improves their margin. A person facing eviction with no lawyer at all goes from zero to something. That gap — between the marginal improvement for the well-served and the categorical change for the unserved — is where nearly all the real upside sits. Roughly half the world has no reliable access to essential health services. An AI that's 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 people actually have.
The same logic runs through science. Research is bottlenecked on expensive cognition — reading everything, generating hypotheses, sifting candidates. Protein structure prediction is the concrete case where this already paid: a problem that took years per protein now takes hours, and that's not a forecast, it happened. Whether that generalizes to drug discovery and materials is genuinely open.
Now the discipline. Adoption is fast and measured returns are not showing up yet. MIT's study of 300+ enterprise AI initiatives found roughly 95% produced no measurable P&L impact on $30–40 billion of spending. That's not proof AI doesn't work — general-purpose technologies historically take a decade-plus to show in the statistics, exactly as electricity did. But it does mean anyone promising you a transformed world on a schedule is guessing. Meanwhile the BIS warns the buildout resembles railway mania and dotcom: real breakthroughs that still drew more capital than returns justified.
So the honest shape: the plausible gift is access — expertise for people who had none. The plausible cost is concentration, since whoever owns the models owns the tollbooth on that expertise. Both are live. Neither is settled.
An example that makes it click
Think about what a printing press did — and who it did it for.
Before 1450, a book cost roughly a year of a scribe's life. Rich people had books. Everyone else didn't. The press didn't make books better; a printed Bible was uglier than a hand-copied one. It made them cheap. And that barely helped the duke who already owned forty books. It utterly transformed the blacksmith's daughter who had never held one.
That's the shape of the bet. AI probably won't give a Harvard professor much they didn't have — they already have researchers. It might give a farmer in a village with no doctor someone to describe the rash to. The press also fueled a century of religious wars, because cheap information isn't automatically good information. Both parts came in the same box. They usually do.
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.
- The IEA projects data centre electricity roughly doubling from about 415 TWh in 2024 to about 945 TWh by 2030, with associated CO2 emissions of 300-500 Mt by 2035 — under 1.5% of energy sector emissions.
- The BIS Annual Economic Report (June 2026) notes the five largest hyperscalers will spend over $1 trillion on AI capex across 2025-2026, outpacing their earnings and free cash flow.
- The BIS compares the AI boom to canal mania, 1840s railway mania, 1920s electrification, and the dotcom boom — all genuine breakthroughs that attracted capital beyond justifiable returns.
- Historical precedent: general-purpose technologies such as electricity took roughly 20-40 years from introduction to measurable aggregate productivity gains, so absence of current returns is weak evidence either way.
- Scale AI's model shows the labor dependency underneath: modern models are trained partly on human-produced preference data generated by large global contractor workforces.
▶ The 60-second explainer (script)
What will AI do for humanity? Strip out the drama and it does one economic thing: it drives the cost of cognitive work down, hard. Everything else follows from that. And here's the key insight — when something expensive gets cheap, the winners aren't the people who already had it. They're the people who didn't. A law firm partner already has research assistants. A cheap AI lawyer improves their margin. Someone facing eviction with no lawyer at all goes from zero to something. That's not an improvement, that's a category change. 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 people actually have. Same story in science. Protein structure prediction went from years per protein to hours. That's not a forecast, that already happened. Now the discipline. Measured returns aren't showing up. 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 in the statistics. But it means anyone promising a transformed world on a schedule is guessing. So: the plausible gift is access. The plausible cost is concentration — whoever owns the models owns the tollbooth on that expertise. Both are live. Neither is settled.
What authoritative sources say
People also ask
What's AI's single biggest potential benefit?
Access to expertise for people who have none. The gain for someone who already has a doctor, lawyer, or tutor is marginal. For the billions who don't, going from nothing to a flawed-but-real advisor is a categorical change — that's where the upside concentrates.
Has AI actually delivered measurable benefits yet?
Thinly. MIT found ~95% of 300+ enterprise AI initiatives showed no measurable P&L impact. Protein structure prediction is a genuine, verified win. Broad economic returns haven't materialized — though general-purpose technologies historically take decades to show up in statistics.
Will AI cure diseases?
It's already accelerated one bottleneck: predicting protein structures dropped from years per protein to hours. Whether that translates into approved drugs is unproven — biology's slow steps are trials and safety, which AI doesn't shortcut.
What's the strongest argument against optimism here?
Concentration. Cheap expertise is only a gift if it's widely distributed. If a handful of firms own the models, they own the tollbooth on cognition itself — and the same technology that democratizes access could centralize power over it.
Is the environmental cost worth it?
The IEA projects data centres reaching ~945 TWh by 2030 with emissions under 1.5% of the energy sector total — real but not dominant. Whether that's worth it depends entirely on benefits that, as of 2026-07, remain largely unmeasured.