What can AI do?
AI is strongest where answers can be checked automatically — performance on the SWE-bench coding benchmark jumped from about 60% to near 100% in one year. It writes, codes, translates, summarizes, sees, and speaks at useful quality. It cannot reliably know what is true, take responsibility, or judge what matters. 88% of organizations now use it.
Why — the first-principles explanation
The useful way to think about this is not a list of tasks but a single dividing line: can the answer be checked by a machine?
Where it can — code that either compiles and passes tests, math with a provable answer, a translation with a reference — AI improves ferociously fast. The reason is mechanical: if you can grade an answer automatically, you can generate millions of attempts, score them, and train on what worked, with no human in the loop. That is why SWE-bench Verified went from roughly 60% to near 100% in a single year. Nobody had to read those solutions.
Where the answer cannot be automatically checked — is this strategy wise, is this essay good, is this diagnosis right, does this joke land — progress is far slower, because every training signal requires a human to produce it, and humans disagree. This one distinction predicts more about AI capability than any list of features.
That framing also explains the failures people find surprising. A language model generates plausible text, not true text. It has no separate database of facts to check against — truth and fluency are the same operation to it, which is why a confident fabrication looks identical to a correct answer from the inside. This is not a bug awaiting a patch; it is what the mechanism does. Any use where being wrong is expensive needs verification bolted on the outside.
So, concretely, what it does well today: drafting and rewriting text, writing and debugging code, translating, summarizing long documents, transcribing speech, describing images, generating images and short video, extracting structure from mess, answering questions about documents you give it, and tutoring. What it does badly: knowing what is true without a source, arithmetic without a calculator, staying consistent across long tasks, understanding what you actually meant, and anything requiring accountability. It cannot want anything, and it cannot be blamed — which matters, because responsibility cannot be delegated to something that cannot hold it.
Adoption tells you this is real rather than hype: 88% of organizations now use AI, and 70% use generative AI in at least one business function. But note the shape of that adoption — it clusters in drafting, coding, and summarizing. It clusters exactly where a human checks the output before it matters.
An example that makes it click
Think of a brilliant intern who has read essentially everything and remembers none of it precisely. Ask for a first draft, a translation, a summary of forty pages, or a working script, and you get it in seconds — genuinely good.
But this intern has one strange trait: they never say "I don't know." Asked for a court case that supports your argument, they produce one with a citation, a date, a judge. Sometimes the case doesn't exist. And they say it in exactly the same confident voice as when they're right — because to them, producing plausible-sounding text is the job. That's why you use the intern for drafts and never for final answers nobody's going to check.
Key facts
- Performance on SWE-bench Verified, a software-engineering benchmark, rose from roughly 60% to near 100% in a single year (Stanford AI Index 2026).
- Organizational AI adoption reached 88%, with generative AI used in at least one business function at 70% of organizations (Stanford AI Index 2026).
- Four in five university students now use generative AI (Stanford AI Index 2026).
- Language models generate statistically plausible text rather than verified fact — they have no internal fact-checking step separate from generation.
- Pew Research found AI assistants often misrepresent news content, demonstrating the reliability gap in unverified factual output.
- The 2026 AI Index states that AI capabilities are advancing more quickly than our ability to measure and manage them.
▶ The 60-second explainer (script)
What can AI do? Forget the list of features. There's one line that predicts almost everything: can the answer be checked by a machine? Where it can — code that compiles and passes tests, math with a provable answer, translation against a reference — AI improves ferociously fast. And the reason is mechanical. If a computer can grade the answer, you can generate millions of attempts, score them all, and train on what worked. No humans needed. That's why a software engineering benchmark went from about sixty percent to near a hundred in a single year. Nobody read those solutions. Where the answer can't be auto-checked — is this strategy wise, is this essay good, does this joke land — progress crawls. Because every training signal needs a human to make it, and humans disagree. That same line explains the failures that surprise people. A language model generates plausible text, not true text. There's no separate fact database it consults. Truth and fluency are the same operation to it — which is why a confident fabrication looks identical to a correct answer from the inside. That's not a bug waiting for a patch. That's the mechanism. So: it drafts, codes, translates, summarizes, transcribes, sees, tutors. It cannot know what's true without a source, can't do arithmetic without a calculator, can't stay consistent over long tasks, and cannot be blamed. That last one matters — you can't delegate responsibility to something that can't hold it. And look at where adoption actually landed: eighty-eight percent of organizations use AI, clustered in drafting, coding, summarizing. Exactly where a human checks the work before it counts.
What authoritative sources say
People also ask
Can AI think?
It produces outputs that look like reasoning and often function like it. Whether that constitutes thinking is a philosophical question with no test to settle it. For practical purposes, judge it on results, not on the label.
Why does AI make things up?
Because generating plausible text is the whole mechanism. There is no separate step where it checks facts. A confident fabrication and a correct answer are produced identically, which is why they sound the same.
What is AI genuinely best at right now?
Drafting text, writing and debugging code, translating, summarizing long documents, transcribing speech, and extracting structure from messy data — all tasks where a human reviews the output before it matters.
Can AI do math?
Language models are unreliable at raw arithmetic because they predict tokens rather than calculate. Modern systems mostly solve this by calling an actual calculator or code interpreter — which works well.
Will AI replace my job?
The evidence so far shows task-level change rather than wholesale replacement, and credible economists disagree sharply on the eventual scale. Anyone quoting a single confident number is hiding the spread between estimates.
The same question, asked other ways
- What does AI do?1,600/mo
- What is AI used for?1,300/mo
- What can I do with AI?1,000/mo