Reading and writing stopped being special skills a long time ago; they became the floor. AI literacy is going the same way, which is awkward, because most people still think it means something it doesn't.
It does not mean knowing how models are built. You don't need linear algebra to be literate any more than you needed to study printing presses to be literate in the ordinary sense. AI literacy is knowing what these tools do well, knowing what they do badly, and having a habit for telling the difference before the difference tells on you.
And it's becoming a workplace basic the way spreadsheet fluency did. Employers now screen for it at the application stage, managers assume it in the same breath as email, and the gap between people who have it and people who don't is quietly becoming a gap in who gets the interesting work. The good news is that the skill is smaller and more learnable than the marketing suggests. Here's what's actually in it.
What models do well, and what they don't
The "well" list is short and consistent: drafting, summarizing, transforming, brainstorming. Turning a rough paragraph into a clean one. Pulling the three decisions out of a six-page thread. Converting notes into a table. Giving you twenty starting ideas when you had none. These are tasks where a plausible draft is most of the value and a human eye finishes the job.
The "badly" list matters more. Models are bad at guaranteed truth — they will produce a confident sentence with nothing behind it. They're bad at judgment with consequences: hiring calls, medical decisions, anything where being wrong costs someone something real. And they're bad at anything that needs accountability, because a model can't hold responsibility. It can assist the person who does, but it can't be the person who does.
The underemphasized half: output review
Most AI literacy advice stops at prompting. The other half is reviewing, and it's the half that keeps you employed.
Concretely, reviewing means checking claims instead of assuming them — if a summary cites a number, that number has a source or it doesn't. It means asking for assumptions: "what did you take for granted here?" surfaces the places a draft went off the rails. And it means knowing when a human expert has to sign off. A first draft from a model is fine for the internal memo. The client-facing contract, the published figure, the decision that affects someone's job — those get expert eyes, every time.
Building it in a week
This is a one-week project, not a semester:
- Take one free course. Elements of AI or AI for Everyone, both free to audit, will do. Not for the certificate — for the vocabulary.
- Do five everyday tasks with a model: summarize a long email, rewrite a clumsy paragraph, brainstorm options for a real decision, compare two documents, build a checklist from a procedure you know.
- Write one reusable prompt for a task you actually do weekly, with context and constraints in it, and keep it.
- Build the verification habit: for every output that matters, check one claim against a source. That's it. One check, every time.
The half-life problem
Here's the honest caveat: literacy has a half-life. The specific tool you learn this month will look different in a year, and the prompt that works today will need rewriting. What outlives any of it is the habit of verifying — the reflex that treats every output as a claim to be checked rather than a fact to be copied.
That reflex is the actual skill. Tools are seasonal; skepticism is durable. Which means the best time to build the habit was last year, and the second-best time is this week.
We built After AI around the same question from the other side — which parts of a job machines can reach and which they can't, with every figure cited to a public source.
No comments yet