What AI can and can't do

You practiced separating real powers from wishful thinking. Typical can (today’s tools, with caveats): draft and rewrite text, brainstorm options, translate and summarize, generate images or code sketches, spot patterns in large piles of data, personalize recommendations. Typical can’t (even when the demo looks magic): guarantee truth, hold lasting authentic feelings, replace your accountability, know your private life unless you pasted it, reliably invent accurate fresh facts without a source, or silently “know” what happened after its training cutoff without search/tools.

Run every shiny claim through three filters:

1. Is it pattern work? “Rewrite this in a friendlier tone” → usually yes. “Will this stock hit $200 Tuesday?” → not a honest can.
2. Does truth matter legally or medically? Summary of a public article → maybe help. “Is this mole cancer?” → can’t be your doctor.
3. Is the output checkable? Subject lines you can skim → fine. A fake citation you can’t open → danger zone.

Concrete pair: AI can turn bullet notes into three wedding toast drafts. It can’t know Uncle Ray’s tenderness — and it might invent a childhood memory that never happened. You keep the editor’s pen.

• “Just let AI handle the report” — draft structure (can) vs submit unaudited numbers (can’t).
• An ad promises “AI that never hallucinates” — sales, not physics.
• Before a long prompt: park final calls, secrets, and live facts on the can’t column.

“Can’t” means “don’t treat as authority,” not “never useful.” “Can” means “in the right lane, with your review.” Browsing tools change the edges; the principle stays: check outputs that matter.

Redraw a mini board on paper: three cans, three can’ts, using your life (job, school, home). Compare it to the playground — any surprises?