What would be an appropriate task for using generative AI?
Generative AI fits tasks where output is easy to check and mistakes are cheap: drafting emails, summarizing documents you have, brainstorming, rewriting tone, writing code you'll test, translating rough text. It fits badly where being wrong is expensive and verification is hard — medical, legal, or financial decisions, and anything requiring facts you can't confirm.
Why — the first-principles explanation
The right test isn't "can AI do this?" — it's "can I tell quickly if it did it badly?" Generative AI produces plausible output. Plausible and correct overlap most of the time, and the gap is invisible from the text alone. So the value of any AI task depends entirely on how cheaply you can close that gap.
That gives a clean rule. Appropriate tasks have a cheap verification step. Code is ideal: you run it, and it either works or doesn't. Summarizing a document you already have is ideal: the source is right there. Drafting an email is ideal: you'll read it before sending. Brainstorming is ideal: you're picking from options, not accepting one.
Inappropriate tasks invert this. Ask for facts you don't already know and can't easily check, and you get fluent output with no error signal — the worst possible combination. This is why AI-drafted legal briefs keep producing sanctions: the citations look perfectly formatted, and checking each one costs more than writing the brief. NIST published a Generative AI Profile in July 2024 precisely because these risks — confabulation, data leakage, and synthetic content — are structural, not incidental.
There's a second axis: sensitivity. Anything you paste into a consumer AI tool may be stored and, depending on the terms, used for training. Client data, medical records, unreleased financials, and passwords fail this test regardless of how good the output would be.
So: cheap to verify + low stakes + not confidential = green light.
An example that makes it click
Think of generative AI as a fast, confident intern who has read everything and remembers none of it precisely. Give that intern a stack of papers and ask for a one-page summary — great, because you can spot-check it against the stack. Ask them to draft a thank-you note — great, you'll read it anyway. Ask them to brainstorm 20 names for a product — great, you're picking, not trusting.
Now ask the same intern, from memory, for the exact citation of a 1987 tax ruling. They'll give you one. It will have a volume number, a page number, and a confident tone. It may not exist. The intern isn't lying — they're pattern-matching what a citation looks like. The task was wrong, not the intern.
How to do it
- Ask: can I check this answer in under a minute? If no, it's a poor AI task.
- Ask: what happens if it's silently wrong? If someone gets hurt, sued, or misdiagnosed, don't delegate it.
- Ask: does the input contain confidential, personal, or regulated data? If yes, don't paste it into a consumer tool.
- Give the AI the source material instead of relying on its memory — summarizing a document you provide is far safer than asking it to recall facts.
- Treat every output as a draft. Verify names, numbers, dates, quotes, and citations against a primary source before use.
- Keep a human decision-maker for anything with legal, medical, financial, or safety consequences.
Key facts
- NIST released the Generative AI Profile (NIST-AI-600-1) on July 26, 2024, identifying risks unique to generative AI including confabulation and synthetic content.
- Generative AI produces output by predicting likely next tokens; it has no built-in step that verifies factual accuracy.
- The EU AI Act, in force since 2024, prohibits certain AI uses outright (such as social scoring) and classifies others — including employment and credit decisions — as high-risk with mandatory human oversight.
- EU obligations for general-purpose AI models became applicable August 2, 2025.
- Tasks with automatic verification — code that compiles and passes tests — carry the lowest risk, because errors surface immediately.
▶ The 60-second explainer (script)
Here's the question people get wrong. They ask, can AI do this task? Wrong question. The right one is: can I tell quickly if it did it badly? Because generative AI produces plausible output. Plausible and correct overlap most of the time — and the gap is invisible from the text. It looks identical either way. So the whole value depends on how cheaply you can close that gap. That gives you a clean rule. Good AI tasks have a cheap verification step. Code is perfect: you run it, it works or it doesn't. Summarizing a document you already have is perfect: the source is right there in front of you. Drafting an email is perfect: you'll read it before you hit send. Brainstorming is perfect: you're choosing from options, not swallowing one. Bad tasks invert this. Ask for a fact you don't know and can't check, and you get fluent output with no error signal. That's the worst possible combination. It's exactly why AI-written legal briefs keep getting lawyers sanctioned. The fake citations are perfectly formatted. Checking each one costs more than writing the brief yourself. NIST published a whole Generative AI Profile in July 2024 because these risks are structural — not bugs someone will patch. There's one more filter. Sensitivity. Whatever you paste into a consumer AI tool may be stored, and depending on the terms, used for training. Client data, medical records, unreleased numbers, passwords — those fail no matter how good the output would be. So the rule is three words long. Cheap to verify. Low stakes. Not confidential. Hit all three, go ahead.
What authoritative sources say
People also ask
What are the best everyday tasks for generative AI?
Drafting and rewriting text, summarizing documents you supply, brainstorming options, explaining unfamiliar concepts, translating rough text, and writing code you'll run and test.
What should I never use generative AI for?
Final medical, legal, or financial decisions; anything where you can't verify the facts; and anything involving confidential, personal, or regulated data pasted into a consumer tool.
Is it OK to use AI for work emails?
Generally yes — you read the draft before sending, so verification is nearly free. Avoid pasting in confidential details, and check your employer's AI policy first.
Why is coding considered a good AI task?
Because verification is automatic. Code either compiles and passes your tests or it doesn't, so errors surface immediately instead of hiding inside fluent prose.
The same question, asked other ways
- Which kinds of tasks count as generative AI tasks?2,900/mo