Capabilities and limits
A field guide to what today's models are reliably great at, what they're unreliable at, and how to tell which zone you're in.
The fastest way to waste time with AI is to use it for the wrong job. The fastest way to get value is to aim it at the tasks it's reliably good at. This lesson is the field guide.
The green zone: reliably strong
- Transforming text you provide — summarize, rewrite, change tone, translate, structure. When the facts come from your input, the model just reshapes them. Error rate: low.
- Drafting from a clear brief — emails, job descriptions, meeting agendas, project plans. You review and edit; it does the blank-page work.
- Explaining established concepts — 'explain a 401(k) match like I'm a new hire.' Textbook knowledge is deeply represented in training data.
- Brainstorming and critique — generating options, poking holes in a plan, arguing the other side. Volume and perspective, cheaply.
- Extracting structure from mess — pull names, dates, amounts, and action items out of a rambling email thread into a clean table.
The yellow zone: useful but verify
- Specific facts, numbers, citations — the model may blend, misremember, or invent them. Always check anything you'd be embarrassed to have wrong.
- Recent events — models have a training cutoff (the date its training data ends). Some tools bolt on web search; know whether yours did, and check the sources it cites.
- Arithmetic and precise logic over many steps — it usually gets there, but silently dropping a row or sign is a classic failure. Give it a calculator (many tools run code now) or verify.
- Niche domains — the rarer the topic in public writing, the thinner the model's knowledge and the smoother its bluffing.
The red zone: don't delegate
- Decisions with legal, medical, or financial consequences — AI can brief you; a qualified human decides.
- Anything requiring information it doesn't have — it cannot know your Q3 numbers, your org chart, or what your CEO said yesterday unless you paste it in.
- Final accountability — 'the AI wrote it' has never once been accepted as an excuse. You ship it, you own it.
I'm going to describe a task. Tell me honestly: is this the kind of task where a language model is reliably strong, needs verification, or shouldn't be trusted alone? Explain why in two sentences. Task: [describe a real task from your job]
Models are surprisingly good at assessing their own reliability zones when asked directly. Try it with three tasks from your week.
Output quality looks identical across all three zones — same confident tone, same clean formatting. The zone is a property of the task, not the answer. Classify the task before you read the answer, not after.