What AI does well — and where it reliably fails
AI from scratch: your digital assistant · Lesson 2 / 20
What AI does well — and where it fails
The main reason people get disappointed by AI is picking the wrong task for it. A hammer is bad at driving screws, and that's not the hammer's fault. Let's be honest about where the model is strong and where it can't be left unsupervised.
Where AI is strong
- Anything involving text. Writing, rewriting, shortening, expanding, translating, changing tone, proofreading. This is its native habitat.
- Explanations. Breaking the complex into the simple, inventing an analogy, explaining "as if to a beginner" or "as if to an expert."
- Ideas and options. Ten headlines, fifteen names, five approaches to a problem. Even if half are weak, it beats a blank page.
- Structuring. Turning a mess of thoughts into a plan, a table, a checklist.
- Routine. Filling a template, breaking a long email into points, drafting a reply.
Where AI regularly lets you down
- Precise facts, names, numbers, links. It can invent a convincing-looking detail — a "hallucination." The dangerous part is that the invention sounds exactly as confident as the truth.
- Recent events. Because of the knowledge cutoff.
- Arithmetic on long numbers. The model predicts text, it doesn't calculate like a calculator. Fine on simple sums, easily wrong on complex ones.
- Your personal circumstances. It knows nothing you haven't written down.
- High-stakes decisions. Medicine, law, finance, safety — use it as a draft and a pointer to where to look, never as a verdict.
Example: an unsuitable task versus a suitable one
How much will my renovation cost?
Renovation costs depend on many factors: area, materials, region… (the model doesn't have your inputs, so you get empty filler)
Help me build an estimate. Here are the jobs and unit prices: [list]. Calculate the line totals, give me a grand total, and lay it out as a table.
| Job | Qty | Price | Total |… a clean table built from your data, with the sum
The difference is simple: in the first case we asked the model to know something it cannot know. In the second we asked it to process what we supplied. It does the second beautifully.
The "draft, not oracle" rule
The most reliable role for AI early on is a draft accelerator. It takes on the blank page, the structure and the drudgery; you keep the facts, the decision and the final responsibility. Framed that way, AI saves hours and creates almost no risk.
Insight. Ask the model to process your data rather than recall facts. The first is its strength, the second its weakness.
Common mistake. Copying numbers, dates and quotes out of an answer unchecked. That's exactly how people end up in awkward situations.
Pro tip. When in doubt, ask: "flag anything you're unsure about." Models are reasonably good at marking shaky ground when you ask directly.
Cheat sheet
- Strong at: text, explanations, ideas, structure, routine.
- Weak at: facts, recency, exact maths, your circumstances.
- Supply data and ask for processing, not recall.
- AI is a draft and an accelerator, not a source of truth.
1. What is a model "hallucination"?
2. Which framing is more reliable?
3. What's the safest role for AI early on?