Start here2 min read
What AI Can and Cannot Do Today
A realistic guide to generative AI in 2026: where it excels, where it is mediocre and where you should not trust it, with concrete examples.
Between those who say AI does everything and those who say it is useless, reality sits in the middle and moves fast. This guide sums up, with examples, what AI assistants do well today, what they do so-so and what you shouldn’t hand over to them. It will help you use them where they shine and avoid surprises.
Where it excels
- Writing and rewriting. Emails, reports, summaries, ads, changes of tone (“warmer”, “more formal”), style edits. Its strongest suit.
- Summarising and extracting. Pulling out the essentials of a long document, a transcribed meeting or an email thread, or turning text into a table.
- Explaining. From “what is inflation” to a rental contract, adapting to your level and answering questions tirelessly.
- Translating and adapting. Good-quality translation between common languages, keeping the tone.
- Coding. Writing, explaining and fixing code is one of the areas where it has advanced most. Coding agents already build whole apps under supervision.
- Brainstorming. Titles, angles, names, interview questions: a tireless thinking partner.
- Reading images and documents. Current models are multimodal: they read screenshots, photos of a whiteboard, charts or scanned PDFs.
Where it is mediocre (use with supervision)
- Very specific facts without a source. Figures, dates, quotes, references or laws: it may get them right or invent them with the same confidence. Web search helps if enabled, but keep checking.
- Long calculations “in its head”. For arithmetic, better that it uses a code tool or a spreadsheet; reasoning models fail less, but are not infallible.
- Very recent events. Its knowledge has a cutoff date; anything later it only knows by searching the web.
- Texts with a distinctive voice. It writes correctly but tends to sound generic. Give it samples of your style and edit afterwards.
- Long tasks without review. Agents can work for hours, but the longer the task, the more checking the result matters.
What you shouldn’t hand over
- Medical, legal or financial decisions without a professional. Use it to understand, prepare questions or summarise, not to decide.
- Anything you can’t verify. If you couldn’t spot a mistake in the answer, don’t treat it as true.
- Personal or confidential data of other people or your company, unless your organisation allows it with suitable tools.
- Checking whether a text was written by AI. Detectors often fail and have led to unfair accusations.
Why it works this way
A language model predicts text from patterns learned from huge amounts of data. That gives it impressive fluency and useful reasoning, but it has no record of what is true: it generates what is plausible. That is why it shines at language and structure and stumbles on exact facts. Tools (web search, code execution, documents you provide) make up for much of that.
A practical rule
Before asking AI for something, ask yourself: would I recognise a bad answer? If yes, go ahead: it will save you time. If not, use it to learn and prepare, but check a reliable source before acting.
Transparency is now also a legal duty: since 2 August 2026, the EU AI Act requires telling people when they are talking to an AI and labelling content such as deepfakes.
Frequently asked questions
Does AI understand what it says?
Not like a person. It predicts text from learned patterns, which lets it reason surprisingly well in many cases, but it has no experience of the world and does not check whether something is true on its own unless it uses tools such as web search.
Will it replace my job?
Today it automates tasks, not whole jobs. People who learn to delegate repetitive tasks to it and review its work usually save time; the risk is higher for jobs made up almost entirely of those tasks.
