What AI is genuinely good and bad at

Ask a chatbot to draft a formal email and it will do it better than most adults, in seconds. Ask it to multiply two six digit numbers, something a two dollar calculator does perfectly, and it may confidently give you a wrong answer. If AI were simply smart or simply dumb, using it well would be easy. It is neither, and the map of where it is strong and weak is stranger than people expect.

Where it is genuinely strong

• Drafting and rewriting: emails, essays, summaries and outlines in any tone, produced in seconds and easy to iterate on.
• Summarising: condensing a long document into key points, because summaries of documents are a pattern it has seen countless times.
• Translation and tone shifts: between languages, and between registers, like turning casual notes into formal prose.
• Code: writing, explaining and debugging programs, since huge amounts of code and discussion of code were in the training data.
• Pattern tasks in general: classifying, extracting, reformatting and finding structure in messy text.

The common thread is that these are all tasks where producing text shaped like good examples from training is exactly what is needed. That is the machine's home ground.

Where it is genuinely weak

• Facts without sources: it generates plausible statements, not verified ones, so specific names, dates, quotes and citations need checking.
• Current events: its training data has a cutoff, so anything after that point is invisible unless the tool searches the web.
• Precise arithmetic: it predicts what an answer should look like rather than calculating, so long calculations often go wrong.
• Knowing what it does not know: it rarely says I am not sure, because hedging was not what fluent training text usually did.

Look at the two lists together and a single rule appears. The model is strong wherever the job is to produce text shaped like good examples, and weak wherever the job needs something the shape of the text cannot supply. A summary, a rewrite, a chunk of code: these are text patterns through and through, so prediction handles them well. A verified fact, today's date, the exact product of two big numbers: these have a correct answer that does not depend on what text usually looks like, so prediction can only approximate it. Even the fourth weakness, rarely admitting doubt, comes from the same place, because confident flowing explanations filled the training data far more often than an honest I am not sure did, so a hedge is simply a less likely continuation and the model reaches for it far less than it should.

Why the arithmetic breaks

Take the arithmetic case slowly, because it is the clearest window into the whole problem. Ask for 738,491 times 6. A calculator applies a procedure, working digit by digit and carrying as it goes, and it is right every time. The model has no such procedure running. It has seen a great deal of text with numbers in it, so it has a feel for what a product of those numbers looks like, roughly how many digits and roughly how it starts, and it generates a plausible looking result. Plausible is not the same as computed. The leading digits might be right and a middle digit wrong, giving an answer that looks convincing and is off by thousands. This is also why many chatbots now quietly hand maths to a real calculator tool in the background, because the makers know the raw model predicts numbers rather than working them out.

The jagged frontier

Researchers describe the boundary of AI ability as a jagged frontier. With a person, skills cluster: someone who writes a brilliant legal summary can safely be assumed to manage basic arithmetic. With a language model, that assumption fails. Capability juts out surprisingly far in some directions and collapses surprisingly close in others, and the strong and weak spots sit side by side with no warning between them. A model can explain university level chemistry and then miscount the letters in a word. The practical skill of using AI is learning where the jagged edge runs for your tasks, and the only way to learn it is to test.

A common misconception

A tempting assumption is that these weaknesses are just early bugs, and the next, bigger model will simply fix them. Some do soften as models improve, but the core ones are built into the approach, not into this year's version. A system whose entire method is predicting likely text will always, at bottom, be guessing at facts rather than checking them, so a bigger version guesses better but still guesses. The reliable fixes come from wiring the model to something else: a search tool for current events, a calculator for arithmetic, a real database for facts that must be exact. Expecting prediction alone to become perfect prediction of the truth misunderstands what the machine is.

You can watch the jagged edge in something close to home. Ask a chatbot to explain the offside rule, or how preferential voting works in an Australian election, and it will usually do a clean job, because both are explained thousands of times in stable, well documented text. Then ask it who won your local council ward last month, or the current price of a specific item at your nearest shop, and it may invent a confident answer, because that information is recent, local and probably never sat in its training data. Same tool, same sentence structure, opposite reliability, and nothing in the wording warns you which one you are getting.

This map is the bridge to the rest of the AI literacy lessons. The case study at the end of this topic traces exactly this split through two answers from one chatbot, one solid and one confidently wrong. And the later lessons on using AI well are really about living on this frontier on purpose: leaning on the model where it is strong, checking it where it is weak, and never letting fluent wording stand in for a fact you have not verified.