What AI can and can’t do

This is what chatbots are best at. Give it your rough points and say who it’s for, and you get a clean draft in seconds.

Try this: Say what tone you want, then edit the draft so it sounds like you. It only knows the dates, names and amounts you give it.

Which one: Any of them. Meta AI in WhatsApp is enough for short messages. ChatGPT, Claude and Gemini are better for long letters.

It’s good at layout and wording, and at matching your experience to a job advert you paste in.

Try this: Give it your real work history and the advert. Read every line afterwards, because it sometimes adds skills or achievements you never mentioned.

Which one: ChatGPT, Claude or Gemini. The free plans are enough.

Paste or upload the text and it pulls out the main points well. It’s much weaker when it has to summarise from memory.

Try this: For anything that matters, like a contract or a policy, ask it to quote the part it’s relying on, then read that part yourself. Summaries sometimes drop the one clause you needed.

Which one: ChatGPT, Claude and Gemini all read PDFs on their free plans, with daily limits. Google’s NotebookLM is free and made for this.

It explains at any level, as many times as you ask, and never loses patience. On well-known subjects like school maths, how a loan works or what inflation is, it’s reliable.

Try this: Say what you already know and ask for an example. Be more careful with niche or very recent topics, where it’s more often wrong.

There’s no wrong answer here, so its weakness with facts doesn’t matter. It gives you twenty options in seconds and you choose.

Try this: Ask for a lot, then ask for more like the ones you liked. Search for a business name before using it. The chatbot doesn’t check whether someone already has it.

For the world’s big languages, chatbots translate about as well as dedicated translation apps, and can keep the tone.

Try this: Contracts, certificates and immigration papers still need a sworn translator. Offices won’t accept an AI translation.

Tell it what you sell, the price and who buys it, and it writes a week of posts in a minute.

Try this: Give it real details: location, prices, what makes you different. Without them every post sounds like every other AI advert.

Which one: Any of them. Meta AI is already in WhatsApp, where you probably post anyway.

It’s good at finding calm, polite wording when you’re too annoyed to. Paste the message and say what outcome you want.

Try this: Take out names and account details before pasting. Check that the reply doesn’t promise anything you can’t deliver.

Organising and tidying what you’ve already written is safe ground. The facts come from you and it supplies the structure.

Try this: Paste everything in, messy as it is, and name the format you want. Check it hasn’t filled gaps with details nobody said.

Describe what you want in plain words and it writes the formula and explains it. You can test it straight away, so mistakes show up fast.

Try this: Tell it which columns hold what. If the formula gives an error, paste the error back and it will fix it.

Which one: ChatGPT, Claude or Gemini. Gemini inside Google Sheets and Copilot inside Excel do it in place, mostly on paid plans.

Give it your notes or a topic and it sets practice questions, marks your answers and explains what you got wrong.

Try this: Use your own notes or past papers so it tests what your course covers. It doesn’t know exactly what ZIMSEC or Cambridge examiners give marks for.

Which one: NotebookLM (free, from Google) builds quizzes from notes you upload. ChatGPT and Gemini both have a study mode.

It saves teachers hours on first drafts: schemes of work, exercises at three difficulty levels, a marking guide.

Try this: Give it the syllabus objective and the grade. Work through its answer key yourself, since it gets the odd answer wrong.

Paste the job advert and it plays the interviewer, asks likely questions and tells you how to improve your answers.

Try this: Use voice mode if your app has it, so you practise speaking and not typing.

Good at breaking a big job into steps: a wedding, a move, a study timetable, a product launch.

Try this: Give it your dates and budget. Treat any prices, opening times or travel times it mentions as guesses until you’ve checked them.

Describe the problem or paste the error message and it walks you through fixes step by step. It’s strong on common phone, Windows and Wi-Fi problems.

Try this: Before any step that resets or deletes something, ask it what the step does.

It’s usually right about well-known things and sometimes confidently wrong, and it sounds exactly the same either way. It can invent names, dates, numbers and quotes.

Try this: Ask it to search the web and show its sources, then open the source. Don’t repeat a figure you haven’t checked.

Which one: ChatGPT, Gemini, Claude and Perplexity all search the web and link their sources. Perplexity is built around that.

A chatbot’s own knowledge stops months before you use it. It only knows today’s exchange rate or fuel price if it searches the web, and for Zimbabwe it often lands on old pages.

Try this: Go to the source for rates, fuel and tariffs: the RBZ, ZERA, the bank or the operator. If you do ask a chatbot, look at the date on the page it quotes.

It’s good at explaining a diagnosis, a test result or a medicine in plain words. It can’t examine you, and it can miss something serious or frighten you over nothing.

Try this: Use it to understand and to prepare questions for the doctor, not to decide. Chest pain, trouble breathing or a sick baby means the clinic, not a chatbot.

It explains what a clause in a lease or contract means in plain language. It knows far less Zimbabwean law than American or British law, and may quote laws that don’t apply here or don’t exist.

Try this: Ask what to look out for and what to ask a lawyer. If money, property or your job is at stake, see a Zimbabwean lawyer.

It can explain what VAT, PAYE and presumptive tax are. Rates, thresholds and deadlines change with every budget, and it often has old ones.

Try this: Get the current rule from ZIMRA’s notices or an accountant. Then use AI to help you understand it.

Its Shona is understandable but often stiff, with odd word choices. Ndebele is weaker and can drift towards Zulu. There’s far less Shona and Ndebele writing online for AI to learn from than English.

Try this: It’s fine for getting the gist of a message. Have a fluent speaker read anything you’ll publish or send to customers.

Which one: No clear winner. We haven’t found a published test comparing them in Shona or Ndebele, so try your text in two of them.

It explains methods well. The main chatbots now run a calculator behind the scenes for big sums, but slips still happen in long working.

Try this: Ask it to show every step and check the final number yourself. For money sums in a business, use a spreadsheet.

It will write one. It will be bland, it may contain made-up references, and your teacher has probably read fifty like it. You also skip the learning the assignment was for.

Try this: Use it to explain the topic, suggest an outline and criticise your own draft. Check your school’s rules: many treat undeclared AI writing as cheating.

A plain chat answer can invent realistic-looking books, articles and court cases. With web search on it finds real ones, but may get wrong what they say.

Try this: Open every source. Confirm it exists and says what the AI claims before you cite it.

Which one: Use something that links its sources: Perplexity, or the deep research mode in ChatGPT or Gemini (limited on free plans). NotebookLM answers only from documents you give it.

It makes striking pictures in seconds. Words inside the picture still come out misspelt at times, it struggles to keep a logo identical from one version to the next, and Zimbabwean faces, places and products often look generic.

Try this: Good for a quick poster or social post. For a logo you’ll print on signs and packaging, have a designer finish it.

Which one: ChatGPT and Gemini make a few images a day on their free plans. Meta AI makes them inside WhatsApp. Claude doesn’t make pictures.

Take a photo and it reads printed text well and can explain a form or a letter. Handwriting and blurry photos lead to mistakes, especially in numbers.

Try this: Check every amount and date against the original. Cover ID and account numbers before you take the photo.

Which one: ChatGPT, Gemini, Claude and Meta AI all accept photos.

Upload a spreadsheet and it can total, chart and spot trends by actually running the numbers. It can also misread a messy sheet: the wrong column, or rows it skipped.

Try this: Compare its totals with ones you already know. Clear column headings help a lot.

Which one: ChatGPT, Claude and Gemini accept spreadsheet uploads. Free plans allow only a few a day.

It produces the structure and the wording fast. The market sizes, costs and prices it fills in are guesses unless you supplied them, and there’s little Zimbabwean data online for it to draw on.

Try this: Bring your own numbers: your costs, your prices, your customers. Replace every figure you didn’t give it.

It writes working code for small jobs and can build a simple site or tool. The tools on this site were built with AI coding tools. Bigger projects still need someone who can tell when it’s wrong.

Try this: Start small and test each piece. Don’t put anything that handles payments or customer data online without a developer looking at it.

Which one: ChatGPT, Claude and Gemini all write code on free plans. The tools that work inside a whole project, like Claude Code, Codex and Cursor, are mostly paid.

It gives quick, useful comments on writing and can mark against a rubric you supply. Its marks aren’t consistent: the same essay can score differently on a second try.

Try this: Use it for comments, and decide the mark yourself. Leave students’ names out of what you paste.

A bot that answers common questions from your price list is doable. It needs setting up on WhatsApp’s business platform, usually by a developer, and it will sometimes tell customers the wrong thing.

Try this: Start with the free quick replies and away messages in the WhatsApp Business app. If you add a bot, keep it to questions with fixed answers and make it easy to reach a person.

It’s a patient sounding board and good at laying out your options. It only knows what you tell it, and it tends to agree with whoever it’s talking to.

Try this: Ask it to argue the other side. Then talk to someone who knows you.

Many people find it a comfort to talk to, especially late at night. It isn’t a trained counsellor, it tends to agree with you, and it can’t act if you’re in danger.

Try this: Use it alongside people, not in place of them. If things are heavy, talk to someone you trust, a clinic, or a service like the Friendship Bench.

Nobody can, and AI can’t either. It has no inside information and no view of the future. An “AI” that claims to predict the rate, forex trades or bets is a scam.

Do this instead: Ask it to explain what moves a rate or a price. That it can do.

“AI trading bots” and “AI investment platforms” that promise steady returns are the old pyramid scheme with a new label. Real AI tools save you time on work you already do. They don’t earn on their own.

Do this instead: If you’re asked to deposit money or recruit others, walk away. Use AI to work faster at something people already pay you for.

A chatbot has no connection to your EcoCash wallet, bank account or ZESA meter. Ask and it will either say so or guess.

Do this instead: Use the official app or USSD code. Never type a PIN or password into a chatbot. A real AI assistant won’t ask for one.

“AI agents” that click through websites for you exist, mostly on paid plans, and they’re slow and make mistakes. They don’t work with EcoCash, ZESA tokens or most Zimbabwean sites.

Do this instead: Let it find and compare the options, then do the paying yourself. Don’t give a chatbot your card or wallet details.

Chatbots know very little about Zimbabwean businesses. They may make up a name or a phone number, or suggest a place that closed years ago.

Do this instead: Ask in a local WhatsApp or Facebook group, or look on Google Maps and read the recent reviews.

AI detectors are unreliable. They flag honest work as AI, especially from people writing in a second language, and miss AI text that’s been lightly edited. A chatbot asked “did you write this?” just guesses.

Do this instead: Teachers: ask the student to explain the work, or have part of it written in class.

It can’t tell reliably. Chatbots guess, and fake pictures, videos and cloned voices now pass a casual look.

Do this instead: Check where it came from and whether a news source you trust has reported it. If a “relative” sends a voice note asking for money, call them back on their usual number.

It can’t examine you, doesn’t know your whole situation, can’t sign anything, and isn’t answerable if it’s wrong.

Do this instead: Use it to prepare. Arriving with clear questions gets you more out of the visit you’re paying for.

Treat a chatbot like a stranger who takes notes. On free plans, what you type may be reviewed or used to train later versions unless you switch that off in settings.

Do this instead: Leave out ID numbers, passwords, customers’ details and company secrets. Look for the data or privacy controls in the app’s settings.