5. The limits, mistakes and bias of AI

One of the strangest things an AI can do is called hallucination. This does NOT mean the AI sees things that are not there, like in a dream. In AI language, hallucination means the system confidently states something that is completely false — as if it were a fact.

For example, Abebe asks an AI assistant to list five Ethiopian economists who have published books on agriculture. The AI might give five confident, realistic-sounding names — but some or all of them could be invented. The AI does not know it is wrong. It is not lying on purpose; it is generating text that sounds plausible based on patterns it learned.

This is dangerous because the answer looks real. The name, the book title, the university — all sound believable. If Abebe copies that list and sends it to his professor, he may be embarrassed or worse.

Rule of thumb: Never trust an AI's specific facts — names, dates, statistics, quotes, prices — without checking them from a reliable source.

AI assistants like ChatGPT and Claude learned from text that was collected up to a certain date — this is called the training cut-off. After that date, they do not automatically learn new things. They are, in a sense, frozen in time.

This means:

• If you ask about a new law passed in Ethiopia last month, the AI may not know about it.

• If the price of teff or the dollar-to-birr exchange rate has changed recently, the AI will give you an old figure.

• If a well-known person has passed away recently, the AI might still speak of them as living.

Some AI assistants (like Gemini and the newer versions of ChatGPT) can browse the internet to get fresher information. But even then, they can make mistakes — finding the wrong page or misreading what they find.

Whenever you need current, precise information — prices, exchange rates, laws, news — always check a fresh source directly: the Ethiopian National Bank website, a current newspaper like Addis Fortune, or Telebirr's official app.

Bias in AI means the system gives unfair, unequal, or stereotyped results for certain groups of people. This happens because AI learns from human-written text — and human text is full of historical prejudices, stereotypes, and inequalities.

Here are real examples of AI bias:

• Language bias: Many AI systems were trained mostly on English text. Their answers about Ethiopian culture, Amharic proverbs, or local businesses may be less accurate or miss important context.

• Gender bias: An AI asked to imagine 'a successful entrepreneur' may always picture a man. Asked about 'a nurse', it may always picture a woman.

• Geographic bias: AI knows far more about cities like New York or London than about Addis Ababa or Mekelle. Questions about local Ethiopian markets, prices, or customs may get less accurate answers.

• Cultural bias: What is polite, funny, or correct varies across cultures. An AI trained mostly on Western content may give advice that does not fit Ethiopian norms.

Bias does not mean the AI is evil or intentionally unfair. It reflects the patterns in the data it was trained on. But the result can still harm people if they trust AI outputs without thinking critically.

There is a fourth important limit people often forget: AI does not actually understand what it says. It works by predicting which word or phrase is most likely to come next, based on enormous amounts of text it has seen. It has no real feelings, no lived experience, and no common sense the way humans do.

This shows up in several ways:

• AI may give a technically correct answer that completely misses the point of your question.

• AI cannot sense when something is urgent or emotionally sensitive.

• AI may repeat wrong information with the same confident tone as correct information.

• AI does not know what it does not know — it rarely says 'I am guessing here.'

Sara ran a small shop near Bole and asked an AI for advice on pricing injera for a café. The AI gave a detailed, professional-sounding answer about profit margins. But the prices were based on markets in a completely different country. Sara only realised this after she had already printed new menus.

Key lesson: AI is a very fast, very fluent text generator. It is not a wise advisor who knows your situation.

Knowing the limits of AI does not mean you should stop using it. It means you should use it like a smart first draft, not like a final authority. Here are five practical habits for safer, smarter AI use:

1. Ask for sources. Type: 'Where did you get this information? Can you give me a link or a source I can check?' A good AI will admit when it cannot point to a verified source.

2. Test the answer. If the AI says the price of a bag of teff is 850 birr, check a market app or call your local supplier. A quick real-world check takes 30 seconds.

3. Break complex tasks into steps. Instead of asking 'Is this business idea good?', ask 'What are the common risks for a new small café in Addis Ababa?' Specific questions give more useful, checkable answers.

4. Read critically. Does the answer feel too certain? Is it ignoring your specific context? Does something not quite make sense? Trust your own judgement.

5. Know when not to use AI. For medical symptoms, legal disputes, or financial decisions affecting your savings — talk to a real doctor, lawyer, or financial adviser. AI is not a substitute for professional expertise.