Skip to content
Tim van der Wijst · AI guides
Start hereLesson 1 of 13 — AI from zero

What AI is and isn't

The difference between AI, machine learning and chatbots, what they're good and bad at, and why an AI sometimes tells you nonsense with total confidence.

In short

A chatbot like ChatGPT is generative AI: a large language model that predicts text based on patterns from enormous amounts of text. It is strong at rewriting, summarising and explaining, but weak at exact facts and sources, and can convincingly make things up. Always check facts, numbers and sources.

  • Beginner
  • 6 min
  • free
  • Fact-checked on 1 November 2026

After this lesson you’ll know: how AI, machine learning and a chatbot like ChatGPT relate to each other, what such a chatbot is good and bad at, and why you should never trust it blindly.

One word, four layers

Section titled “One word, four layers”

“AI” is an umbrella term with increasingly specific techniques underneath:

flowchart TB
  subgraph AI["Artificial intelligence (AI)"]
    direction TB
    subgraph ML["Machine learning"]
      direction TB
      subgraph DL["Deep learning"]
        direction TB
        GEN["Generative AI<br/>chatbots (LLMs), images, music"]
      end
    end
  end
  • AI is the broad idea: getting computers to do things that would normally require intelligence.
  • Machine learning is the technique behind most modern AI: a system learns patterns from examples instead of fixed rules, and uses them to predict new cases. Example: a spam filter that learns from millions of emails.
  • Deep learning is machine learning with large neural networks: many layers of calculations stacked one after another.
  • Generative AI creates something new: text, images, music, code. Chatbots like ChatGPT, Claude and Gemini run on an LLM (large language model): a huge model trained on text.

What a chatbot is good at (and what it isn’t)

Section titled “What a chatbot is good at (and what it isn’t)”

Strong at language and patterns, weak at exact, up-to-date facts.

Good at Less good at
Rewriting, shortening, translating text Exact facts, figures and quotes
Summarising a text that you provide Recent events (unless it’s allowed to search)
Brainstorming and developing ideas Citing sources and references that actually exist
Explaining things at your level Doing maths without a calculator tool
Writing and explaining code Knowing what it doesn’t know

Why AI sometimes talks nonsense

Section titled “Why AI sometimes talks nonsense”

A hallucination is text that sounds believable but is factually wrong.

Why? In 2025, researchers at OpenAI explained that during training and testing guessing is rewarded and “I don’t know” isn’t. Just like a student who always fills in something on a multiple-choice exam.

It really does go wrong:

  • Mata v. Avianca (US, 2023). Two lawyers submitted a filing citing court cases that didn’t exist, made up by ChatGPT, including an entirely fabricated case “Varghese v. China Southern Airlines”. One of them later stated: “My reaction was, ChatGPT is finding that case somewhere.” The judge imposed a fine of $5,000.
  • Air Canada (Canada, 2024). The chatbot on the airline’s website gave a customer incorrect information about a fare discount after a death in the family. Air Canada argued that the chatbot was “a separate legal entity”, responsible for its own statements. The tribunal didn’t accept that: the company is responsible for all information on its website, and had to pay.
  • As of 1 November 2026, a public database run by legal researcher Damien Charlotin already counted 2,246 court decisions in which fabricated AI material played a role, and that number is still growing.

🔍 Hallucination detective

An AI gave this answer. What's true, and what's made up? It all sounds equally sure.

  1. In English text, one token is about 4 characters on average.

  2. The free Gemini app can see 1 million tokens at once.

  3. In Mata v. Avianca, lawyers were fined $5,000 for cases ChatGPT made up.

  4. Under the AI Act, chatbots only need to say they're AI from 2028.

  5. A language model learns during your conversation and updates its knowledge.

  6. According to Anthropic, Claude Mythos 5.1 is the same model as Claude Fable 5.1, with different safeguards.

  7. Grok 4.20 was released after Grok 4.5.

  8. The Dutch Data Protection Authority warned against the DeepSeek chatbot in February 2025.

Five misconceptions

Section titled “Five misconceptions”
Misconception How it really works
“It knows things.” It has learnt patterns from text: incomplete, sometimes out of date, sometimes made up.
“It looks everything up on the internet.” Only if that feature is switched on and it uses it. Otherwise it draws on its training.
“It’s always right.” No. Mistakes aren’t an exception but a known, structural problem.
“It remembers everything about me.” The model doesn’t learn from your conversation. Some apps do store notes about you and send them along in later conversations (more in the next lesson).
“It’s just simple autocomplete.” The basic principle is predicting the next bit of text, but research by Anthropic shows that a model can plan ahead, for example by picking a rhyming word before it writes the line.

Checklist

0/4

Your progress is only stored in this browser.

Test yourself

  1. 1What is an LLM?

  2. 2A chatbot cites a source with a title, author and year. What do you do?

  3. 3According to OpenAI researchers, why do language models sometimes make things up?

Next lesson: what an LLM actually does when it gives you an answer.