Chapter 09 · Risks & limits

What AI gets wrong — and why that matters

The most important thing you can know as an AI user is not how impressive these systems are. It's where they fail — and why. Many of the failures are not random bugs. They are structural features of how these systems work.

The hallucination problem — demonstrated

Hover over highlighted text to see what's wrong. This is a real category of AI failure — not a rare glitch.

A confident-sounding AI response — with hidden problems
User asked:
"Who invented the World Wide Web and when did they publish their proposal?"
AI responded (hover the highlighted parts):
The World Wide Web was invented by Tim Berners-Lee
✓ Correct
, a British scientist working at CERN
✓ Correct
in Switzerland. He published his foundational proposal in March 1989
⚠️ Actually: he submitted it in March 1989, but it was formally circulated in May 1990. Sounds precise, is slightly off.
. The proposal was titled "Information Management: A Proposal"
✓ Correct
and was initially met with the response from his supervisor: "Vague but exciting."
✓ This quote is real — but notice how confidently the AI mixed correct and slightly incorrect information throughout.
Why this happens:
Language models are trained to produce fluent, plausible text — not to produce only true text. They don't "look things up." They generate responses based on patterns learned during training. A response that sounds authoritative and internally consistent gets generated — regardless of whether each detail is accurate. The danger is that fluency and accuracy look identical from the outside.

When to trust AI — a rough guide

AI reliability varies enormously by task type. This is a rough calibration, not a guarantee.

The five structural limits

It generates false information with complete confidence

Language models are trained to produce fluent, plausible text — not to produce only true text. When they don't have good data on something, they don't say "I don't know." They generate an answer that sounds authoritative and fits the context. The answer may be entirely fabricated. For anything consequential — facts, citations, medical or legal information — verify independently.

It inherits the biases in its training data

AI learns from human-generated data — and human-generated data reflects human history, including its inequities. Systems trained on historical hiring decisions inherit the biases in those decisions. This isn't intentional — the system is learning what was actually in the data. But the consequences are real: AI systems used in hiring, lending, and criminal justice have been shown to systematically disadvantage certain groups.

It has no common sense about the physical world

A language model can write a technically accurate essay on why bridges don't fall down, then fail a question any five-year-old would answer correctly. It has processed text about physics — it hasn't experienced gravity or weight. This produces a peculiar failure mode: confident, well-written answers to questions the system has no actual understanding of.

It can be used to deceive at unprecedented scale

The capabilities that make AI useful — generating fluent text, creating realistic images, cloning voices — also make it the most powerful deception tool ever built. A system that generates a thousand personalized phishing emails in the time it takes to write one, or produces a realistic video of someone saying something they never said, changes the threat landscape in fundamental ways.

It doesn't know what it doesn't know

A chatbot may hedge carefully on a question it actually knows well, and speak with equal confidence on a question it's simply inventing. The absence of "I'm not sure" is not evidence of accuracy. Developing a feel for when a response is likely reliable — versus when it's likely invented — is one of the core skills of AI literacy.

"None of these are reasons to avoid AI. They are reasons to use it the way you'd use any powerful tool — with clear eyes about what it's good for and where it fails."