AI Ethics, Bias & Limitations
Everything in this guide so far explains how AI works. This module is about how it can go wrong — and what to do about it. Understanding these limitations isn't a footnote; it's part of understanding AI at all, because "a model that learns patterns from data" inherits every flaw already present in that data.
Bias
Recall Module 3's warning: garbage in, garbage out. If historical hiring data reflects past discrimination, a model trained to "predict who gets hired" will learn to reproduce that discrimination — not because the algorithm is malicious, but because it faithfully learned the pattern in the data it was given. Well-documented real-world cases include facial recognition systems performing worse on darker skin tones (due to underrepresentation in training data) and recruiting tools downgrading resumes associated with women.
Bias can creep in from the training data (who or what is over/under-represented), from the labels (whose judgment defined "good" vs. "bad"), or from how a system is deployed (used on a population very different from what it was trained on).
Hallucination
As Module 10 explained, a language model generates the statistically most plausible next token — it does not have a built-in fact-checking step. When a model states something false with complete confidence, that's called a hallucination. It happens because fluent-sounding text and factually-correct text are not the same objective, and a model optimized primarily for the former can produce very convincing versions of the latter that are simply wrong. This is precisely why verifying important AI-generated claims against a reliable source matters, especially for anything consequential (medical, legal, financial, or factual claims you plan to rely on).
Other real limitations
- Distribution shift — a model trained on one kind of data can perform poorly when the real world changes or differs from that data (e.g. a fraud model trained before a new scam pattern emerges).
- Lack of true reasoning — as Module 1 noted, narrow AI systems, including today's LLMs, can fail at logic or arithmetic that seems "simpler" than tasks they're impressively good at, because their strength comes from pattern recognition at scale, not from step-by-step logical reasoning the way a person (or a calculator) does it.
- Adversarial examples — tiny, often human-imperceptible changes to an input (a few pixels in an image, a subtly reworded sentence) can sometimes cause a confidently wrong prediction.
- Environmental and economic cost — training large models consumes significant compute, energy, and money, and automation driven by AI has real effects on jobs and industries.
A practical checklist for using AI responsibly
- Verify important claims. Treat AI-generated facts, citations, and numbers as a draft to check, not a final answer, especially for anything you'll act on.
- Consider who's represented in the data. If a system makes decisions about people, ask whether its training data reflects the population it's being used on.
- Keep a human in the loop for consequential decisions. Hiring, lending, medical diagnosis, and legal decisions benefit from AI as an assistant, not an unchecked decision-maker.
- Be transparent about AI use where it affects other people (e.g. disclosing AI-assisted content when relevant, or that a chatbot isn't a human).
- Test on the edge cases, not just the typical cases, before trusting a model in an important workflow.