AI systems now help decide what news you see, whether a loan gets approved, and what a photo appears to show. Understanding where these systems can go wrong is just as important as understanding how they work.
Previous articles in this series touched on two honest limitations of AI systems: they can inherit bias from their training data, and language models can state incorrect information confidently. Those aren't minor footnotes — they're the starting point for an entire field called AI ethics, which asks a harder question than "does it work?": who does it work well for, who might it fail, and what happens when it's wrong?
AI systems have moved from research labs into decisions that affect real people's lives: screening job applications, assisting medical diagnoses, moderating what content is shown or hidden online, and helping decide loan or insurance approvals. As covered earlier in this series, a machine learning model learns patterns from data rather than following explicitly written rules — which means the responsibility for how it behaves is shared between the data it learned from, the people who built it, and the people who deploy it, in ways that are often less visible than a traditional program's rules would be.
As the machine learning article explained, a model learns whatever patterns exist in its training data — including unfair ones, if they're present. This has caused real, documented problems: hiring tools trained on a company's past hiring data have been found downgrading candidates in ways that echoed historical hiring imbalances; facial recognition systems have shown measurably lower accuracy for some demographic groups than others, because those groups were underrepresented in the training images.
A biased AI system usually isn't the result of anyone deliberately programming unfairness — it's the model faithfully learning patterns that were already present in real-world data, then repeating and sometimes amplifying them at scale. This is precisely why the "garbage in, garbage out" principle matters so much more in machine learning than in traditional software.
Training a capable AI model typically requires enormous amounts of data, which raises real privacy questions: Was the data collected with people's knowledge and consent? Could a model's output ever reveal private details from its training data? Systems that personalise content — recommending videos, ads, or products — also generally rely on tracking behaviour over time, which is convenient but means a detailed profile of someone's habits and interests is being built and stored somewhere.
The same generative capability that lets an LLM write a helpful email can also produce convincing but entirely false text, and similar generative techniques can produce fabricated images, audio, or video — often called deepfakes — realistic enough that people can struggle to tell them apart from genuine recordings. This creates a genuinely new challenge: as AI-generated content becomes harder to distinguish from real content, verifying what's authentic becomes more effortful for everyone, not just experts.
Responsibility for an AI system's behaviour is generally shared across several groups, rather than resting on any single one.
| Stakeholder | Typical Responsibility |
|---|---|
| Developers | Choosing representative training data, testing for bias and failure cases before release |
| Companies deploying AI | Being transparent about AI use, providing human review for high-stakes decisions |
| Users | Verifying important outputs, understanding a tool's limitations, using it appropriately |
| Regulators | Setting legal standards for safety, privacy, and accountability |
Governments and organisations worldwide are actively working out how to govern AI, and there is genuine, reasonable disagreement about the right balance. Some argue for strict, precautionary regulation to reduce the risk of harm before problems occur, given how quickly these systems are being deployed. Others argue that heavy-handed rules risk slowing beneficial innovation and that flexible, evolving guidelines paired with industry self-regulation can respond faster than formal law. Different regions have taken noticeably different approaches — this is an active, still-developing area rather than a settled one, and worth following independently as it changes.
A hospital using an AI tool to help flag high-risk patients might find it performs excellently on the population it was tested on, but less reliably for a group underrepresented in its training data — which is exactly why responsible deployments generally keep a human clinician reviewing the AI's suggestions, rather than acting on them automatically.
None of this is a reason to distrust AI outright — the same systems raising these questions are also solving real problems, as earlier articles in this series covered. It's a reason to engage with AI the way you'd engage with any powerful tool: understanding what it's good at, where it can fail, and who's accountable when it does. That combination of curiosity and healthy skepticism is exactly the mindset this entire Computer & AI subject has been building toward.
10 questions. Select an answer for each, then submit to see your score instantly.