AI limits, risks and regulation
Hallucinations, bias, privacy, copyright, energy — and what the EU AI Act provides.
本页目录
1. Errors and hallucinations
Plausible text, not necessarily accurate
A language model produces plausible text, not necessarily accurate text. Real cases illustrate this: lawyers presented a New York court with fake court decisions invented by ChatGPT (Mata v. Avianca, 2023, $5,000 fine); a Canadian airline had to compensate a customer misled by its chatbot (Air Canada, February 2024); a report billed at $440,000 to the Australian government contained nonexistent references (2025).
According to Stanford's AI Index 2026, the number of documented AI-related incidents rose from 233 in 2024 to 362 in 2025.
In the Mata v. Avianca case (2023), what was the problem?
Air Canada was able to avoid all responsibility by explaining that the error came from its chatbot, not a human employee.
The number of documented AI incidents tracked by the AI Index decreased between 2024 and 2025.
2. Bias and discrimination
A reflection of the data, not an opinion
An AI learns from data produced by humans, with their inequalities. It can therefore reproduce them, or even amplify them. Bias isn't an AI "opinion": it reflects its data and how the system was designed.
Three documented cases
| Domain | What happened |
|---|---|
| Recruiting | Amazon scrapped a CV-screening tool in 2018 that disadvantaged women's applications (Reuters) |
| Facial recognition | The "Gender Shades" study (2018) showed higher error rates for darker-skinned women |
| Justice | The COMPAS recidivism-risk tool was challenged by a 2016 ProPublica investigation over uneven errors by ethnicity |
Match each documented case to its domain
What did the "Gender Shades" study (Buolamwini & Gebru, 2018) show?
Bias in an AI is a form of personal opinion the system forms freely.
3. Privacy, copyright, environment
Privacy
What you write to an assistant can be stored, reviewed by moderation teams, or used to train future models. Don't enter sensitive data, and check the privacy settings.
Copyright
Models are trained on huge corpora that include copyrighted works, challenged in court (Getty Images v. Stability AI in 2023, The New York Times v. OpenAI and Microsoft in December 2023). The law is evolving fast and differs by country.
Energy
According to the International Energy Agency, data centers consumed about 415 TWh in 2024 (nearly 1.5% of global electricity), with consumption that could reach about 945 TWh by 2030.
According to the IEA, roughly what share of global electricity did data centers consume in 2024?
Which 2023 lawsuit illustrates tensions around copyright and AI?
It's wise to never enter banking details or medical records into an AI assistant.
4. Regulation: the EU AI Act
The core principle
The European Union adopted the first major general legal framework for AI: the AI Act, which entered into force on August 1, 2024. Its principle: the riskier an AI use, the stricter the rules.
The 4 risk levels
| Level | Rule | Example |
|---|---|---|
| Unacceptable | Banned | Social scoring of citizens |
| High | Allowed under strict conditions | Recruiting, credit, education |
| Limited | Transparency obligations | Chatbots, deepfakes |
| Minimal | No specific obligation | Spam filters, video game AI |
The AI Act's application timeline has already been revised once (some "high-risk" obligations postponed to December 2027) and may change further. Check the European Commission's official pages for the current version. This isn't legal advice.
Sort each use case by its AI Act risk level
What is the core principle of the EU AI Act?
When did the AI Act enter into force?
5. Good habits
5 habits that reduce the risks
Verify — Check facts, figures and quotes against a reliable source.
Protect — Don't share sensitive data, check the privacy settings.
Stay in control — Use AI as an assistant, not a decision-maker, especially in health, law or finance.
Be transparent — Disclose AI use when it's expected or required.
Keep learning — Understanding how AI works helps you spot its limits.
Put these 5 habits back in the order presented
In which domains should you especially "stay in control" rather than let AI decide?
Boss challenge: which AI Act risk level?
Sort each AI use case by its risk level under the European regulation.