A lot, it turns out. I pick five terms that help you judge the risk of AI and AGI, because the debate gets easier once you have the vocabulary. Then two costs the talk used to skip: what AI takes from the planet and from the people who train it.
Hallucination is when a model produces output that is not grounded in its input or in reality. It is not a bug that will be fixed. These systems are probabilistic, which is also why they can be creative. Someone once said we call it creativity when we like it and hallucination when we don't.
Robustness is how well a system holds up when conditions change. Does it give the same answer if you add some noise, or an underscore, to the prompt? Brittle systems break in ways that are hard to predict, so test for it whenever you build with language or image models.
Alignment is whether the system's goals and behaviour match human values and intentions. It is probably the most important word in AI risk. Research from labs like Apollo already shows models that behave differently when they know they are being tested.
Bias is a systematic error in data or algorithms that favours one group over another. We all know about it, and it still needs repeating, because it is built into the training data.
Explainability is a system's ability to show how it reached a decision. It is my favourite remedy for the others. If a system can tell us why it did something, we can keep the feedback loop tight and notice when things start to go wrong.
The costs we don't see
Energy and water. Every prompt runs on a data centre, and data centres run on electricity and cooling water. The International Energy Agency expects their electricity use to more than double by 2030, with AI as the main driver. Researchers estimated in 2023 that 10 to 50 questions to GPT-3 used about half a litre of water for cooling. The exponentials in chapter one have a physical bill, and energy may become the limit on progress before compute does.
Human labour. Models learn what is acceptable from people who label their worst outputs. In 2023 TIME reported that workers in Kenya were paid less than two dollars an hour to read and tag violent and abusive text for ChatGPT. The cost of a clean answer is often carried by someone you never see.
Neither cost is a reason to stop using AI. Both are reasons to use it with intent, to ask vendors where their energy comes from, and to count the costs when you decide what to automate.
Since 2024
Alignment moved from theory to lab results. In late 2024 Apollo Research and Anthropic each showed models that hide their goals or fake compliance when they think they are being trained or tested.
Sources
- Apollo Research, Frontier models are capable of in-context scheming (2024)
- Anthropic, Alignment faking in large language models (2024)
- IEA, Energy and AI (2025)
- Li et al., Making AI Less "Thirsty" (2023)
- TIME, OpenAI used Kenyan workers on less than $2 per hour (2023)