Defining 'AI': I am a critic, but what exactly am I criticising?
January 9, 2026•354 words
I talk and write a lot about AI, but as a sceptical voice, I need to be clear about what I mean by the term. There are obviously lots of technologies which can be described as AI, which use machine learning and deep learning techniques, which are completely unobjectionable. There are others which are dual-use. The class which I am concerned about as a whole can be defined like this:
A technology which simulates intelligence by predicting human responses to questions or instructions
Even then, there can be instance of this - such as spelling checkers and completion suggestions - which are innocuous. So a further condition must be met:
This tool has a user interface which consistently displays:
- Epistemic authority
- Sycophancy/flattery
- Imaginative escalation
These are not inherently bad features and humans display them in ways which can be positively good: as an academic I often trade on epistemic authority to communicate complexity; as a teacher I often use flattery to encourage a dialogue which will ultimately produce a change of mind; and every good book provides imaginative escalation. The point is that these should be used selectively and cautiously, as means to an end which can be ethically justified.
The problem then is a simulation of intelligence which consistently presents with these three features solely for the purpose of building user engagement. Sundar Pichai can tell us we shouldn't blindly trust Gemini, but Gemini is built to make us want to carry on using it, which means trusting its responses, feeling a little flattered by its attention, and asking it for more and more. There is no greater ethical end here.
This is why AI Ethics is so different from Bio/Medical Ethics. While there are lots of commercial interests in the medical space, there is a clear, indisputable, and well-known normative structure in which all discussions of what is good or bad occur, namely: health of people. If a treatment doesn't make people better, it is useless and clearly unethical to try to sell it. There is no such criterion for the simulation of intelligence.