The generative turn

Lesson 3 of 5 in What Is AI? From Rules to Reasoning Machines.

Until around 2020, AI governance mostly meant governing predictors — systems that scored, ranked, and classified. Then generative AI arrived: systems that produce text, images, audio, video, and code. It did not just add a new product category; it broke the assumptions most governance frameworks were built on.

At the centre sits the foundation model: one enormous model, trained once on a vast slice of the internet, then adapted — via fine-tuning, via prompts, via APIs — into thousands of downstream products. One model's flaws become everyone's flaws. Regulators responded by inventing a new legal object: the EU regulates general-purpose AI models directly, separately from the systems built on them.

How a large language model actually works is disarmingly simple to state: given a sequence of text, predict the next token. That's the whole training objective, repeated across trillions of words. Everything an large language model appears to do — answer, summarise, translate, write code — emerges from that objective plus post-training alignment (RLHF) that tunes it toward outputs people rate as helpful.

Hold onto the governance consequence: an LLM models plausible language, not verified fact. When it states something false with perfect fluency — an invented court case, a fabricated citation — it is not malfunctioning. It is doing exactly what it was trained to do: produce likely-sounding text. This is why hallucination is managed as a permanent risk (grounding, verification layers, disclosure, use-case limits), never dismissed as a bug awaiting a patch.

Generative AI also created new governance categories that now run through every jurisdiction you will study:

• Synthetic media and deepfakes — hence disclosure and labeling laws (EU AI Act Art 50, China's labeling rules, election-deepfake statutes).
• Training-data provenance — whose text and images went in, under what right? The copyright fights and data-governance duties land here.
• Compute as a proxy for capability — the EU presumes systemic risk (GPAI) for models trained above 10²⁵ FLOPs; a number of laws now trigger on training compute.
• Release strategy — open-weight models democratize access and can't be recalled; laws split on how to treat them.

Key terms: foundation model, large language model, hallucination, general-purpose AI, deepfake, open-weight model

Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.