Skip to content
Multigrid

AI governance, policy & society

How the rules around AI are actually built — which institutions have real authority, how each policy instrument bites, and where the evidence is genuinely unsettled.

Most writing about AI policy is advocacy. It argues for an outcome and selects the evidence that supports it, which is a legitimate thing to do and a useless thing to read if what you want is to understand how the machinery works.

These pages take the opposite approach. They describe policy instruments the way the rest of this library describes inference: mechanically. What does a compute threshold actually catch, and what walks past it? What does a provenance manifest prove, and what does it only appear to prove? Who bears the loss when a liability rule changes, and which contract term moves it back? Where a question is settled by evidence, the page says so and names the work. Where it is a disagreement about values, the page says that too, states the strongest version of each side, and leaves the conclusion to you.

Nothing here is legal advice, and nothing here should be relied on as a statement of what any rule currently says. Regulation in this area moves faster than a written page; the structure is durable, the text is not.

Who Governs AI? A Map of the Actual Institutions

The bodies with a claim to govern AI, sorted by the only thing that distinguishes them — what happens if they decide you are doing something wrong.

4 min read

Compute Thresholds in Regulation, Explained

What a training-FLOP threshold actually measures, the arithmetic that turns one into GPU-hours, and the four things it reliably fails to catch.

4 min read

Open Weights vs Closed Models: The Policy Argument

Both cases for and against open model release, stated as their own proponents state them, with the empirical disagreements separated from the value ones.

5 min read

Export Controls and the Global Chip Map

How semiconductor export controls work as a legal instrument — the four keys a licence turns on, why they reach beyond the issuing country, and why performance thresholds invite a design-around loop.

4 min read

AI Safety Institutes and What They Do

The institutional form of a government AI evaluation body: what it can compel, what it can only request, and the routes by which its technical output becomes binding on someone.

4 min read

Model Cards and System Cards: Reading Them Critically

What the model card genre was designed to disclose, what launch documents actually disclose, and how to read one for the absences rather than the contents.

4 min read

Algorithmic Accountability and the Right to Explanation

What explanation duties in law actually require, the scholarly dispute over whether a right to explanation exists at all, and why explanations of model behaviour are hard to make faithful.

4 min read

Bias in Language Models: Measuring It Properly

The families of bias measurement, the documented weaknesses of each, and the formal result that makes an unqualified claim of fairness impossible.

4 min read

AI and Employment: What the Data Shows So Far

The four study designs used to measure AI's effect on work, what each can and cannot establish, and how to tell an exposure estimate from an outcome.

4 min read

Automation and the Reorganisation of Knowledge Work

Why task-level analysis predicts more than job-level analysis, which task properties actually determine automatability, and what shifts inside an organisation when generation gets cheap.

4 min read

AI in Education: Cheating, Tutoring and Assessment

Three problems that get merged into one, and for each assessment redesign option, the specific validity constraint it removes and the cost it adds.

4 min read

Synthetic Media, Deepfakes and Provenance

Two opposite strategies for synthetic media — detect the fake or authenticate the real — and precisely what a provenance manifest does and does not prove.

4 min read

AI Watermarking: Does It Work?

How text and image watermarking actually work, the entropy constraint that limits text watermarks, and the published attack literature including forgery.

5 min read

Detecting AI-Written Text: Why Detectors Fail

The precision arithmetic that governs any detector at classroom or platform scale, why the errors fall systematically on some writers, and what a defensible use looks like.

4 min read

The Energy and Water Cost of Inference

Why published per-query energy figures differ by orders of magnitude, how to build the estimate yourself from throughput and power, and which term dominates.

4 min read

Data Centres, Grids and the AI Build-Out

Why 'where the electricity comes from' has three defensible answers, what actually constrains data centre build-out, and who pays for the connection.

4 min read

Concentration of Power in AI: Who Owns the Stack

The AI stack layer by layer, the structural forces pushing each toward or away from concentration, and what competition law can actually reach.

4 min read

Digital Sovereignty and National AI Strategies

The four different things 'sovereign AI' is used to mean, the threat model each addresses, and the one each conspicuously does not.

4 min read

Liability: Who Is Responsible When AI Causes Harm

How the existing liability doctrines apply to model-caused harm, why the product-versus-service question decides so much, and where contract quietly reallocates the risk.

4 min read

Content Moderation by AI: Scale, Errors and Appeals

The arithmetic that makes moderation errors unavoidable at scale, why the threshold is a value choice, and what procedural fairness would actually require.

5 min read

Other topics