AI policy in sixteen jurisdictions, as at 3 October 2026
Push is the pressure on a government to act. Bite is what its policy then does. The navigator scores sixteen jurisdictions on both, compares how closely the bite of each jurisdiction's policy follows the push on its government, and shows what the gap says about where policy is likely to go next.
The pressure on a government to act on AI. It comes from rival states, trading partners, industry, voters, incidents and the government's own economic and security aims. Push explains why a policy exists.
Chatbot harms to minors pushed United States states to legislate. Competitiveness concerns and pressure from Washington pushed the EU to delay its high-risk rules.
What the policy does to the firms and people it applies to. It is judged by effect on the reference date and not by legal text: a rule bites when it is in force, specific, monitored and enforced, and a program bites when money or capacity is committed.
A United States export control letter took two frontier models offline within days. The EU's high-risk rules were law and constrained no one.
All push and bite scores are judgements made from the research report, for sixteen jurisdictions as at 3 October 2026. They are not measurements. Government intent is scored. Self-regulation by firms is not.
Compare jurisdictions
Dot area shows private AI investment in 2025 as a share of GDP and dot colour shows legal form. Hover or focus a dot for its scores. The Scoring and evidence tab explains how each score is set.
Question 1 of 2
Each box combines the kind of push on the government with the kind of bite in its policy.
These patterns come from sixteen scores made by judgement, so treat them as indications and not proof.
| How policy bites | Push | Bite | Jurisdictions |
|---|
Question 2 of 2
Each box combines how strong the push on the government is with how strong the bite of its policy is.
These patterns come from sixteen scores made by judgement, so treat them as indications and not proof. Investment figures exist for twelve of the sixteen.
| How strongly policy bites | Push | Bite | Jurisdictions |
|---|
Look ahead
The previous tab shows where each jurisdiction's policy stands in 2026. This page uses the same push and bite scores to look ahead, in three themes: where policy is likely to change next, which jurisdictions face the same pressures, and how policy has moved since 2024 and is scheduled to move to 2028. The scores are judgements that have not been tested against outcomes, so treat what follows as leads and not predictions.
Theme 1 of 3
Policy tends to catch up with the pressure on a government, with a delay. When the pressure and the policy response differ, policy is likely to change to close the difference. Two measures show it. The gap compares their strength, and the lean compares their direction. Choose one to see what it means.
The gap between the pressure a government faces and the strength of its policy response. A positive gap means the pressure is stronger than the policy response. A negative gap means the policy goes further than the pressure calls for. A gap of 10 points or more counts as a clear difference.
Whether policy points the same way as the pressure. A lean near zero means it does. A positive lean means policy leans further towards regulating AI than the pressure on the government asks for, and a negative lean means it leans further towards supporting AI than the pressure asks for.
The gap and lean are calculated from sixteen scores made by judgement, so treat them as indications and not proof.
| Jurisdiction | GapStrength of pressure minus strength of policy | LeanDirection of policy minus direction of pressure | What it suggests |
|---|
Theme 2 of 3
Governments face eight kinds of pressure on AI policy. Grouping the sixteen jurisdictions by their strongest pressures gives four groups.
Pressures to promote AI
Pressures to protect people
A jurisdiction goes in the first group that fits: strategic rivalry rated 2 or 3, then rights tradition rated 2 or 3, then public harm rated 2 or 3, and then the rest. The pressures shown for each group are those that most of its members rate 2 or 3. The ratings are judgements made at a single date, so treat the groups as indications and not proof.
| Pressure | Signs it is rising | Likely policy response | How quickly policy follows |
|---|
How quickly policy follows is a judgement, from the research, of how fast a rise in each pressure has led to binding rules. Fast means months, slow means years, and medium falls between them.
Theme 3 of 3
For each jurisdiction, these charts show how the pressure on its government and its AI policy have changed since 2024, and how they would change to 2028. The projection includes only changes that already have a date, such as a law coming into force, so a jurisdiction with no dated change stays at its 2026 level.
| Jurisdiction | 2024 | 2025 | 2026 | 2027, projected | 2028, projected | What moved, 2024 to 2026 | What is assumed, 2026 to 2028 |
|---|
| Jurisdiction | 2024 | 2025 | 2026 | 2027, projected | 2028, projected | What moved, 2024 to 2026 | What is assumed, 2026 to 2028 |
|---|
The 2024 and 2025 ratings are made now, in hindsight, from dated events in the research, and they are less certain than 2026 because the research concentrated on 2025 and 2026. The 2027 and 2028 ratings are projections, not forecasts of likelihood. Where a law or a notice sets a date, the change is applied on that date, and otherwise the 2026 rating is kept. Projections that depend on a bill passing say so. A cell with a shaded edge is a change of pattern from the year before.
Australia
Where Australian AI policy stands in October 2026, how it got here, and four ways it could change next, with the signs that would show each one.
Now
Since 2024
Next
What happens: Parliament passes the national AI standards in 2027, with the Digital Duty of Care and the second privacy reforms.
Signs to watch: the standards bill is introduced, as planned for early 2027, and the Joint Select Committee on AI reports by 30 November 2026.
What it means: regulation becomes strong for the first time, with duties on businesses that train large AI models or run large data centres. This is the scheduled path, and the chart above assumes it.
What happens: a serious harm leads to a fast rule aimed at that harm, as with the under-16 minimum age.
Signs to watch: findings from the Senate and Joint Select Committee hearings on the Medicare incident, or new findings on AI companions or automated decisions.
What it means: regulation rises quickly but narrowly, enforced by an existing regulator: eSafety for online safety, the OAIC for privacy or the ACCC for consumer law.
What happens: the government pauses or waters down planned rules to attract investment, as it did when the guardrails were set aside in December 2025.
Signs to watch: calls to pause rules, relief for developers, faster data centre approvals and new funding rounds.
What it means: support rises and regulation stalls. Rules dropped this way can be brought back later.
What happens: new laws pass, but regulators do not enforce them.
Signs to watch: cuts or turnover at regulators, enforcement that does not follow a breach, and an AI Safety Institute that stays advisory.
What it means: regulation looks strong in law but stays weak in effect. This framework scores policy by its effect, so it would score as weak.
| Pressure | Rating for Australia | Signs it is rising | Likely policy response |
|---|
Scoring and evidence
Every score is a judgement made from the research report. On the charts, push means the pressure on a government and bite means what its AI policy does. Each section says how a measure is scored, and the full ratings and evidence sit in the table under it.
Each of eight pressures is rated 0 to 3 for how strongly it acted on policy in 2025 and 2026: 0 no evidence, 1 stated only or weak evidence, 2 documented, 3 dominant. Pressure to promote AI is the total of the four promotion pressures, and pressure to protect people is the total of the four protection pressures, each out of 12. On the charts, the direction of push is the protection total as a share of both totals, and the strength of push is both totals as a share of 24.
Pressures to promote AI
Pressures to protect people
| Jurisdiction | Push direction | Push strength | To promote AI | To protect people | Basis for the scores | ||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Rivalry | Sovereignty | Growth | Industry | Rights | Harm | Labour | Control | ||||
Policy has two sides. Regulation is how hard rules bind the actors they apply to. Support for AI is how much the state commits to building and adopting AI. On the charts, the direction of bite is regulation as a share of both, and the strength of bite is the average of the two. The sides describe what the policy does to firms, not why: a constraint can serve a promotion aim, as United States export controls do.
Each component is rated 0 to 3 for the main instruments operating on the reference date. The regulation score is 60 per cent the lowest component and 40 per cent the average, scaled to 100, because the components form a chain and a missing link disables the rest.
| Jurisdiction | Regulation | Form | Duty | Specific | Control | Detect | Sanction | Enforce | Basis for the scores |
|---|
Each component is rated 0 to 3. The support score is the plain average scaled to 100, because these measures substitute for each other and do not form a chain.
Investment. Public funding, state compute and data centre programs.
Adoption drive. Diffusion targets, government adoption, procurement preference and skills programs.
Regulatory relief. Active loosening: delays, exceptions, sandboxes, pre-emption or withdrawal of planned rules.
Strategic backing. State-owned vehicles, export promotion, trade measures and deals with developers.
| Jurisdiction | Support for AI | Investment | Adoption | Relief | Backing | Bite direction | Bite strength | Basis for the scores |
|---|
Dot size shows private AI investment in 2025 as a share of GDP. It is a measured proxy for how much AI weighs in the economy and is not a measure of AI's contribution to output, for which no comparable statistic exists. The area of each dot is proportional to its exact share, so a dot with twice the share has twice the area. Shares run from 0.03 per cent to 5.1 per cent, and California, at the top of that range, is larger than the biggest circle in the legend. Small hollow dots have no figure.
| Jurisdiction | Investment, USD billion | GDP, USD billion | Share of GDP | Note |
|---|
Investment: Stanford AI Index 2026, figure 4.2.8, from Quid data on companies that have raised more than USD 1.5 million. GDP: IMF World Economic Outlook, April 2026, estimates for 2025.
All push and bite scores are judgements made from the research report. They are not measurements, and no published index was used. Dot size is the only measured quantity on the charts.
The investment figures count venture and private funding of AI companies by where the company is based. They favour jurisdictions that host developers and say little about adoption, so China and Japan appear small.
Strength of push partly reflects how much evidence the research found. A factor with no evidence is rated 0, so jurisdictions with less coverage (India, Brazil, the Gulf states, Africa, Montana) and those with little independent analysis of motives (Singapore, Japan) will score low on push strength whether or not the pressure is weak.
The nine boxes on each chart come from cutting each axis at one third and two thirds. The cut points are arbitrary, and a jurisdiction near a line could fall either side with a one-point change in a rating. A jurisdiction exactly on a line is placed in the higher cell. The dotted lines compare two scales built in different ways, so distance from the line is indicative.
The push factors condense ten of the report's fourteen drivers. Institutional structure, state capacity, diffusion and electoral politics shape the form of a policy more than its direction and are left out. Support for AI was not a focus of the research, and compute and industrial policy were covered in less depth than regulation.
Regulation is scored on whichever instruments operate on AI products today. In Australia, the United Kingdom, Singapore, India, Brazil and Canada those are mostly online safety, privacy or intermediary laws. Detection and enforcement ratings rest on limited evidence.
Scores reflect what governments do and intend. Self-regulation by firms, such as a developer pausing training, is not counted as bite, and neither is its absence.
United States states are shown as two examples from opposite ends of the range of approach. Montana's scores are inferred from the text of its Act only.