On 20 January 2025, the US executive order that required frontier AI developers to report safety testing to the federal government was rescinded. Six months later there was a new national AI Action Plan, a renamed standards body, and a markedly different tone coming out of Washington. Anyone who glanced away from AGI policy news for a quarter missed a genuine reversal, not just a change of press releases.
The beat moves like that: sideways in bursts, then all at once. Announcements pile up faster than anyone can verify them, and the words doing the most work — “frontier”, “systemic risk”, “AGI” itself — mean different things to the people signing the documents. What follows is a map of where the rules actually stand, built from decisions that have been made rather than commitments that have been floated.
Washington swapped precaution for procurement
The 2023 order rested on a simple mechanism: training compute above a threshold triggered reporting obligations. That mechanism is gone. In its place sits an approach that treats AI leadership as an industrial and national-security question first, with safety handled through standards work and voluntary commitments rather than binding rules.
Two administrative moves tell most of the story. The AI Safety Institute became the Center for AI Standards and Innovation, and its remit shifted away from evaluating model risks toward promoting American standards abroad. The Commerce Department’s AI duties now read closer to export promotion than oversight.
States stepped into the vacuum
Congress has not passed comprehensive AI legislation. States have. California’s SB 53, signed in September 2025, requires the largest frontier developers to publish safety frameworks and report critical incidents to the state’s emergency services office within 15 days. Texas passed its own framework earlier in the year, and Colorado’s law, delayed once already, lands in mid-2026.
The result is a patchwork that companies are quietly engineering around. A California-grade disclosure regime is becoming the de facto national baseline, mostly because nobody wants to maintain two compliance tracks.
Brussels is still going, just slower
Europe took the opposite bet: legislate first, argue about interpretation later. The AI Act’s general-purpose model obligations took effect on 2 August 2025, with the Code of Practice published that July as the practical route to compliance. Signatories included Amazon, Google, Microsoft, OpenAI, Anthropic and xAI. Meta declined, calling the process overreach.
The penalties are not decorative — up to 7% of global turnover for prohibited practices, 3% for most other breaches. The systemic-risk threshold sits at 10²⁵ FLOP of training compute, a bar fewer models clear each quarter.
The friction is timing. High-risk obligations were pencilled in for August 2026, and the Commission has since signalled it wants to push them back, an admission that the technical standards those rules depend on aren’t ready. Enforcement capacity is the other soft spot. National market surveillance authorities are being staffed right now, not in 2027, and their budgets are modest.
The measurement problem keeps getting worse
Every serious policy debate about AGI runs into the same wall: nobody agrees on what counts as progress. Benchmarks saturate within months of release. Lab-reported evaluations can’t be independently checked at the frontier. “AGI” has no settled technical definition, which is exactly why it works so well as a contractual term — the 2023 Microsoft–OpenAI agreement reportedly tied it to $100 billion in profit, a financial line rather than a cognitive one.
That definitional fight isn’t academic. The terms written into contracts, charters and statutes determine who gets to deploy what, and the ongoing dispute over how AGI gets defined shapes everything downstream of it.
Some thresholds did move — longer task horizons, better performance on novel research-style problems, cheaper reasoning at scale. Tracking which AGI milestones genuinely shifted rather than merely got talked about is a more useful exercise than reading capability claims at face value.
The policy consequence is that rules get written against proxies — compute, revenue, incident reports — because those are measurable and capability isn’t. A defensible second best. It also means a lab can be fully compliant while fielding a system nobody outside the building really understands.
What companies are doing while they wait for clarity
Frontier labs have converged on a similar playbook: publish a safety framework, commission an external review, hire a policy team, and negotiate the details behind closed doors. That’s not nothing. It’s also uneven, and the gap between the published framework and the internal practice is where most of the risk lives — the parts of responsible AI that teams routinely skip tend to be the unglamorous ones: red-teaming documentation, escalation paths, who has authority to halt a deployment.
One underrated development is incident reporting. California’s 15-day window creates the first structured dataset of serious AI failures in the world. Whatever accumulates there in the next two years will shape federal legislation more than any white paper has.
Four things worth watching over the next twelve months
- Federal pre-emption. A serious push to override state AI laws would reset the compliance map overnight. Watch whether it lands as standalone legislation or gets attached to something must-pass.
- Europe’s high-risk delay. A formal postponement past August 2026 would signal that enforcement ambition is outrunning enforcement capability.
- The incident data. Early California filings will reveal how often frontier systems fail in ways developers consider reportable — and how narrowly they interpret “critical”.
- The next summit cycle. India hosts the follow-up to Paris in early 2026. The Paris summit ended without a joint declaration after the US and UK declined to sign; whether that fracture widens matters for any future international agreement.
The paperwork matters more than it looks
It’s tempting to treat all of this as inside baseball, disconnected from what the models can actually do. The connection is tighter than it appears. Systems that handle multi-step work across unfamiliar domains are already in circulation, and general purpose AI’s real capabilities and real limits set the boundaries of what any regulation is trying to govern.
Get the capability picture wrong — in either direction — and the rules miss. Overestimate and you build a compliance apparatus around systems that don’t exist, which is roughly what happened with early autonomous-vehicle timelines. Underestimate and the first serious incident catches everyone unprepared.
The distinction that keeps getting blurred in policy documents is the one between AGI and what comes after it. Most current frameworks are written for something roughly human-equivalent in scope, while the rhetoric around them often describes something far beyond — the line between AGI and superintelligence is doing a lot of unacknowledged work in these debates.
If you build or buy AI systems, the practical takeaway is narrow but real: know which jurisdiction your deployment touches, keep documentation you could hand to a regulator on short notice, and treat incident reporting as a design requirement rather than an afterthought. The rules are still being written. The organisations that fare best will be the ones already holding the paperwork when they land.