Frontier AI Safety Timeline: 2023-2026

Last verified: September 20, 2026. Frontier AI safety moved from a single voluntary declaration to five overlapping binding statutes in under three years. The Bletchley Declaration in November 2023 opened the period with a non-binding statement signed by 28 countries; by May 2026, California, Colorado, New York, Texas, and South Korea all had their own binding frontier-AI statutes in force or scheduled, several of them layering duties onto the very same handful of developers. This page lists every major milestone in date order — declaration, statute, or report — with a short summary and a link to CASRAI’s own deep-dive guide on that exact topic.

The timeline

1–2 November 2023 — Bletchley Declaration

At the first global AI Safety Summit, held at Bletchley Park in the UK, 28 countries plus the EU signed a joint statement on frontier-AI risk — the first time China joined the US in a multilateral AI-safety text. It set no binding obligations, but it created the mandate (agreed by the same 30 signatories) that later produced the International AI Safety Report. See CASRAI’s Bletchley Declaration vs Seoul Declaration comparison for the full signatory list and what changed by 2024.

1 May 2024 — Utah AI Policy Act takes effect

Utah’s SB 149 took effect, becoming one of the first US state laws to specifically address generative AI — not as a broad risk-management framework, but as a consumer-disclosure statute requiring businesses to tell people when they’re interacting with generative AI in specific regulated contexts. See CASRAI’s Utah AI Policy Act guide for exactly what it requires and who it covers.

21–22 May 2024 — Seoul Declaration and Frontier AI Safety Commitments

At the AI Seoul Summit, 11 governments signed a leaders’ declaration on international cooperation, and — in a separate document — 16 AI companies signed the Frontier AI Safety Commitments, pledging to publish safety frameworks before deploying frontier models. That pledge is the direct precursor to the statutory Frontier AI Framework requirement that California’s SB 53 and New York’s RAISE Act later made mandatory. See CASRAI’s Seoul Frontier AI Safety Commitments guide for the full signatory list.

1 August 2024 – 2 August 2030 — EU AI Act’s staggered phase-in

The EU AI Act entered into force on 1 August 2024, but its obligations apply on a staggered schedule rather than all at once: prohibited-practice bans and AI-literacy duties applied from 2 February 2025; General-Purpose AI model obligations applied from 2 August 2025 (with a 2 August 2027 compliance deadline for models already on the market). The high-risk-system deadlines were revised by the 2026 “AI Omnibus” — high-risk obligations under the Annex III “direct listing” route (biometrics, education, employment, critical infrastructure, and more) now apply from 2 December 2027, rather than the original 2 August 2026 date; the Annex I product-safety route follows on 2 August 2028; and systems used by public authorities have until 2 August 2030. See CASRAI’s EU AI Act high-risk compliance checklist for the current deadlines and what’s due before each one.

29 September 2025 — California SB 53 signed

Governor Newsom signed SB 53, the Transparency in Frontier Artificial Intelligence Act — the first US law to regulate frontier AI developers specifically on the basis of catastrophic risk, rather than on bias, discrimination, or consumer-facing harm. Its core obligations, including a 15-day incident-reporting duty, took effect 1 January 2026. See CASRAI’s California SB 53 explainer for who it covers and what it requires.

18 December 2025 — UK AISI publishes its first Frontier AI Trends Report

The UK AI Security Institute (AISI) published its first Frontier AI Trends Report, aggregating two years of its own government-led evaluations — internal testing of more than 30 frontier AI systems released between 2022 and October 2025 — into a single public analysis. See CASRAI’s UK AISI Frontier AI Trends Report guide for what its evaluations found.

19 December 2025 — New York RAISE Act signed

Governor Hochul signed the original RAISE Act (Responsible AI Safety and Education Act) the day after AISI’s report. A chapter amendment narrowing its thresholds and penalties toward SB 53’s followed — introduced 6 January 2026, passed 11 March 2026, and signed 27 March 2026 — and the law as amended takes effect 1 January 2027, with a tighter 72-hour incident-reporting clock than California’s. See CASRAI’s New York RAISE Act guide for the law as finalized by that amendment.

1 January 2026 — Texas TRAIGA takes effect

The Texas Responsible Artificial Intelligence Governance Act (TRAIGA), enacted as House Bill 149, took effect. Unlike Colorado’s risk-assessment model, TRAIGA doesn’t create a general duty to assess AI systems before deploying them — it bans a short, specific list of AI uses outright and layers narrower disclosure duties on top. See CASRAI’s Texas TRAIGA guide for the full list of banned uses.

January 2026 — Korea AI Basic Act in force

South Korea’s AI Basic Act came into force, making it the most concretely binding AI-specific statute in a major market outside the US, the EU, and China. It requires a mandatory risk assessment for “high-impact” and generative AI systems, and requires developers and deployers to designate a Korea-based local representative. See CASRAI’s Korea AI Basic Act guide for the obligations that apply even to organizations with no Korean office.

3 February 2026 — International AI Safety Report 2026 published

The second full edition of the International AI Safety Report — chaired by Turing Award-winning computer scientist Yoshua Bengio and written by more than 100 independent experts under the mandate the Bletchley Declaration’s 30 signatories agreed in 2023 — was published ahead of that year’s AI summit. See CASRAI’s International AI Safety Report 2026 guide for what the 2026 edition found.

14 May 2026 — Colorado AI Act reenacted as SB 26-189

Colorado’s original AI Act, SB 24-205 (signed 17 May 2024), never actually took effect: an August 2025 special-session bill pushed its start date from 1 February 2026 to 30 June 2026, and before that date arrived, the legislature repealed SB 24-205’s substantive provisions and reenacted them under a new bill, SB 26-189, signed 14 May 2026. Unlike SB 53 or the RAISE Act, Colorado’s law was never a frontier-model statute — it regulates any developer or deployer, of any size, whose automated system influences consequential decisions about individual people. See CASRAI’s Colorado AI Act guide for what the reenacted law actually requires and the enforcement date that applies now.

Why the CASRAI NIKOLAI project tracks this timeline

NIKOLAI is CASRAI’s own frontier-AI-safety dictionary — an independent, unendorsed reference work, not a standard that any lab, evaluator, or regulator has adopted or endorsed. Its N7 (“Incidents”) track exists specifically because incident reporting is where this timeline’s shift from voluntary framework to binding law shows up most concretely: SB 53’s 15-day clock and the RAISE Act’s 72-hour clock, both listed above, don’t share a common definition of what triggers the clock, who receives the report, or how the emergency (24-hour) exception works. NIKOLAI’s N7 element “incident reporting deadline and recipient” records that gap as a rule with four sub-properties — trigger condition, clock length, recipient, and follow-up obligation — because, per NIKOLAI’s own documentation, “no two sources currently share identical timing.” Every row in that element’s crosswalk table, including its reading of Anthropic’s Advanced AI Framework alongside SB 53 and the RAISE Act, is a shadow mapping — CASRAI’s own independent reading of what each source has published — rather than a declaration confirmed by the organization itself; see the full N1–N10 track map for how the other nine tracks organize the rest of this timeline’s vocabulary.

Frequently Asked Questions

Is the Bletchley Declaration still relevant, now that binding laws exist?

Yes, as the origin point rather than as active obligation. Bletchley itself created no binding duties, but the mandate its 30 signatories agreed in 2023 is what produced the International AI Safety Report, and the Frontier AI Safety Commitments signed at the follow-up Seoul Summit are the direct precursor to the statutory Frontier AI Framework requirement now written into SB 53 and the RAISE Act.

Which of these milestones are actually binding law, and which are voluntary?

Bletchley, Seoul, the International AI Safety Report, and the UK AISI Trends Report are declarations, pledges, and research outputs — none create legal obligations. The EU AI Act, California SB 53, the New York RAISE Act, Texas TRAIGA, the Korea AI Basic Act, and the Colorado AI Act are binding statutes.

Why does Utah’s milestone appear before the Seoul Summit, even though Seoul is the bigger event?

This page orders every entry strictly by the date of the specific fact cited — Utah’s law took effect 1 May 2024, three weeks before the Seoul Summit’s 21–22 May 2024 dates. Seoul remains the more consequential milestone for frontier-model governance specifically; the ordering reflects calendar date, not significance.

Whose incident-reporting deadline is shortest?

New York’s RAISE Act requires a report within 72 hours of determination. California’s SB 53 allows 15 days from discovery for most incidents, tightening to 24 hours only where there’s an imminent risk of death or serious injury — the same 24-hour emergency threshold the RAISE Act also applies.

Is the EU AI Act’s 2 August 2030 date final, or could it move again?

It’s the current adopted date, not a proposal — the 2026 “AI Omnibus” is a binding amendment to the Act, not a draft. But the Omnibus itself already shows these dates can move: it pushed the Annex III high-risk deadline from 2 August 2026 to 2 December 2027 after the Act was already in force, so a compliance program should treat “current” as meaning current as of this page’s last-verified date, not permanent.

What’s the single most consequential entry on this list for a compliance program covering multiple US states?

The 27 March 2026 RAISE Act chapter amendment, because it narrowed New York’s thresholds to nearly match SB 53’s — meaning a frontier developer can now draft one safety-framework program for both states, provided its incident-response plan is built around New York’s tighter 72-hour clock rather than California’s 15-day one.