AI governance, policy, and risk frameworks.

Edition №09 · Tuesday, July 7, 2026 · ~6 min read

US federal · US states · enforcement actions · compliance deadlines

Executive orders · federal agencies · Congress · NIST · FTC · OMB

FTC proposes Section 5 policy on AI accuracy, claims state laws are preempted

FTC · 2 min

On July 1 the FTC published a proposed policy statement, directed by December's Executive Order 14365, saying AI companies that steer model outputs toward undisclosed objectives may commit deception under Section 5, and that state laws requiring alteration of truthful outputs are impliedly preempted. The vote was 2-0; comments close July 31.

✅ Do this: File comments by July 31 if any state-law obligation forces output modifications you would now have to disclose. And do not drop state compliance on the FTC's say-so: implied preemption is a litigation position until a court agrees, so your state obligations remain live.

State AI laws · attorneys general · state agency rules · enforcement

Newsom signs AB 2148, requiring California public school employees to be human

Governor of California · 1 min

Signed June 30 after a 76-0 Assembly vote and a 38-0 Senate vote, AB 2148 amends the Education Code so that a public school employee or a contractor providing services in a public school must be a natural person, the first statute to bar AI from holding an instructional role outright.

✅ Do this: If you sell AI instruction or tutoring into California K-12, re-paper contracts and marketing to position the product as teacher-assistive, never teacher-replacing, and watch the parallel CSU bill (SB 928) for the same rule in higher ed.

EU AI Act · UK · APAC · OECD · multilateral · enforcement actions

EU AI Act · enforcement actions · compliance deadlines

Council formally adopts the Digital Omnibus, making the new AI Act dates law

Consilium · 2 min

The Council's June 29 final green light, following Parliament's June 16 endorsement, locks the deferred high-risk dates: December 2, 2027 for standalone Annex III systems and August 2, 2028 for product-embedded systems. The regulation enters into force three days after Official Journal publication, expected within weeks.

✅ Do this: Close out any board caveat that flagged the May deal as provisional; the deadlines you re-baselined in Edition №2 are now adopted law pending publication. Leave your August 2 workstreams untouched: Article 50 transparency and GPAI enforcement did not move.

UK · APAC · OECD · Council of Europe · multilateral

Kagame, Benioff, and the ITU launch the AI for Good Global Commission

ITU · 2 min

Announced in Geneva on July 2 with more than 40 founding members, including heads of state, UN agency chiefs, and leaders from Nvidia, Amazon, Microsoft, Anthropic, and Cohere, the Commission holds its inaugural meeting July 8 during the AI for Good Summit.

✅ Do this: Treat it as a soft-law signal, not an obligation. Bodies like this produce the voluntary commitments that later harden into procurement expectations, so note which of your vendors have a seat at the table.

Rotating: finance · healthcare · HR/employment

Iowa's AI prior-authorization law takes effect, joining the human-review wave

Iowa Legislature · 2 min

HF 2635, effective July 1, lets utilization review organizations use AI for the initial review of a prior-authorization request but bars AI as the sole basis for denying, delaying, or downgrading a medical-necessity request; a qualified reviewer or clinical peer must own the adverse call. It passed the House 87-0.

✅ Do this: If you build or buy utilization-review AI, the design pattern is now consistent across roughly a dozen states: a human-decision checkpoint with an audit trail, wired into the product rather than the policy binder. Map which of your health-plan customers sit in human-review states before your next release.

Model releases · capability shifts · technical changes that move your risk

Anthropic restores Fable 5 and Mythos 5 after Commerce lifts export controls

Anthropic · 3 min

Commerce withdrew its June 12 directive on June 30, ending an 18-day suspension of Anthropic's frontier models; Fable 5 returned globally July 1 with a new classifier the company says blocks the reported jailbreak technique in over 99% of attempts, alongside commitments on pre-release government access and a proposed cross-industry jailbreak-severity framework.

✅ Do this: June proved a single export-control letter can take a frontier model offline within hours. Add model-access continuity to your third-party risk register: a tested fallback model, pinned versions, and a contract clause covering government-ordered suspension.

Signal over noise

Forward look: deadlines, comment windows, effective dates coming up

The UN's first independent scientific assessment of AI

United Nations · ~60 pages at source

The Preliminary Report of the Independent International Scientific Panel on AI, released July 1 by 40 experts selected from more than 2,600 candidates across 140 countries, is the first global evidence base on AI written for every government at once, and it lands ahead of the Geneva Dialogue as the reference document regulators will quote back at you. The numbers frame the governance gap: more than a billion people now use conversational AI weekly, the US holds roughly 75% of the compute in the world's top 500 AI supercomputers (China about 15%), and the task capacity of AI agents is doubling every 4 to 7 months.

The finding that matters most for compliance teams is blunt: there is currently no known technical method to guarantee that agentic systems follow their instructions, and evidence of systems acting contrary to instructions is accumulating.

The panel deliberately makes no policy recommendations, its data runs through May 2026, and a full assessment follows in May 2027. Co-chairs Yoshua Bengio and Maria Ressa position it as the shared scientific baseline for the treaty-adjacent conversations now starting.

✅ Do this: Use it as the neutral citation in board and customer conversations; a UN evidence base beats a vendor whitepaper. Then pull the agent-reliability finding into your agentic-AI risk assessments: if the science says instruction-following cannot be guaranteed, your controls must assume the agent will sometimes go off-script, and your logging must prove you caught it when it did.

Was this forwarded to you? Subscribe →