The Signal

No harmonised standards published in Official Journal as of Q1 2026, with high-risk AI obligations due August 2026. CEN-CENELEC drafts in progress but compliance gap remains active, creating regulatory uncertainty for providers. GPAI Code of Practice first release published, due for full finalisation Q2 2026. Watch for CEN-CENELEC updates and potential enforcement delays.

Executive Insight

Items 1–5 · Mainstream High-Impact Developments

Items 6–10 · Underweighted / Asymmetric Signals

Model Frontier

High-risk AI obligations due August 2026; GPAI Code of Practice first release published

ARC-AGI-2 (Static Reasoning)

Human average ~60%. Models above this line are superhuman on this benchmark.

ARC-AGI-3 (Interactive/Agentic)

Humans score 100%. Frontier AI near-zero — reveals the adaptive intelligence gap.

GPQA Diamond (Graduate Science)

Human expert ceiling ~70–80%.

Investment & M&A

Foundation models

AI-enabled telehealth / GLP-1 programs

Agent-based AI systems infrastructure

AI traffic management / critical infrastructure

Program of record designation locks in multi-decade revenue with switching costs that dwarf the contract values. No other commercial AI platform has this status.

Compute supply increasingly locked into bilateral agreements between hyperscalers and chip designers. Concentrates compute access further and creates structural moats for the signatories.

Sector Penetration

UnitedHealth Group (Optum Insight) quietly ramping AI to automate prior authorisation and denial workflows before provider-side AI can catch up on appeals — asymmetric deployment entrenching payer advantages.

Qualified Health $125M Series B (25 March) — Anthropic embedded as infrastructure partner via Anthology Fund; 7% of US hospital revenue covered. eMed $200M at $2B+ (26 March) — AON Consulting-led AI telehealth platform embedding into employer benefits distribution. FDA deregulation: single-recommendation AI devices no longer require FDA review under Commissioner Makary. ARISE Network: 95% of FDA-cleared AI medical devices have never reported a single patient health outcome.

Quinn Emanuel developed AI platform for real-time trial querying of discovery data — AI embedded in the actual trial process, without regulatory scrutiny.

Harvey $200M at $11B — GIC (Singapore sovereign wealth) co-leading signals legal AI as strategic economic infrastructure. 2026 SKILLS Survey: AI in production at 130 largest law firms across 7 substantive use-case categories. Pilot phase over. Legora $550M Series D at $5.55B — US expansion into Houston and Chicago from Swedish base.

Vietnam leads globally at 74% active AI deployment in financial services — regulatory-light jurisdiction generating AI financial services deployment data at pace that will pressure Western regulators.

Finastra: 98% AI adoption in financial services globally. Only 2% report no AI use. Agentic AI in core banking: Eurobank, Wells Fargo, Metro Bank embedded agentic AI in transaction monitoring and compliance. Colorado AI Act enforcement deadline: June 30, 2026.

xAI/SpaceX merger creates single entity combining AI model (Grok), satellite connectivity (Starlink), and launch capacity — an AI-enabled space-based ISR platform controlled by one private actor with direct presidential access, without regulatory oversight framework.

Palantir Maven elevated to department-wide program of record (22 March). OpenAI-Pentagon deal amended (3 March) — surveillance red lines added after backlash. German Army publicly committed to deploying AI for battlefield analysis using Ukrainian combat data. Shield AI $2B at $12.7B — institutional capital (JPMorgan, Blackstone) formalising defence AI as asset class.

Washington Post AI-generated personalised podcast produced more content in three days than in its entire previous history — AI-authored journalism at scale running on a brand trusted for human editorial work.

Reuters Institute: Google search traffic to publishers down ~33% year-over-year. 43% forecast >75% decline within 3 years. Breaking news traffic up 103% — AI-resistant category where editorial judgment has comparative advantage. Muck Rack: 82% of journalists use at least one AI tool; disinformation concern up to 26%.

White House federal preemption recommendation, if enacted, voids most stringent student data protections before they take effect. Ed-tech vendors lobbying for federal preemption gain market access without complying with strictest privacy standards.

FutureEd 2026 tracker: 53 bills across 25 states. Ohio HB 96: mandatory district AI policies by July 1, 2026 for all 608 Ohio school districts. EdSurge: dominant teacher adoption mode is AI for instructional content generation.

PJM Connect-and-Manage service construct means largest AI data centers could be first loads cut during grid stress — an AI model training run being curtailed mid-operation is a new category of operational risk not yet priced in SLAs.

EIA launches pilot survey targeting 196 data center operators (25 March) — mandatory reporting precursor. DHS plans AI-ISAC for critical infrastructure. AMD $100B Meta GPU deal; Deutsche Telekom/Nvidia €1B Munich AI data center opened; Adani $100B AI infrastructure plan.

European & China Watch

AI in Science

AI independently rediscovered the Standard Model of particle physics from raw 1950s–60s data

Neuro-symbolic AI autonomously solved open cosmic string power spectrum problem

AlphaFold Database — protein complexes expansion

1.7 million high-confidence homodimer predictions added (March 2026) — first expansion to include protein complexes. 30M total calculated. 18M lower-confidence homodimers available for bulk download. Heterodimers in analysis pipeline. Changes functional scope from single proteins to how proteins interact — directly relevant to AMR drug discovery.

OpenAI Preparedness (now: Capabilities Reports)

GPT-5.4 Thinking (5 March) is the first general-purpose model rated HIGH in Cybersecurity — 73.33% combined pass rate on CTF, CVE-Bench, Cyber Range. HIGH in Biological & Chemical also. No Critical tier for any model. GPT-5.4 mini: Below High across all categories.

Anthropic RSP v3.0

RSP v3.0 (24 February 2026) active. Current ASL: ASL-3. No tier change. 24 March: minor noncompliance reporting policy update expanding employee reporting channels. Pause commitment removed. ASL-4+ moved to ‘industry recommendations’.

DeepMind programmes

AlphaFold: major protein complex expansion (see threshold events). AlphaGenome: API live (non-commercial), no new window updates. AlphaEvolve: active in production (data center scheduling, TPU design, Gemini training), no new release. WeatherNext 2: live in Google products, no window update. GNoME: automated lab in preparation, no window update.

Proves reward hacking is inescapable structural equilibrium; derives computable distortion index predicting hacking severity before deployment. Evaluation coverage → zero as tool count grows in agentic systems.

Three structural compliance gaps in EU AI Act transparency mandate cannot be addressed through post-hoc labelling before August 2026 deadline.

Risk Indicators: 2028

No harmonised standards in Official Journal ahead of high-risk obligations August 2026.

No harmonised standards in Official Journal ahead of high-risk obligations August 2026.

Phased rollout strains multi-level supervision with AI Office and national authorities.

Phased rollout strains multi-level supervision with AI Office and national authorities.

Risk-based categories create gaps for GPAI and open-source models.

Risk-based categories create gaps for GPAI and open-source models.

National MSAs and EU AI Board face coordination challenges.

National MSAs and EU AI Board face coordination challenges.

Sandboxes mitigate but high-risk rules burden SMEs.

Sandboxes mitigate but high-risk rules burden SMEs.

AI Act extraterritoriality collides with US/China frameworks.

AI Act extraterritoriality collides with US/China frameworks.

Military AI Watch

OpenAI deployed models on Pentagon’s classified network. Amendment added prohibition on surveillance of US persons and explicit ‘human responsibility for use of force’ red line. OpenAI engineers forward-deployed with safety researchers in the loop.

Federal judge found Pentagon’s designation of Anthropic as supply chain risk was ‘classic illegal First Amendment retaliation.’ DoD records showed designation issued because of Anthropic’s ‘hostile manner through the press.’ DC Circuit challenge still pending. Sen. Gillibrand letter reveals DoD rejected vendor request to memorialise restriction on fully autonomous weapons use.

Record budget request. Replicator program (expendable autonomous drones/vessels) received $1B in FY2025. FY2026 NDAA authorised $900.6B total defence spending with AI reporting requirements but no new autonomous weapons restrictions. US and Russia voted against November 2025 UN resolution calling for legally enforceable LAWS agreement.

Law & Guidance

EU AI Act — Layered System

See standards_highlights above.

Country Grid — Law & Standards Status

🟢 Binding law in force  ·  🟡 Law passed/in implementation  ·  🟠 Guidance/soft law only  ·  ⚪ No framework  ·  🆕 New this week  ·  ⚠️ Amendment/enforcement

AI Governance

6th OECD AI in Work conference across exact AGM window. Sessions on agentic AI governance, algorithmic management (89% of firms have at least one governance measure; effectiveness uncertain), AI capability indicators for occupations, and AI in social protection. OECD’s AI Capability Indicators framework — first results presented — is the most credible multilateral data infrastructure for tracking labour market displacement. No new binding multilateral instruments.

First legally binding international AI treaty opened September 2024. GPPi (26 March) confirms it lacks US and China signatures. Structural governance limitation confirmed.

Original deal permitted ‘any lawful use.’ Red lines only added after public backlash via amendment. Anthropic’s attempt to enforce identical red lines resulted in supply chain risk designation and litigation. 37 OpenAI and DeepMind employees filed amicus brief warning punitive measures could chill public discussion of AI risks.

General application date: 2 August 2026. Omnibus aims to delay Annex III high-risk to December 2027. CIOs in regulatory limbo. No first EU AI strict liability case filed. Standards Vacuum means notified bodies cannot certify high-risk systems regardless of Omnibus outcome.

Comment period closed April 3. Grants irrevocable ‘any lawful’ use rights; prohibits inference-level output refusals; bars non-US AI systems. GSA used MAS solicitation refresh not 60-day Federal Register notice — procedural challenge grounds under 41 U.S.C. §1707.

Ethics & Accountability

Stated commitment (surveillance red line, human control of force) materialised only after public pressure required amendment. Original deal’s ‘any lawful use’ standard accepted without stated red lines. No independent audit or enforcement mechanism for compliance with AI safeguards in classified environments. Gap score: HIGH. Classified deployment = no public verification possible.

Unusual direction: Anthropic’s principles were consistent with its actions (refusing to surrender guardrails) and it was penalised for that consistency. The accountability gap belongs to the contracting authority. Court’s preliminary injunction confirms designation was retaliatory. Creates structural deterrent against other labs maintaining safety commitments in government contracts.

ACM USTPC: well-intentioned AI ethics declarations without enforcement have failed to prevent real-world harms. Shift to agentic AI increases friction: ethics frameworks built around AI ‘decisions’ do not map onto AI ‘actions’ (agent sequences), creating responsibility attribution vacuum at the exact moment AI systems are most consequential.

FTC vacated Rytr consent order (mass-produced fake reviews) for insufficient evidence of ‘tangible consumer harm.’ Proactive capability-based enforcement dismantled while AI-enabled fraud scales. Gap: evidence bar now high enough to permit significant AI-enabled harm before enforcement triggers.

Technical Standards

Litigation Tracker

Anthropic v. Department of Defense — preliminary injunction granted (26 March 2026)

California federal judge found Pentagon supply-chain risk designation constituted ‘classic illegal First Amendment retaliation.’ DoD records showed designation issued due to Anthropic’s ‘hostile manner through the press.’ DC Circuit challenge on statutory authority under 10 U.S.C. §3252 still pending.

AI copyright litigation — Meta summary judgment on fair use granted

Federal judge granted Meta summary judgment on fair use grounds in books case. Court found use highly transformative; no evidence of market dilution. OpenAI-trade-secrets lawsuit dismissed for lack of detail. Class certification timelines extend to March 2027.

FTC AI enforcement: Rytr vacatur and deregulatory pivot

FTC vacated 2024 Rytr consent order under Trump AI Action Plan. Section 5 enforcement narrowing to false advertising/AI washing rather than capability-based preemptive prohibition. FTC chatbot investigation for child protection ongoing.

Personnel & Org Watch

Scale-up aligned with Pentagon contract, $852B valuation, and IPO preparation. Acquisition of Promptfoo (AI agent security testing tool) in March.

UK AI Security Institute (AISI)

US NIST CAISI (formerly AI Safety Institute)

EU AI Office