A short reading list for policy readers on AI sovereignty in 2026

The rule. Every entry is a primary source or a peer-reviewed publication. No newspaper op-eds. No podcast episodes without a linked transcript. No think-tank white papers without a named author and an institutional footer.

On hardware level enforcement

Petrie and colleagues, Embedded Off-Switches for AI Compute, arXiv:2509.07637 (September 2025). Petrie, Aarne, Ammann, Dalrymple, FlexHEG: tamper-proof guarantee processors for AI chips (ARIA-commissioned), arXiv:2506.15093 (April 2025). Odeh, US Patent Application 2026/0010411 and 2026/0010780, hardware-level ethics enforcement (filed July 2025, published January 2026). Read in that order. Each establishes a different mechanical vector for putting safety architecture into silicon.

On the export-control frame

US Commerce/BIS actions on Fable 5 and Mythos 5, 12 June 2026 (Harvard Law Review blog for the commentary; primary orders on the BIS website). CSA whitepaper on the same. The Reuters coverage of 12 June 2026. Read the primary orders first; the commentary is compact and helpful only after you have seen the actual text.

On EU sovereignty policy

European Commission communication on the Tech Sovereignty Package, 3 June 2026, including Chips Act 2.0 and the CADA framework. Virkkunen's public statements on the same date. The primary source is the Commission's communication; the CNBC coverage of 3 June 2026 has the "we want to be sure nobody has a kill switch" quote in context.

On G7 and trusted-partners framing

The G7 Évian communiqué of June 2026. Macron's intervention, primary transcript through the Élysée and secondary through broadbandbreakfast.com and CNBC. Read Macron's intervention before reading the commentary on it; the commentary has consistently misquoted the specific line.

On formal insufficiency of external evaluation

Gumbau Mezquita, The Undecidability of AGI Alignment, arXiv:2606.28639 (26 June 2026). SHA-256 8e4483b1287146da6d725e6078ed4bf708be602f4bd2751b9d0e040a45b1c9a7. The proof is short; the paper is longer. Read the theorem statement and the trilemma before the discussion.

On cognitive-diversity-as-protective-factor

Hernández-Espinosa, Abrahão, Witkowski, Zenil, PNAS Nexus 5(4) pgag076 (14 April 2026); the earlier arXiv preprint at 2505.02581 for the version circulating from May 2025. The peer-reviewed paper is the citation; the preprint is the earlier date.

On what a governance instrument would look like

Paper X of the ARC/Eden programme (OSF DOI 10.17605/OSF.IO/6C5XB), for the beta over k stability condition. The public research evidence spine at /research/evidence-spine.html records the kill conditions. Read the two together; the paper without the dashboard is theory, and the dashboard without the paper is a list.

What the reader keeps

Eight entries. Each is a primary source with a named author and a stable identifier. If you read only three, read Petrie 2025, the Commission Tech Sovereignty communication of 3 June 2026, and Gumbau Mezquita 2026. Those three will let you sit in any 2026 policy conversation on this topic with something specific to say.