EU AI Act: first systemic-risk GPAI evaluations due September 15

The week of September 10–17, 2026 sits inside the first real enforcement cycle of the EU AI Act's GPAI regime. Sequence of events (FACT unless labelled):

• 31 July 2026 (announcement, pre-window context): Commission press release "Commission starts enforcing AI Act rules and new transparency requirements" — as of 2 August 2026 the Commission (via the AI Office) can exercise its GPAI enforcement powers: request information and documentation (Article 91), conduct evaluations of GPAI models (Article 92), request measures including risk-mitigation and market-recall steps (Article 93), and impose fines of up to 3% of global annual turnover or €15 million, whichever is higher (Article 101). Algorithmic-transparency rules (Article 50: chatbots, deepfakes) became enforceable on the same date.
• 1 September 2026: the Commission sent its first information requests to 30+ AI companies, in two streams: (1) safety- and security-related requests under Article 91 (systemic-risk models: evaluations, risk mitigation, incident reporting, cybersecurity) and (2) copyright/transparency requests under Article 53 (training-content summaries, copyright compliance policy). (COMMISSION ACTION per law-firm and policy-blog reporting — Allegiance Law, Magica, EU Perspectives; the requests themselves were announced by the Commission, i.e., a company/institutional claim, echoed by secondary sources.)
• 8 September 2026: TechTimes reports OpenAI filed the first serious-incident report under Article 55(1)(c) with the AI Office, with OpenAI's chief scientist reportedly admitting a monitoring gap in incident detection. (SECONDARY REPORTING per TechTimes; not independently confirmed with the AI Office. INTERPRETATION: the first public exercise of the GPAI serious-incident-reporting pipeline.)
• 10 September 2026: POLITICO Europe covers the EU's response to frontier-AI "extinction warnings", including the AI Office's contracting of evaluation capacity with METR (Model Evaluation and Threat Research) — the office is building the technical muscle to use its new Article 92 evaluation power credibly.
• 15 September 2026 (the in-window event): per Enterprise DNA (8 Sep 2026) and the September press wave, providers of systemic-risk GPAI models (presumed over the ~10^25 FLOP training-compute threshold) were due to submit their first-round systemic-risk model evaluations to the AI Office — red-teaming/adversarial-testing methodology and results (Art 55(1)(a)), systemic-risk assessments (Art 55(1)(b)), plus the related transparency package: energy-consumption disclosures and copyright training-summary template compliance (Art 53). On the same date the first compliance inspection wave of high-risk AI systems began across the Union — targeting HR/resume screening, retail banking/credit scoring, and healthcare/AI triage — co-led by the French CNIL, German BfDI and Spanish AESIA together with the AI Office, spanning the 24 Member State market surveillance authorities (MSAs).
• 16 September 2026: Lawfare (Eliška Andrš) publishes "You Don't Have to Sell It to Be Bound by It — GPAI and the EU AI Act", a governance analysis arguing Chapter V binds providers of systemic-risk GPAI models even where the model is deployed internally rather than "placed on the market" — significant because GPAI obligations are typically framed around market placement, yet Articles 51/55 attach systemic-risk duties to the model's high-impact capabilities regardless of how it is made available.

The Act-vs-implemented deadline distinction (FACT + INTERPRETATION): Article 113(b) made Chapter V (Articles 51–56) applicable from 2 August 2025 — so model evaluations, adversarial testing, serious-incident reporting and cybersecurity were already legal obligations before September 2026. What changed in the current window is that the enforcement machinery became operative (2 August 2026 powers), the Commission began issuing Article 91 requests (1 September 2026), and the 15 September milestone is the first concrete administrative submission date for evaluation documentation inside that machinery. No calendar date for evaluation submissions appears in the Act itself; "September 15" is the AI Office's implemented first-milestone date as reported by the press. Any statement that "the AI Act requires evaluations by September 15" as a statutory deadline is therefore an oversimplification of a real operational deadline.

• First enforceable frontier-model compliance milestone in the world (discovery's core claim): earlier deadlines (prohibitions Feb 2025; GPAI obligations Aug 2025) had no operational supervision behind them; September 2026 is the first time frontier labs face a concrete, inspector-backed submission date for evaluation evidence, backed by a functional fining power (Art 101).
• Hands-on supervision of frontier labs became real: the AI Office now runs an actual enforcement cycle — information requests (Art 91), evaluations (Art 92), requested measures (Art 93) — against the operators of the most capable models, with METR-style evaluation capacity contracted in (POLITICO Europe).
• Adversarial testing moved from research practice to regulated deliverable: red-teaming results are now documentation a lab must produce on a schedule, to a standard the regulator can interrogate — not just a safety-benchmark scorecard.
• Inspection wave for high-risk AI systems began in parallel: HR, retail banking and healthcare deployers (not just model providers) now face MSA inspections requiring evidence of AI Act compliance documentation — the first sectoral enforcement of the deployer-side regime (postponed Annex III obligations to 2 Dec 2027 do not stop inspections of systems whose obligations applied from Aug 2025/2026).
• Copyright/transparency enforcement started: first Article 53 information requests (training-content summaries, copyright policy) pulled frontier providers into the copyright pipeline the Act created — a separate fight from safety evaluation.

• Before: Chapter V rights and duties existed on paper since 2 Aug 2025; the GPAI Code of Practice (assessed adequate; signatories incl. OpenAI, Anthropic, Google, Meta, Microsoft) was the voluntary vehicle; the AI Office could recommend but not compel; no information requests, no inspections, no fines for GPAI providers (Art 101 excluded from the 2 Aug 2025 penalties date and became operable with the general application from 2 Aug 2026); serious-incident reports were a contractual, not enforceable, expectation.
• Change (2 Aug–15 Sep 2026): enforcement powers live (2 Aug); first Article 91 information requests to 30+ companies (1 Sep); first Article 55(1)(c) serious-incident report filed (8 Sep); systemic-risk evaluation submissions due (15 Sep); first MSA inspection wave in HR/banking/healthcare begins (15 Sep); evaluation capacity being contracted (METR).
• After: frontier labs operate under a standing EU supervision cycle — submissions, evaluations, follow-up requests, and the credible prospect of Article 101 fines (€15M/3% worldwide turnover) for intentional/negligent failure; evaluation evidence becomes a regulated input into both EU enforcement and the labs' own public safety narratives; deployments of high-risk AI in EU HR, banking and healthcare face inspector scrutiny; the precedent exists for the rest of the world's regulators (US, UK, Canada, Japan) to benchmark against.

Legal mechanics (FACT, per Regulation (EU) 2024/1689 as amended by the Digital Omnibus, Regulation (EU) 2026/1744):

1. Classification (Arts 51–52): a GPAI model is classified as presenting systemic risk either by Commission decision or — presumptively — where cumulative training compute exceeds 10^25 FLOP (Art 52; criteria in Annex XIII; Scientific Panel can alert the Commission under Art 90). Designation is by implementing act; providers may rebut with documented justifications (Art 51(2)).
2. Obligations (Art 55(1)(a)–(d)): providers of systemic-risk GPAI models must, in addition to Arts 53–54: (a) perform model evaluation in accordance with state-of-the-art standardised protocols and tools, including conducting and documenting adversarial testing, to identify and mitigate systemic risks; (b) assess and mitigate possible systemic risks at Union level, including their sources, from development, placing on the market, or use; (c) track, document and report serious incidents and corrective measures without undue delay to the AI Office (and national authorities as appropriate); (d) ensure adequate cybersecurity for the model and its physical infrastructure.
3. Enforcement powers (Arts 88–94): the AI Office supervises GPAI providers (Art 88); monitoring actions (Art 89); power to request documentation and information (Art 91); power to conduct evaluations of the model, including with independent experts / third-party evaluation support such as METR (Art 92); power to request measures, including risk mitigation and market recall (Art 93); procedural rights of providers (Art 94).
4. Fines (Art 101): up to €15 million or 3% of total worldwide annual turnover, whichever is higher, when the Commission finds intentional or negligent infringement of GPAI provisions, failure to comply with Art 91 requests or supply of incorrect/incomplete/misleading information, failure to comply with Art 93 measures, or failure to make the model available for an Art 92 evaluation. Court of Justice retains unlimited jurisdiction (Art 101(5)). Per the Commission's own framing and Praxikon's reading, Article 101 became operable on 2 August 2026 (excluded from the earlier Chapter XII application date).
5. Compliance routes (Arts 55(2), 56): adherence to the GPAI Code of Practice assessed as adequate demonstrates compliance (not a full presumption of conformity) until harmonised standards are published; non-signatories must show alternative adequate means (e.g., a gap analysis vs the Code).
6. Timing (Art 113; Omnibus amendments): in force 1 Aug 2024; prohibitions from 2 Feb 2025; Chapter V GPAI from 2 Aug 2025; most remaining provisions (incl. GPAI enforcement powers) from 2 Aug 2026; new Omnibus-added Article 5 prohibitions (sexual imagery/CSAM-related) from 2 Dec 2026; Annex III high-risk obligations deferred to 2 Dec 2027 and Annex I to 2 Aug 2028 by the Omnibus — Chapter V GPAI untouched by the deferral; models on the market before 2 Aug 2025 must fully comply with Chapter V by 2 August 2027 (Art 111(3)).
7. The Sep 15 milestone: reported by Enterprise DNA (8 Sep 2026) and the September press wave as the first submission deadline for systemic-risk evaluation documentation (adversarial-testing methodology + results under Art 55(1)(a), systemic-risk assessment under 55(1)(b), plus energy-consumption and copyright training-summary transparency under Art 53) — an implementation date set inside the AI Office's supervision agenda rather than a date found in the Regulation (INTERPRETATION; the source basis is press reporting, not a directly verified AI Office notice).

• The first enforceable frontier-model compliance deadline in the world (discovery's why_it_matters): California's SB 53/AB 54-style laws and voluntary commitments set precedents, but no other jurisdiction had, by Sep 2026, combined a hard model-evaluation obligation with information powers, an inspection apparatus and a turnover-linked fine. The EU is defining what "regulated frontier AI" means operationally.
• Evaluation evidence is now a regulated asset: adversarial-testing outputs move from safety-team internal documents into regulator-submitted records with legal consequences — reshaping how labs resource red-teaming.
• A working model for supervision: the AI Office–METR contract (POLITICO Europe) shows the EU buying third-party evaluation capacity, effectively outsourcing the "ARIA-style" technical scrutiny the Act's authors knew the office could not build in-house — a model other regulators will copy.
• Copyright enforcement enters the frontier: Article 53 training-content-summary requests put EU copyright law into the frontier-training pipeline — the second front of this enforcement cycle alongside safety.
• Signal to enterprises and deployers: inspection wave = deployer-side diligence becomes enforceable now, in the highest-visibility sectors (HR, banking, healthcare) — not in 2028.

• The AI Office can compel evaluation documentation, run its own model evaluations (Art 92) with third-party support (METR), and order risk-mitigation measures (Art 93) — a full supervisory loop over the most capable models.
• Serious-incident transparency: Article 55(1)(c) reports create a regulator-facing record of frontier-model incidents (first exercised by OpenAI on 8 Sep 2026 per TechTimes).
• Turnover-proportional fines make compliance economically non-optional for the largest labs.
• Downstream deployers get an authoritative compliance signal to anchor procurement and risk due diligence (who has submitted, what gaps the AI Office flags).
• Non-EU regulators gain a live experiment on which to calibrate their own regimes (US federal preemption fights, UK's pro-innovation posture, Japan's soft-law approach).

1. Which providers were (a) presumed systemic-risk by the 10^25 FLOP threshold and (b) formally designated by implementing act — and which actually submitted evaluation documentation by Sep 15, 2026?
2. What exactly did the Sep 15 submission package require, and is that requirement documented in an AI Office notice, a Code of Practice annex, or only in press reporting?
3. Will the AI Office publish evaluation outcomes or a compliance register, or stay behind Article 78 confidentiality?
4. When will the first Article 92 evaluation run, with which external evaluator (METR or another), and on which model?
5. Will the first Article 101 fine land in 2026–27, and on which ground (infringement, info-request failure, requested-measure failure, or denial of Art 92 access)?
6. Does the internal-deployment reading (Lawfare) survive judicial contact — i.e., are internally-deployed systemic models bound by Chapter V?
7. How will FLOP-threshold disputes be adjudicated (training-compute definitions, model clustering, rebuttal evidence)?
8. Which sectors get inspected second (after HR/banking/healthcare), and what enforcement findings will the 24 MSAs actually surface?
9. How will the Code of Practice be updated (Art 56(8)) now that enforcement is live, and will harmonised standards for GPAI evaluation emerge before the next milestone?
10. What will the pre-2-Aug-2025 grandfathered models (full Chapter V compliance due 2 Aug 2027) do differently versus new releases?

• Late Sep–Oct 2026 (near-term, FACT-adjacent/prediction): AI Office review of Sep 15 submissions; first documented Art 92 evaluations likely in Q4 2026 (with METR or similar); follow-up Article 93 requests where evaluations find unmitigated systemic risks; MSA inspection feedback in HR/banking/healthcare flowing to the Board and Safeguard procedures as needed.
• 2 Dec 2026: new Article 5 prohibited practices (sexual-imagery/CSAM-related, added by the Omnibus) become applicable — the next hard date after the Sep milestone.
• 2026–27 enforcement escalations (PREDICTION): first Article 101 fines before end-2027 (likely on Art 91 request failures or Art 92 access denials first, since those are easiest to prove); one or more legal challenges to FLOP presumptions at the General Court; a visible AI Office clash with a non-signatory provider over "alternative adequate means".
• 2 Aug 2027: full Chapter V compliance for models placed on the market before 2 Aug 2025 (Art 111(3)) — a second, broader compliance cliff hitting older flagship models.
• 2 Dec 2027 / 2 Aug 2028: Annex III and Annex I high-risk obligations apply (post-Omnibus dates); deployer-side enforcement in full swing; harmonised standards for GPAI evaluation likely consolidated by then.
• 2027+ (PREDICTION): Code of Practice updated (Art 56(8)) with enforcement-learned refinements; other jurisdictions mirror or differentiate from the EU's evaluation-first model; the Sep-2026 wave is retrospectively seen as the moment EU frontier-AI enforcement became operational rather than symbolic.

September 15, 2026 is the day the EU AI Act stopped being a compliance exercise and became an enforcement cycle. The headline — "first systemic-risk GPAI evaluations due" — is directionally true and globally significant, but the honest version contains two corrections the reporting mostly elides: the Act itself never sets a September 15 calendar date (the obligations have bound providers since August 2025; the deadline is the AI Office's implemented first-milestone inside a machinery that only became operational on 2 August 2026), and the actual enforcement teeth — information powers, regulator-run evaluations, requested measures, and €15M/3%-of-turnover fines — are all from Article 101-inflected Articles 88–94 that are being exercised for the very first time. Watch three things: whether the AI Office converts its milestone into published, verifiable facts (it has not yet); whether the first Article 92 evaluation and the first Article 101 fine are credible enough to matter; and whether the scope fight — does "placing on the market" even matter for systemic-risk duties, per Lawfare's analysis — lands in Luxembourg. In one sentence: the world's first enforceable frontier-model compliance deadline is real, it is operational, and the next twelve months will decide whether it becomes the global template or a cautionary tale in transatlantic divergence.

Evidence discipline summary: FACT (statutory provisions and dates as enacted — Art 51/52/53/54/55/56, 88–94, 101, 113 of Regulation (EU) 2024/1689; 2 Aug 2025 / 2 Aug 2026 / 2 Aug 2027 / 2 Dec 2026 / 2 Dec 2027 / 2 Aug 2028 dates; Digital Omnibus Regulation (EU) 2026/1744 deferrals; Art 101 fines 3%/€15M; Commission enforcement powers from 2 Aug 2026 per official press release); COMPANY CLAIM (Commission announcement of info requests to 30+ companies, echoed by law-firm summaries; OpenAI chief-scientist monitoring admission as reported by TechTimes); INDEPENDENT EVIDENCE (Lawfare scope analysis; POLITICO Europe on AI Office–METR; Enterprise DNA on the Sep 15 milestone and inspection wave; Slaughter and May and CSA enforcement-readiness briefings; Praxikon penalty-structure analysis); INTERPRETATION (Sep 15 as implemented/operational milestone vs statutory date; internal-deployment scope; Annex III inspection sequencing); PREDICTION (first Art 101 fines, FLOP-threshold litigation, CoP update, transatlantic divergence). The discovery-named E&T article could not be verified and was excluded. No rumours used.