⚡ Quick Summary
A preprint dated 2 June 2026 and written by authors affiliated with Google DeepMind and the Centre for the Governance of AI, this position paper argues that frontier AI governance centred on models becomes less effective as capability progress is increasingly driven by 'non-model gains' — improvements independent of advances in the base model. It formalises three vectors: inference gain (scaling compute at test time), systems gain (post-training enhancements such as scaffolds and tool use), and asset gain (combining a model with restricted assets such as classified data or specialised hardware). Three further prospective vectors are discussed: embodiment, continual learning and diffusion.
The paper contends that risk management resting mainly on pre-deployment evaluation and mitigation is undermined by these gains, and sets out responses: an evolutive elicitation standard, partnerships to improve the elicitation stack, post-deployment monitoring and forecasting of capability overhang, plus governance beyond the model level — system, entity, agent and cloud governance — with societal resilience as a complement. Table 1 maps governance strategies to capability paradigms, and the paper closes with a call to action on metrics, monitoring, forecasting and complementary governance research.
🧩 What’s Covered
The document runs to six sections with a reference list; the major parts, in order:
- Introduction (Section 1): frames the question of the limits of and alternatives to model-level governance, states the aim of informing policy debates, and previews the paper's structure.
- Model-level governance and its limits (Section 2): defines model-level governance as measures informed by a model's capability profile, such as dangerous capability evaluations, access controls for model weights, unlearning and behavioural alignment, and identifies three failure modes — elicitation failure, mitigation failure and overhang cost — focusing on the first.
- Non-model gains (Section 3): analyses inference gain, systems gain and asset gain, citing OpenAI's o1 series, recursive self-aggregation bringing Qwen3-4B-Instruct-2507 on par with o3-mini (high), DeepSeek-V3.2 out-performing Gemini 3 with 1.5 to 2.5 times more tokens, Big Sleep discovering a zero day, and a Claude Code scaffold built by a Chinese state-sponsored group.
- Future limits (Section 3.4): embodiment gain, continual learning and diffusion effects, including monoculture and cascading failures.
- Governance strategies (Section 4): enhanced model-level measures — an evolutive elicitation standard, partnerships, post-deployment monitoring via agent benchmarks such as GAIA, AgentBench, SWE-Bench, MLE-bench and Cybench, and forecasting — summarised in Table 1.
- Beyond model-level governance (Sections 4.2.1–4.2.5): system, entity, agent and cloud governance, the last covering KYC, content-based monitoring and monitoring of computational patterns, plus societal resilience.
- Alternative views and call to action (Sections 5 and 6): two objections answered, then four asks on metrics, monitoring, forecasting and complementary governance.
💡 Why it matters?
For people governing, assessing or auditing frontier AI, the paper supplies both a vocabulary for a risk that pre-deployment evaluation alone does not capture and a set of places to intervene. It explains why model-level evaluations may stop reflecting what downstream actors can elicit, and points to elicitation standards, monitoring of agent benchmarks and public channels, threat intelligence sharing, and cloud-level signals. It links these to instruments the document cites, including post-market monitoring under the EU AI Act Code of Practice, GPU export controls and the US AI Action Plan's requirement that federally funded researchers use nucleic acid synthesis providers with screening protocols.
❓ What’s Missing
As a position paper it formalises concepts rather than measuring them: no metrics, thresholds or methods for sizing non-model gains are given. The authors state that magnitudes are hard to assess, especially for asset gain and diffusion effects, where even establishing a baseline is difficult. Feasibility is left open throughout: cloud governance is said to face contractual, legal, technical and commercial hurdles; entity governance may privilege incumbents or devolve into formalistic compliance; system governance must handle many heterogeneous providers. The paper offers no implementation steps, timelines or cost estimates, and its analysis is not tied to a single jurisdiction.
👥 Best For
Governance and policy analysts tracking the limits of model-level regulation; frontier model developers and safety teams designing elicitation, post-deployment monitoring and forecasting; and compliance or assurance staff who need a structured case for governing systems, organisations, agents and cloud deployments alongside the model.
📄 Source Details
Full title: Position: Comprehensive AI governance requires addressing non-model capability gains. Authors: Arthur Goemans, Dan Altman, Noemi Dreksler, Jonas Freund, Milan Gandhi, Zhengdong Wang, Sarah Cogan, Sebastien Krier, Demetra Brady, Lewis Ho and Allan Dafoe, affiliated with Google DeepMind and the Centre for the Governance of AI. Marked 'Preprint. June 2, 2026' and stamped arXiv:2606.00047v1 [cs.CY] 1 May 2026. 12 pages, in English, with a reference list. The extraction covered all 12 pages. No publisher or document URL is printed.
