The Invisible Game: Structural Frictions in AI Governance Strategy

In a recent essay published on LessWrong, the co-founder of the French Center for AI Safety (CeSIA) details a critical strategic imbalance within the AI safety community: an over-investment in visible, public-facing advocacy at the expense of invisible, institutional policy work. For technical stakeholders, funders, and policy strategists, this analysis exposes a structural friction between the academic culture of open intellectual production and the relationship-driven realities of executive state power.

The Illusion of Visibility in Policy Influence

The AI safety ecosystem has historically relied heavily on what the source terms the "outsider game." This strategy prioritizes public intellectual production, including open letters, press campaigns, think-tank reports, and peer-reviewed academic papers. While these mechanisms are highly visible to the public and generate significant engagement within the technical community, the author argues that they are frequently invisible to the actual decision-makers who shape state and international policy. The core friction lies in how impact is measured and perceived. Technical and academic communities default to quantifiable metrics such as citation counts, media impressions, and the volume of high-profile signatories. However, the machinery of government operates on a fundamentally different frequency. The author posits that the community exhibits a systemic bias against invisible types of work, operating under the assumption that broad public awareness automatically translates into targeted policy action. In reality, public pressure often fails to penetrate the specific ministerial cabinets, working groups, and bureaucratic layers where regulatory frameworks are actually drafted, debated, and implemented.

The Executive Branch and Institutional Integration

Drawing on firsthand experience transitioning from technical safety initiatives-such as EffiSciences and ML4Good-to co-founding CeSIA, the author highlights the disproportionate impact of the "insider game." This approach involves direct, often unpublicized engagement with executive branches and international forums. The source cites specific involvement with the OECD Hiroshima AI Process Reporting Framework (HAIP), United Nations forums, the EU AI Office Code of Practice, and French ministerial cabinets. A critical insight from the text is that effective governance work in the executive branch complements legislative efforts but requires fundamentally different, less public skillsets. Legislative work often involves public hearings, media management, and broad political posturing. In contrast, executive work relies on technical advising, drafting administrative codes, negotiating definitions, and building long-term trust with career civil servants. The author notes that technical professionals often find this invisible game alien and slightly distasteful initially, reflecting a cultural barrier that prevents highly qualified technical talent from entering the bureaucratic spaces where their expertise is most needed to shape practical implementation.

Strategic Implications for Ecosystem Resource Allocation

The dynamic described in the source points to a major strategic blind spot in how the AI safety ecosystem allocates funding, talent, and institutional support. If the most impactful policy work occurs behind closed doors, current philanthropic and organizational models that reward visible intellectual production are misaligned with actual policy leverage. This structural friction suggests that the ecosystem must aggressively diversify its talent pipeline. Rather than solely funding researchers to produce public-facing white papers, there is a pressing need to support operators willing to embed themselves within state apparatuses and international standards bodies. This requires a shift in how the community defines and rewards success, moving away from public attribution toward quiet efficacy. Furthermore, the reliance on the outsider game carries the risk of long-term marginalization. As AI governance transitions from theoretical debates to concrete regulatory implementation-such as the operationalization of the EU AI Act and the enforcement of compute thresholds-governments will increasingly rely on trusted insiders rather than external critics. Organizations that fail to build these quiet, institutional relationships risk losing their seat at the table during the critical phases where abstract laws are translated into binding technical standards.

Limitations and Open Questions in the Insider Model

While the source presents a compelling argument for rebalancing strategic priorities toward executive integration, the analysis lacks specific, verifiable case studies demonstrating the efficacy of these invisible interventions. The author asserts that a significant portion of the work that mattered in AI governance has been invisible, but does not provide concrete examples of how specific closed-door engagements directly altered regulatory outcomes or prevented regulatory capture by industry incumbents. Additionally, the precise mechanics of how entities like the EU AI Office Code of Practice or the OECD HAIP operate behind closed doors remain opaque in the text. This lack of transparency makes it difficult for external observers to evaluate the actual leverage technical advisors hold in these forums compared to well-funded corporate lobbyists. The source also briefly mentions hesitations regarding the replication of the ControlAI initiative in France, citing the unique dynamics of executive branch work, but leaves the explicit reasons behind this hesitation unexplored. Without understanding the specific failure modes, political sensitivities, or structural differences between UK and French governance models, it is challenging to extract actionable lessons for other jurisdictions attempting to build insider influence.

The transition from public advocacy to institutional integration represents a necessary maturation phase for AI governance. As the regulatory landscape hardens globally, the ability to navigate ministerial cabinets, advise on technical definitions, and influence international reporting frameworks will likely outweigh the influence of open letters and press campaigns. For the technical community, bridging the cultural divide between open research and quiet bureaucratic operation is no longer just a matter of preference, but a prerequisite for sustained policy impact. Recognizing and funding the invisible game is essential for ensuring that technical safety concerns are embedded into the actual machinery of state governance.

Key Takeaways

• The AI safety community over-invests in visible 'outsider' advocacy, such as open letters, while under-investing in 'insider' executive policy work.
• Effective governance requires navigating ministerial cabinets and international forums, which demands different skillsets than public intellectual production.
• Current funding and talent allocation models are misaligned with actual policy leverage, creating a strategic blind spot in the ecosystem.
• The lack of transparency in closed-door policy interventions makes it difficult to measure the precise impact of technical advisors against corporate lobbying.