AI Governance: The Accountability Question Every Board Will Face

The evidence that changed the conversation

The AI governance framework your board adopts this year is the one your general counsel and risk committee will be defending for the next decade. The conversation shifted from what AI might do to who is accountable when AI acts, and the question facing every GC, CRO, and audit committee changed with it.

At the center of this shift is a landmark Delphi study from MIT FutureTech and the University of Queensland School of Psychology, led by Neil Thompson and Peter Slattery. The researchers asked 272 AI specialists representing 37 countries to assess 24 distinct categories of AI risk. Their judgment under a business-as-usual path is stark: In 18 of the 24 areas, experts saw more than a 1-in-10 chance of a truly catastrophic outcome, defined as more than one million deaths, at least $100 billion in damage, or harm on a similar scale. Even after assuming reasonable, real-world safeguards, five risk categories remained above that 10% threshold.

The takeaway is not that catastrophic outcomes are inevitable, but that many experts believe AI risks now warrant the same level of governance applied to other high-consequence technologies. As Thompson noted, “In any other mature technology field, think nuclear power or aviation, risk at that level would be treated as unacceptable.”

Three frameworks for what governance actually requires

Experts have published sharply different views on what governing that risk actually requires. Three frameworks, each grounded in recent research and implementation work, now define the strategic options boards are debating.

Iyad Rahwan: Moral reasoning requires direct scrutiny

One perspective is that AI’s moral reasoning has to be studied directly, not assumed. In Knowable Magazine (June 17, 2026), Iyad Rahwan argued that as AI increasingly makes decisions on our behalf, understanding how those systems arrive at ethical judgments is no longer optional. They deserve the same scrutiny we’d apply to any decision-maker exercising real authority over people’s lives.

Rahwan’s work also surfaces a deeper challenge: the evidence governance decisions are meant to rely on may itself be compromised. In Nature Communications (June 25, 2026), Rahwan and coauthors documented “LLM Pollution,” the growing contamination of behavioral research as participants increasingly rely on language models to answer surveys instead of responding themselves. If the evidence informing AI governance is itself compromised, governance becomes the problem before regulation does.

Beth Simone Noveck: Governance is an infrastructure question

Another framework reframes governance as an opportunity rather than a constraint. In a Route Fifty interview (June 15, 2026), Beth Simone Noveck, New Jersey’s former chief AI strategist, argued that the greater risk is allowing a handful of companies to define AI governance by default. Her vision of “democratic AI” calls for public-purpose infrastructure and accountability to be built in before governance structures become entrenched.

Noveck’s framework asks boards a different set of questions: Who designed the system you’re governing? Whose values are encoded in it? What public infrastructure would your organization need to govern AI in a way that reflects stakeholder accountability rather than vendor dependency?

Neil Thompson: Risk management at the scale of other high-consequence technologies

The Delphi study itself implies a third framework: that AI governance should adopt the procedural rigor already applied to nuclear power, aviation, and pharmaceuticals. Thompson’s research establishes the empirical baseline for that approach. When experts see catastrophic risk at the levels documented in the study, the governance question is not whether to act, but whether existing structures are sufficient.

This framework asks boards to assess AI risk through the same lenses applied to other technologies where failure has systemic consequences: independent review, pre-deployment stress testing, clear lines of accountability, and regulatory structures that evolve with the technology.

The question leadership teams should be debating

The Thompson-Slattery study makes one thing clear: the governance framework your organization builds now will determine how you defend your decisions when AI acts in ways you did not anticipate. The question is not whether to govern AI, but which framework your board can implement with discipline and defend with evidence.

Is your governance framework grounded in direct scrutiny of AI’s decision-making, in the public accountability structures Noveck describes, or in the procedural rigor Thompson’s findings imply? The answer matters, because the next decade of litigation, regulation, and board liability will test that choice.

Iyad Rahwan, Beth Simone Noveck, and Neil Thompson are each exclusively represented by Stern Strategy Group for keynotes, board sessions, and confidential advisory work. Reach out through sternstrategy.com to discuss who fits your conversation.