Project Cassandra

As artificial intelligence becomes increasingly capable and autonomous, humanity faces a fundamental challenge: how to retain meaningful control over systems whose decisions and actions may have profound consequences for individuals, institutions and society.

Effective governance requires more than principles, voluntary commitments or regulation of individual AI products. It calls for clear responsibilities, enforceable requirements, genuinely independent assurance and institutions with the authority to act. Human oversight must remain meaningful, decisions must be open to scrutiny, and those entrusted with governing AI must themselves be accountable.

Building on the human and social priorities established in Parts I and II, this third part, AI Governance, Assurance and Control, examines how national and international governance arrangements might provide these safeguards. It considers how authority can be exercised responsibly, how AI capabilities can be assessed and how governance must evolve as those capabilities develop, whilst recognising that detailed institutional design will require further specialist work.

Knowing When AI Is Acting for Itself - Examines the growing difficulty of distinguishing between AI systems carrying out human instructions and those exercising increasingly independent judgement or initiative. It considers the need to recognise when AI is acting autonomously, establish clear boundaries for delegated authority and ensure that such authority remains subject to scrutiny and withdrawal when necessary.

Keeping Human Responsibility at the Centre - Examines the principle that responsibility for decisions and actions involving AI must remain identifiable and attributable to people and their institutions. As AI systems gain greater autonomy, it considers how authority may be delegated without surrendering accountability, ensuring that those responsible retain the obligation to justify decisions, address consequences and intervene when necessary.

Preserving Meaningful Human Control - Examines how meaningful human control can be maintained as AI systems become increasingly capable and autonomous. It considers the conditions under which authority may be delegated to AI, requiring evidence that such delegation is justified, clearly defined limits on its scope, continuing oversight and the ability to intervene, suspend or withdraw that authority when necessary.

When Human Control May No Longer Be Enough - Examines the possibility that increasingly capable and autonomous AI systems may eventually exceed humanity's practical ability to understand, supervise or control their behaviour. It considers the limits of conventional human oversight and the need to anticipate such circumstances, developing additional safeguards and governance arrangements before human control becomes ineffective or impossible to exercise.

Making the Transition Visible - Examines the need to anticipate the effects of AI on employment, the economy and society before disruption becomes widespread. Forward reporting of workforce changes, early warnings for affected occupations and assessment of economic consequences would help governments and communities prepare in good time.

Independent Assurance of AI - Examines the need for genuinely independent assurance of increasingly capable AI systems, supported by verifiable evidence rather than reliance on developers' claims or voluntary commitments. It considers how independent scrutiny, testing and continuing assessment might establish whether systems meet mandatory and other defined requirements and remain safe, accountable and subject to effective human control.

Protecting Personal Information in an AI World - Examines the challenges of protecting personal information as AI systems become capable of collecting, combining and inferring increasingly sensitive details about individuals. It considers the need for safeguards governing the use of such information, including people's ability to understand and challenge inferences made about them.

Preventing the Misuse of Increasingly Capable AI - Examines how increasingly powerful AI capabilities might be misused by individuals, organisations or governments, whether deliberately or through inadequate safeguards. It considers the need to anticipate emerging threats, establish proportionate preventive measures and ensure that those developing and deploying AI remain accountable for managing foreseeable risks.

When AI Tells Us What Is True - Examines the implications of relying on AI to provide information, interpret evidence and influence what individuals and societies accept as reliable knowledge. It considers the need for AI-generated conclusions to be supported by traceable evidence, allowing people to examine their foundations, challenge questionable claims and distinguish established facts from interpretations or unsupported assertions.

AI Helping Humanity Govern AI - Examines how AI might assist humanity in monitoring, assessing and governing increasingly capable AI systems, particularly where their complexity and speed exceed the practical limits of unaided human oversight. It considers how AI-assisted governance could strengthen scrutiny and decision-making whilst ensuring that authority, accountability and ultimate responsibility remain with people and their institutions.

Preparing Society for What May Be Possible - Examines how society might prepare for increasingly capable AI before its full consequences become clear. It considers possible changes to employment, institutions and everyday life. It also explores the need for flexible plans and safeguards as opportunities and risks emerge.

National AI Transition Authorities - Examines the establishment of independent, democratically accountable National AI Transition Authorities, with responsibilities extending beyond advice and coordination to include oversight, independent scrutiny and enforceable decision-making powers. These could include the authority to prevent actions that conflict with established requirements for managing the AI transition.

Governing AI Beyond National Borders - Examines the need for an International AI Governance Framework to address the challenges of increasingly capable AI that extend beyond national borders. It considers how nations might establish shared requirements, coordinate oversight and provide international accountability, recognising that effective governance cannot depend solely on individual countries acting independently.

Building International Participation Progressively - Examines how international participation in AI governance might develop progressively, beginning with countries willing to establish shared requirements and oversight arrangements. It considers how cooperation could expand over time, allowing broader participation without delaying essential safeguards until universal agreement becomes possible.

Governing Capability, Not Just Products - Examines why AI governance must address the capabilities of increasingly powerful systems, rather than concentrating solely on individual products or applications. It considers how oversight and safeguards should reflect what AI systems can actually do, including capabilities that may emerge or be combined in ways not anticipated when they were first developed.

Governance That Changes as AI Changes - Examines the need for AI governance to evolve alongside the capabilities and autonomy of the systems it oversees. It considers how requirements, oversight arrangements and safeguards must be regularly reassessed and strengthened as AI develops, ensuring that governance remains effective rather than becoming outdated by technological progress.

Who Watches the Watchers? - Examines how the institutions responsible for governing AI can themselves be held accountable. It considers the need for independent scrutiny, transparency and democratic oversight to prevent excessive concentration of power, conflicts of interest or misuse of authority by those entrusted with protecting society.