My Stance
on Artificial
Intelligence
I am neither a techno-optimist nor an AI doomsayer. I am an academic, a lawyer, and a governance researcher — and my position on AI is forged at the intersection of all three. These are my considered, professional views. They are not provisional.
Overview
Where I
actually stand.
AI is the defining technological development of this century. I engage with it as such — rigorously, sceptically, and with an acute awareness of what is at stake. My work spans the University of Cambridge, UNIDIR in Geneva, and the Campaign to Stop Killer Robots precisely because this technology demands expertise that crosses disciplinary lines.
The dual literacy I have developed — technical understanding of how these systems work alongside legal and policy analysis of how they should be governed — matters because the two are inseparable. A system that cannot be audited cannot be governed. A system that cannot be governed cannot be trusted. And a system that cannot be trusted cannot realise its genuine potential.
The arguments below represent my considered, informed professional position. They will develop as evidence accumulates — any serious intellectual position must — but they are not hedges, and they do not depend on the approval of any government or industry.
AI represents perhaps the most consequential governance challenge in modern history. The pace of capability development has outrun the development of legal norms, institutional oversight, and democratic accountability mechanisms by a significant margin.
The central question is not whether AI should exist or advance — it will, and there are genuine reasons to welcome much of that progress. The question is who controls it, on whose authority, and what legal structures bind it. Those are not technical questions. They are political, legal, and deeply moral ones.
My position is that meaningful human control is not a philosophical aspiration — it is a legal imperative. It must be embedded structurally into both the technical design of AI systems and the international frameworks that govern their deployment.
On AI's Transformative Potential
I am not against AI.
I am for accountability.
Let me be unambiguous on a point that is routinely misread: I am genuinely enthusiastic about what artificial intelligence can achieve. The prospect of AI accelerating medical diagnostics, unlocking scientific discovery at scale, and extending access to legal expertise across the developing world — these are not marginal gains. They represent transformative progress for humanity, and I take them seriously.
Technological potential is not destiny. The history of dual-use technology — from nuclear power to the internet — demonstrates that the benefits of powerful innovations are not automatic. They are the product of deliberate, sustained governance choices made at the earliest possible stage. The time for those choices is now, not after the first major failure.
The failure to govern AI effectively is not a technical problem waiting for a technical solution. It is a political will deficit — actively cultivated by actors who benefit from the resulting vacuum.
— Avi PereraI refuse to accept the argument — advanced, conveniently, by those with the least incentive to accept constraint — that the pace of innovation must necessarily outstrip democratic accountability. That is a choice, made by powerful actors, and it can be unmade through political action and legal instruments.
AI-assisted diagnostic tools have demonstrated accuracy in radiology, genomics, and drug discovery that exceeds specialist performance in controlled settings. With rigorous clinical validation and clear liability architecture, this represents genuine life-saving potential — and a model for responsible deployment in high-stakes domains.
Legal AI has the capacity to close persistent justice gaps in jurisdictions where legal expertise is prohibitively expensive or geographically inaccessible. The conditions for responsible deployment are exacting — accuracy standards, professional oversight, liability frameworks — but the potential to democratise legal access is real and significant.
Protein structure prediction, climate modelling, materials science — AI's capacity to compress the timelines of scientific discovery is extraordinary. These are precisely the use cases that make investment in AI capability worthwhile, and that provide the strongest argument for getting governance right before catastrophic failures force a reckoning.
AI-powered tools in agriculture, energy grid management, and low-resource language processing can genuinely accelerate development trajectories in lower-income economies — provided the technology is accessible rather than proprietary and ring-fenced behind IP regimes that replicate existing global inequalities at scale.
On AI Risk & the Safety Imperative
The risks are not
hypothetical.
The AI safety discourse is commonly framed as a binary between catastrophism and dismissiveness. I reject both. The genuine risk landscape is more granular, more immediate in some respects, and more tractable in others than either pole captures. The harms I regard as most pressing are not purely speculative frontier risks at some indeterminate future point — they are current, compounding, and in several cases already causing demonstrable harm to identifiable people.
The dependency of modern states on interconnected digital infrastructure — power grids, financial systems, health networks — creates attack surfaces of historically unprecedented scale. AI-enabled cyber operations, whether state-sponsored or non-state, represent an asymmetric threat that outpaces current defensive doctrine and regulatory capacity. I regard this as a near-term, under-governed risk of the first order.
AI-generated synthetic media is already degrading the epistemic foundations of democratic discourse. This is not a future risk. It is an ongoing, compounding erosion of shared reality — materialising faster than regulatory responses can track. The implications for elections, judicial proceedings, and the credibility of international communications are severe.
I am not a techno-pessimist about AI's long-run economic effects. But the aggregate-level argument that new technologies generate new jobs does not automatically translate into acceptable outcomes for specific communities most exposed to near-term displacement. The policy response requires proactive distributional intervention — retraining infrastructure, social insurance reform, active labour market policy — currently running well behind the pace of disruption.
The potential for advanced AI to provide meaningful uplift in the development of biological or chemical weapons is, in my assessment, a genuine existential-class risk that warrants treatment commensurate with its severity — including hard capability restrictions, not merely voluntary commitments from frontier labs. Responsible Scaling Policies represent genuine progress; they are not a substitute for binding legal instruments with enforcement mechanisms.
A note on prioritisation: Autonomous weapons appear in a separate section because the governance challenge they present is structurally distinct. LAWS are not merely a risk category to be managed — they represent a categorical threat to the foundational architecture of international humanitarian law, and they require a categorical legal response.
On Autonomous Weapons & Lethal AI
We cannot delegate
the decision to kill
to a machine.
This is the domain in which my research is most concentrated, and where my position is most unequivocal. Lethal Autonomous Weapons Systems — systems capable of independently selecting and engaging targets without meaningful human control — represent the most acute governance challenge at the intersection of artificial intelligence and international law.
The question is not whether autonomous systems can be militarily effective. They demonstrably can. The question is whether they can be legally compliant, morally legitimate, and democratically accountable. In the absence of fundamental legal architecture that does not yet exist, my assessment is that they cannot.
— Avi PereraThe IHL Compliance Failure
The foundational principles of International Humanitarian Law — distinction, proportionality, and military necessity — require the kind of contextual moral reasoning that current AI systems are incapable of exercising in real operational environments. The principle of distinction demands reliable differentiation between combatants and civilians in complex, dynamic settings. The principle of proportionality requires the weighing of incommensurable values — the anticipated military advantage of an engagement against its foreseeable civilian harm.
These are not computational optimisation problems. They are moral judgements — judgements that presuppose contextual understanding, ethical reasoning, and legal accountability that no algorithm currently possesses. The deployment of systems premised on the fiction that they do is generating violations of binding international law.
The Accountability Gap
When an autonomous system commits what would otherwise constitute a war crime, international law as presently constituted has no coherent answer for who bears criminal responsibility. The operator may be too remote. The commander may have been unable to foresee the system's emergent behaviour. The designer may be a private contractor in another jurisdiction. The machine cannot be prosecuted.
This is a structural failure requiring new legal instruments: command responsibility doctrine extended explicitly to AI deployments, mandatory pre-deployment legal review against IHL criteria, and accountability frameworks that cannot be dissolved through operational complexity or system opacity.
The "flash war" risk: Opposing autonomous systems operating at machine speed without human decision-makers in the loop create conditions for rapid automated escalation that no human actor initiates or controls. Deterrence theory is premised on rational human decision-making. Remove the human, and the architecture of international security stability fails.
I advocate for a legally binding international instrument — under the CCW process, or through alternative treaty mechanisms if that process continues to be obstructed — that mandates meaningful human control over all individual targeting and engagement decisions in armed conflict. This is not a call to halt military AI research. It is a call for the law to keep pace with the technology, as it must.
On Meaningful Human Control
Control is not a button.
It is an architecture.
"Meaningful human control" has become the operative term of art in LAWS discourse — and like many such terms, it risks being domesticated into meaninglessness through loose and self-serving application. Certain states and defence manufacturers have begun invoking MHC as a threshold satisfied by minimal human involvement at some point in an operational sequence. That reading is not adequate.
Meaningful human control is a structural condition, not a procedural checkbox. It requires genuine understanding of what a system is doing and why, genuine capacity to intervene before irreversible consequences are generated, and genuine traceability of moral and legal responsibility back to human actors. Each of these elements must be designed into AI systems from the ground up. They cannot be retrofitted onto autonomous architecture after deployment.
Human operators must possess genuine comprehension of a system's operational logic, decision parameters, and real-time environmental context — not mere familiarity with its outputs. Black-box compliance, where a human approves a decision they cannot meaningfully evaluate, does not satisfy this requirement. Meaningful oversight is not possible without meaningful comprehension.
Systems must be architecturally designed with explicit, operable human override pathways genuinely available before irreversible consequences are generated. The timescale of human decision-making — cognitive processing, communications latency, command chain delays — must function as a hard design constraint, not an afterthought optimised away in the name of operational effectiveness.
The chain of moral and legal accountability must, at all points in a system's operational lifecycle, trace back to identifiable human actors — commanders, developers, or deploying states. Legal responsibility cannot be dissolved through algorithmic complexity or distributed across a diffuse network of partially responsible actors until no individual bears meaningful liability.
The Anticipatory Objection Registry: One institutional mechanism I have developed in my academic research is a formal international Anticipatory Objection Registry — an instrument through which states can register advance legal objections to specific autonomous weapons configurations prior to their deployment. This functions both as a deterrent and as a method for generating customary international law norms around permissible control thresholds, without requiring unanimous treaty agreement as a precondition.
On International Governance
Voluntary commitments
are not governance.
The current international AI governance landscape is characterised by a proliferation of principles, guidelines, national strategies, and voluntary commitments — and a stark, consequential deficit of binding, enforceable legal instruments. I regard this asymmetry as the central structural failure of the current governance moment. Voluntary frameworks are, by design, frameworks that actors can depart from when departure is convenient. That is not governance. It is the performance of governance in the absence of its substance.
The UNESCO Recommendation on the Ethics of AI, the EU AI Act, the OECD AI Principles, the G7 Hiroshima AI Process, and the UN Global Digital Compact collectively constitute a substantial body of normative guidance. All represent genuine diplomatic progress. None are sufficient to constrain the actors most capable of generating catastrophic harm, because none carry binding legal force on the actions that matter most.
A binding international framework — ideally under UN auspices, with treaty-level legal force — that establishes minimum mandatory standards for high-risk AI deployments, creates legally enforceable accountability mechanisms with genuine consequences for breach, and provides institutional oversight bodies with both the technical capacity and political independence to function effectively in real time.
The Group of Governmental Experts on LAWS within the CCW framework has been deliberating since 2014. Its consistent failure to produce a binding legal instrument — systematically obstructed by states with vested interests in autonomous weapons development — is a diplomatic failure of the first order. The multilateral process must continue and intensify, but it cannot be treated as the only available pathway to legal constraint.
The decisions being made now about AI governance will shape the architecture of political, economic, and military power for decades. These decisions are disproportionately being made in closed conversations between states and industry, with minimal public deliberation and accountability. AI governance is not a technical matter to be handled by specialists. It is a democratic matter requiring democratic legitimacy.
The governance gap is not primarily a technical problem. It is a political economy problem. The actors with the greatest capacity to close the gap are, in many cases, the actors with the strongest incentive to keep it open.
— Avi PereraOn Corporate Responsibility
Speed-to-market is not
a legitimate safety strategy.
I work extensively with enterprises navigating the operational and legal dimensions of AI deployment. That experience has given me a detailed view of the gap between stated corporate AI ethics commitments and actual institutional behaviour — a gap that is, in many cases, substantial. The structural problem is this: organisations that invest seriously in pre-deployment safety evaluation face a time-to-market disadvantage relative to competitors who do not. This creates a race-to-the-bottom dynamic that cannot be resolved by voluntary ethics commitments. It requires binding minimum standards that apply equally to all market participants.
I have observed a pattern across the AI industry of announcing safety programmes while simultaneously dismissing safety-focused researchers, marginalising internal dissent, and actively lobbying to weaken proposed regulatory requirements. The co-existence of safety rhetoric and safety-undermining practice must be called what it is: a communications strategy, not a safety strategy.
I am also attentive to the structural dimension of regulatory capture — the revolving door between regulatory bodies and the AI industry, and the extent to which industry lobbying has already materially weakened proposed regulatory frameworks in key jurisdictions. These require structural remedies, not just normative appeals to corporate responsibility.
-
→Mandatory pre-deployment risk assessment High-risk AI systems should require independent third-party safety evaluation before market deployment. This is the standard in aviation, pharmaceuticals, and nuclear energy. The exceptionalism that treats AI as categorically different from other high-stakes technologies has no principled basis and must not be allowed to persist.
-
→Enforceable algorithmic audit rights Regulators, researchers, and affected individuals must have legally enforceable rights to audit AI systems whose outputs materially affect them. Commercial opacity cannot shield systems generating consequential decisions from accountability mechanisms when the stakes are sufficiently high.
-
→Director-level personal liability Senior executives should bear personal legal responsibility for foreseeable AI harms arising from systems their organisations deploy, where adequate pre-deployment review was not conducted. Distributing accountability across teams until no individual bears it is not accountability — it is its elimination by organisational design.
On the Path Forward
Governance must move
at the speed of
the technology.
I am not pessimistic about what international governance can achieve. The history of arms control — from the Biological Weapons Convention to the Chemical Weapons Convention to the Ottawa Treaty on Anti-Personnel Mines — demonstrates that binding international instruments can be negotiated, can enter into force, and can measurably constrain state behaviour. The question is not one of legal possibility. It is one of political mobilisation. My research, advisory work, and platform SovDash are all oriented towards the same structural goal: closing the distance between the pace of AI capability development and the pace of the governance frameworks that must contain it.
-
01A binding LAWS treaty — without further delay The CCW process has had over a decade to produce a binding legal instrument on autonomous weapons. The international community must either conclude that process through intensified political pressure or pursue alternative treaty pathways. Every additional year of delay is a year in which lethal autonomous systems are deployed, tested in live conflict, and normalised without legal constraint.
-
02Meaningful human control operationalised into binding standards MHC must move from normative aspiration to specific, legally enforceable technical standards — with clear thresholds for cognitive understanding, intervention capacity, and responsibility traceability, and verification mechanisms attached. Interpretive ambiguity here is not neutrality; it is a resource that will be exploited by those seeking to evade constraint.
-
03Binding corporate liability frameworks The regulatory vacuum that currently allows corporations to deploy high-risk AI systems without bearing proportionate legal consequences for foreseeable failures must be closed through mandatory liability frameworks with genuine enforcement capacity. Voluntary commitments that can be abandoned without legal consequence are not safety mechanisms.
-
04Technical capacity building in governance institutions Regulatory bodies cannot govern what they do not understand. The critical investment is not only in regulatory frameworks but in the technical literacy of the institutions and individuals responsible for implementing them — enabling them to assess, audit, and challenge AI systems on substantive technical grounds.
-
05Democratic legitimacy as a governance requirement The decisions being made now about AI's role in warfare, criminal justice, resource allocation, and political discourse will shape the structural conditions of human life for generations. These decisions require genuine democratic legitimacy — transparent process, public deliberation, and meaningful accountability to affected populations — not just expert consensus arrived at between states and industry in closed forums.
Putting This Into Practice
These positions inform
everything I do.
My stance on AI is not a set of published opinions held at arm's length from professional practice. It is the operational framework for every dimension of my research, consulting, and platform work. The following is how these positions translate into concrete activity — from academic research to civil society advocacy to enterprise advisory work.
My research at Cambridge and through UNIDIR focuses on the legal operationalisation of meaningful human control — translating normative frameworks into specific, enforceable standards that can anchor treaty negotiations and domestic legislation. Current work includes the Anticipatory Objection Registry as a mechanism for generating customary international law norms around permissible autonomy thresholds in weapons systems.
I contribute to the Campaign's LAWS Disarmament Toolkit and broader advocacy strategy, providing technical and legal analysis in support of the civil society dimension of the multilateral disarmament process. The Campaign represents one of the most sustained and effective examples of international civil society mobilisation in the arms control space, and I am proud to contribute to its work.
My platform aggregates geopolitical and AI governance intelligence for high-level stakeholders across government, industry, and civil society. It is built on the conviction that effective governance requires well-informed decision-makers, and that the complexity and pace of the AI policy landscape now demands dedicated synthesis infrastructure rather than periodic reports or general-purpose news aggregation.
I work with organisations at the intersection of AI deployment and regulatory exposure — from algorithmic bias assessment to legal compliance strategy and enterprise risk frameworks. The same principles that underpin my academic research govern my advisory practice: genuine accountability, technical transparency, and rigorous pre-deployment risk analysis rather than post-hoc reputational management.
These questions require
collective engagement.
If you are working on AI governance, international law, autonomous systems, or the ethics of AI deployment — I would like to hear from you. The stakes are too high, and the challenges too complex, for any single researcher or institution to address alone.
Avi Perera · aviperera.com · Cambridge · Geneva · London