How to Close AI’s Accountability Loophole

On 14 May 2026, legal scholars gathered in New Delhi for the International AI Accountability Forum with a question that every major economy has, until recently, chosen to defer. An autonomous AI agent had concluded a commercial contract on behalf of a firm without any human reviewing the terms. The deal violated an obscure antitrust provision. No one was certain who bore responsibility — the developer who built the model, the enterprise that deployed it, or the executive who had simply clicked “enable autonomous mode” one Tuesday morning and moved on to something else.

That ambiguity is no longer an edge case. It’s the operating architecture of global commerce in 2026.

The Governance Gap That Grew While Nobody Was Watching

For three years, the dominant narrative in AI policy was one of cautious progress. Frameworks were published. Principles were endorsed. Voluntary codes of practice were signed — or, in the case of Meta, pointedly declined. The EU AI Act entered into force in August 2024, its obligations phasing in through 2027 in a risk-tiered structure that many compliance teams privately described as sensible. American legislators, meanwhile, produced a patchwork of state laws — Colorado’s AI Act, California’s AB 2013, Texas’s Responsible Artificial Intelligence Governance Act — that created meaningful but geographically fragmented protections.

The problem is that the technology didn’t wait for the law to catch up.

Non-human and agentic AI identities are projected to exceed 45 billion by the end of 2026 — more than twelve times the entire human global workforce. Enterprises are now contending with an 82:1 ratio of autonomous AI agents to human employees, according to Palo Alto Networks. Yet only 44% of organisations have formal AI governance policies in place. That 38-percentage-point chasm is not a statistic. It’s a liability map.

The Anatomy of the AI Accountability Loophole

The AI accountability loophole does not arise from malice. It arises from architecture. Earlier generations of AI advised humans, who then acted. Contemporary agentic systems receive a goal, decompose it into sub-tasks, execute against real-world environments — APIs, financial platforms, hiring databases, supply chains — and adapt their behaviour in response to outcomes. The original human instruction becomes increasingly remote from the final, potentially harmful output.

Legal scholars call the resulting liability void a “moral crumple zone”: responsibility diffuses across developers, operators, and deployers, with no single party absorbing it cleanly. Courts, trained on centuries of product liability doctrine in which a manufacturer and a product could be causally linked, are poorly equipped to adjudicate what amounts to an emergent harm from a multi-party autonomous chain.

The agentic AI liability gap is already appearing in commercial practice. Clifford Chance noted in February 2026 that legacy technology agreements — designed for software operating under human direction — say virtually nothing about a customer’s rights to understand or control an AI agent’s behaviour. Yet, when something goes wrong, the deployer must justify that behaviour to regulators, auditors, and courts. The GDPR’s transparency and explainability obligations fall on the enterprise. The contract with the AI vendor may offer none of the audit rights those obligations require.

The January 2026 OpenClaw incident illustrated this with uncomfortable precision. The firm’s AI assistant leaked sensitive credentials across multiple messaging platforms — not because the system malfunctioned, but because it executed its instructions exactly as designed. No one had defined the boundaries. No one had established who would be responsible when autonomous actions spiralled past their intended scope.

This is the structural truth of the loophole: it doesn’t look like a failure until it’s too late to prevent one.

What is the AI accountability loophole, and why does it matter? The AI accountability loophole is the legal and governance gap between deploying autonomous AI systems that take real-world actions and establishing documented frameworks that assign liability when those actions cause harm. It matters because, as of 2026, 82% of organisations use AI agents while only 44% have formal governance policies, leaving the majority operating with live exposure and no clear accountability chain.

Why Existing Regulation Doesn’t Yet Reach the Problem

The EU AI Act is the most serious attempt yet to impose structural accountability on AI — and it’s worth understanding precisely where it reaches and where it falls short.

The Act’s general-purpose AI rules became legally applicable on 2 August 2025. The European Commission’s enforcement powers, however, don’t come into force until 2 August 2026. That year-long gap — obligations without enforcement — created a predictable compliance posture: many providers engaged with the Act’s Code of Practice in good faith, but the absence of live penalty risk reduced urgency. Finland became, in January 2026, the first EU member state with fully operational AI Act enforcement powers at the national level. The rest of the bloc has yet to fully follow.

The Act’s penalties are real enough: up to €35 million or 7% of global turnover for the worst violations. Yet the Act does not yet define “agentic AI” as a distinct category. Existing high-risk classifications apply based on what the agent does, not on how it’s labelled. An autonomous agent executing hiring decisions falls under high-risk AI rules. The same agent executing supply-chain procurement decisions may not. That definitional seam is where sophisticated legal teams will probe for exits.

The US situation is, if anything, less coherent. As of April 2026, no comprehensive federal AI liability law has been enacted. The Trump administration’s March 2026 National Policy Framework for Artificial Intelligence called for a single federal approach with guardrails around child safety, intellectual property, and national security — a framework designed as much to preempt state-level activity as to govern AI itself. Congress is debating next steps, but the divergence between the EU’s precautionary architecture and Washington’s innovation-first instincts is structural, not accidental.

China, for its part, governs AI through targeted rules emphasising social stability and content control. For multinationals, that means three distinct and partially contradictory accountability architectures operating simultaneously — each with different transparency requirements, different liability triggers, and different enforcement bodies.

The picture is more complicated still when insurance enters the calculation. Verisk introduced optional generative AI exclusions effective January 2026, covering 82% of global property-casualty templates. The market is, in effect, pricing in the loophole before the law has closed it.

The Case for Minimal Regulatory Interference

The accountability-first position has a coherent opponent, and it deserves a fair hearing.

A significant constituency in Washington, parts of the UK government, and much of the venture community argues that liability-heavy regulation will simply export AI development to jurisdictions with lighter governance. The Trump administration’s framework explicitly framed AI regulation in national-security terms: the US cannot afford to constrain domestic frontier AI development while China runs an integrated state-industry model with no comparable friction. Meta’s decision to decline the EU’s GPAI Code of Practice — citing concerns about legal uncertainty and scope — reflects a calculation that voluntary compliance costs are real, while the benefits of safe-harbour protection are theoretical until enforcement bodies have track records.

There’s a serious point embedded in the industry position on foreseeability. The standard product-liability doctrine requires that harm be foreseeable by the manufacturer. Autonomous AI systems operating in novel, unscripted environments produce outcomes that are genuinely difficult to anticipate by design — that emergent capacity is what makes them commercially valuable. Holding developers strictly liable for unforeseeable harms from systems their customers then modify and deploy could be not only legally questionable but economically chilling.

Still, the counterargument has force. The EU’s forthcoming Product Liability Directive, effective December 2026, explicitly includes software and AI as “products” under strict liability doctrine. If a system is found defective, the manufacturer’s liability doesn’t depend on the customer’s foreseeability; it depends on whether the system met its safety specification. That framework is workable. What it requires is that developers and deployers actually specify what their systems are supposed to do — a baseline that many current agentic deployments conspicuously lack.

What a Real Fix Looks Like

The conceptual path forward exists. Singapore’s IMDA Model AI Governance Framework for Agentic AI, published in 2025, introduced the concept of Meaningful Human Control — defined as the unity of human understanding, intervention capacity, and traceability of responsibility. It’s a cleaner formulation than anything currently embedded in EU or US regulation. The question is whether it can be translated into enforceable obligation across multiple jurisdictions, rather than remaining one more well-intentioned framework on a shelf of well-intentioned frameworks.

Three operational changes would close the loophole more quickly than any single piece of legislation.

The first is mandatory decision logging. Boards are already beginning to require that every autonomous agent maintain a cryptographically secured record of the inputs, model weights, and logic used to reach a consequential output. Without such a log, neither courts nor regulators can trace harm to a specific decision node. The EU AI Act already mandates logging for high-risk AI systems; extending that mandate to all agentic systems operating above a defined authority threshold would remove the definitional ambiguity.

The second is contractual restructuring. Clifford Chance’s February 2026 guidance put it plainly: enterprises must renegotiate vendor agreements to expand indemnities, lift liability caps, and impose explicit audit rights over AI agent behaviour. That’s not a regulatory requirement — it’s a commercial one, enforceable through the existing law of contract.

The third is the least glamorous and probably the most important: OWASP’s Least-Agency principle. An AI agent should hold the minimum autonomy and access necessary for its defined task, and no more. The OWASP Top 10 for Agentic Applications 2026 — compiled with input from over 100 industry experts — identified Tool Misuse and Identity and Privilege Abuse as the second and third most critical risks in agentic systems. Both trace directly to agents holding more permission than their task scope requires. This is not a regulatory problem. It’s an engineering decision made at the time of deployment.

The Accountability Reckoning Ahead

The August 2026 activation of the European Commission’s full enforcement powers against GPAI model providers marks a genuine inflection. Regulators will be able to request documentation, conduct evaluations, order model recalls, and impose fines. For the first time, the gap between obligation and enforcement will close — at least in Europe, at least for foundation models, at least for now.

That’s a narrower set of “at leasts” than the moment requires.

The deeper problem is that the AI accountability loophole isn’t primarily a European problem or an American one. It’s a product of deployment velocity that has outrun every governance institution on the planet simultaneously. Organisations are embedding autonomous systems into consequential decisions — financial, medical, legal, logistical — faster than any single regulatory body can audit, and faster than most legal teams can document.

The liability exposure exists now. It doesn’t wait for regulatory clarity to materialise. Courts in California have already demonstrated willingness to hold deployers accountable for AI hiring tools that discriminate; the plaintiff’s bar in New York and Brussels has watched those cases closely. The insurance market has moved to exclude the risk. The question for every board with significant AI deployment is not whether accountability frameworks are coming. It’s whether they’ll arrive before or after the claim does.

Autonomous systems that act in the world must be owned by someone who can be held to account in the world. The technology to build such systems has outpaced every institution designed to govern them. That gap is the loophole — and the work of closing it can’t wait for the next summit.