Is AI Gaining Agency Faster Than the Law Can Assign Responsibility?

My earlier argument on the accountability gap at the heart of AI postulated that delegating authority to AI should never allow humans or corporations to delegate responsibility with it.

A recent US court ruling in Amazon.com Services v Perplexity AI shows just how quickly that principle is going to be tested. The Ninth Circuit considered whether Perplexity could be said to have accessed Amazon’s systems when its AI Assistant acted through a user’s browser. The Court concluded, on the facts before it, that it was the user not Perplexity who accessed Amazon, assisted by the AI. Critically, it observed that however advanced the Assistant might be, legally it remains ‘a tool, not a person’.

That creates a fascinating accountability paradox. AI increasingly possesses agency without legal personality. It can navigate, select, recommend and act, yet the law must ultimately attribute those actions somewhere else. In this case that attribution pointed towards the user. In future cases involving more autonomous agents, the answer may be considerably less comfortable.

Imagine an enterprise AI agent independently negotiating purchases, moving money, altering infrastructure or making decisions affecting customers. When something goes wrong, was it the user who initiated the objective, the company deploying the agent, the developer who created it or the executive who authorised its autonomy?

The Court wisely stressed that its ruling was narrow and did NOT establish a general legal regime for agentic AI or determine wider questions of liability. Indeed, it explicitly recognised that legal understanding will evolve as AI becomes more sophisticated. Therein is precisely the warning.

We are rapidly creating machines capable of exercising delegated authority while our legal frameworks still search for the human hand operating them.

Perhaps the danger is that we are building Frankenstein’s monster with a corporate org chart where the machine acts, the user prompted it, the developer trained it, the company deployed it and the executive approved it, yet when something goes wrong, everyone points enthusiastically at someone else! AI may be the first technology sophisticated enough to make decisions at machine speed while responsibility travels at committee speed. The machines need not become sentient to cause trouble; they merely need humans to become sufficiently inventive at explaining why nobody was actually in charge.

For now I see an emerging principle, simply the greater the autonomy we give machines, the clearer the accountability we must retain for humans.