The Economist: Agent of whom? What happens when the newly agentic citizenry take charge?

So the last shall be first and the first shall be last. The race to build artificial intelligence is so frenetic that declaring any firm the winner is a sure sign it is about to be overtaken — and vice versa.

Meta’s AI investments had been written off by many as wasteful vanity. Yet the same investors now say Muse, the firm’s new personal agent, will soon upend commerce and the world. Muse is at the top of the charts on Apple’s App Store. Meta’s share price has risen by around a third in the past month.

Chatbots are often compared to interns. Muse is more like an indentured secretary. Users will find the difference total: between being read a train timetable and having tickets booked; mulling one’s rights and filing a claim; comparing phone tariffs and overcoming the obstacles companies erect to prevent customers switching.

Markets for goods and services can be surveyed totally and ceaselessly. No form is too long; no claim too small.

Which of the millennium mathematics problems can Muse solve? That is unclear. But prod Muse in the right way and you could wake up with a cheaper car insurance policy.

Might Muse agents end the world? Mark Zuckerberg, Meta’s boss, has dismissed worries about AI safety, much as the White House has. Though not yet the most popular AI product, Muse is surely the first populist one, designed for the everyman of Facebook Marketplace for whom privacy does not matter, existential risks are bunk and everything in the economy is either a bargain or a scam.

The mood in markets is similar to what it was in February. Back then, as it became clear that models could write code better than humans, the question was: what happens when everyone has the coding ability of a Silicon Valley engineer?

Shares in software companies were sold indiscriminately (many have since recovered). Now the question du jour is something like: what happens when everyone has the patience of a monk?

The state will struggle with its newly agentic citizenry. When the cost of identifying and claiming all the rights and benefits rich countries offer their people falls, governments will end up spending more. That is an acute problem for courts, which are already flooded by AI-written filings.

The rule of law demands that citizens can enforce their rights. But also that courts remain open, and judges human. Troublingly, most solutions to this trilemma involve more money or fewer rights.

Companies will suffer in proportion to the profits they receive from ignorance and inertia. Economists have long speculated about why money sits in bank accounts when it could earn far more interest in a money-market fund. Or why insurance customers allow their loyalty to be punished with higher rates, and they don’t simply switch.

It is often said that the West suffers from a compensation culture, but plenty of compensation goes unclaimed. In Europe only around half of potential compensation for delayed and cancelled flights is actually paid out — often with the help of firms that take their own cut from afflicted customers.

When consumers let agents loose on their behalf, firms that rely on subscriptions are less secure; it costs little to test Muse by attempting to cancel an unwanted gym membership or newsletter subscription.

So are businesses that depend on being able to distinguish humans from computers. A qualification earned online must now be worth less. Most would find the idea of flirting with a robot intolerable. Once agents get chatting on dating apps, those apps will be reduced to elaborate matching systems between singletons’ online personas.

States and firms could attempt to reimpose frictions with their own technology. Imagine agents keeping other agents on hold, ignoring and chasing each other in increasingly aggressive email correspondence. (Ironically, such defences will make the internet even more unbearable for humans.)

Shopify, an online-commerce platform, has welcomed Muse agents. Amazon has banned them. If the AI bulls are to be believed, that will soon be costly, since agents will be responsible for a meaningful share of commerce, and because they will be harder to identify: “tick the box to prove your humanity” becomes “click the yellow bikes” becomes “upload a photo with today’s newspaper”.

Unburdened of administrative work, consumers might ask a more fundamental question of their new agents: what is the relationship between Man and Muse? Opinions about the homo-agentic bond will differ. The user of an agent is likely to consider her new assistant an ordinary tech product, for which she does not bear total responsibility.

To the world, of course, the agent and its manager will be quite indistinguishable. Model-makers surely think much of the responsibility lies with the user.

Free agents and free will

Consider three scenarios.

John instructs his agent to write funny posts online, and it proceeds to libel someone; Susan tells her agent to negotiate the sale of some goods but, being prone to exaggeration, it misrepresents them; chatting with his agent, Paul orders a kite for his young son, who then receives, and cuts himself with, a knife.

Careless people operating careless models, you say. But the concepts normally used to apportion blame are stretched by the nature of AI. What does it mean to say that harms caused by something as general and mystical as AI are foreseeable? If the model-maker is blamed, does it matter that its code was written by agents, plausibly designed by another lab?

The tension between autonomy and responsibility is among the most fundamental questions in AI. Mostly this has been discussed in the abstract, in terms of the existential risks and AI consciousness. Now a million agentic citizens let loose on the economy will find out for themselves. If they don’t like the result, they know how to appeal.

The Economist