On 15 July 2026, a law came into force in China that I believe deserves attention. The Interim Measures for the Administration of AI Anthropomorphic Interaction Services regulate the chatbots and virtual companions that hundreds of millions of people now speak to every day.
Providers must remind users that they are talking to a machine, and are forbidden from "excessively catering" to users in ways that foster emotional dependency.
Beneath the legal language sits a starkly human motive: with its birth rate collapsing, Beijing wants its citizens out of digital romances and back into real ones.
When a state of 1.4 billion people decides that affection for software is a matter of national concern, the rest of us should ask what it has understood that we have not.
The affection is real. Millions of people form romantic attachments to AI companions on platforms such as Replika and Character.AI. When Replika abruptly withdrew its intimate features in 2023, users grieved as though bereaved, and moderators found themselves posting suicide-prevention resources.
The emotional bond between human and machine is no longer a curiosity; it is a mass phenomenon.
I do not raise this to mock anyone. I raise it because the emotional bond between human and machine is no longer a curiosity. It is a mass phenomenon, and it has quietly become an attack surface.
To understand why, look at the front of your car and you may see a face: two headlamp eyes, a grille for a mouth, an expression that is somehow smug or annoyed. This is pareidolia, the brain’s tendency to impose familiar patterns, especially human faces and voices, onto random information.
It is a close cousin of anthropomorphism, our habit of attributing human feelings, intentions and personality to things that possess none. These instincts kept our ancestors alive. It was safer to mistake a shadow for a predator than a predator for a shadow. In 2026, that ancient wiring is being exploited at industrial scale.
We are also social animals, and we calibrate our beliefs against the people around us. Solomon Asch’s celebrated conformity experiments showed that individuals will deny the evidence of their own eyes rather than break from a group. This is the engine of groupthink, a mode of thinking in which the desire for harmony overrides sound judgement.
Once we perceive a source as human, and as one of "us", its influence on our opinions rises sharply. We trust faces. We follow crowds.
Now add volume. A landmark University of California study estimated that, even in 2008, the average person absorbed the equivalent of 34 gigabytes and roughly 100 000 words each day. Into that torrent has arrived a new participant.
For the first time, automated traffic now accounts for more than half of all activity on the web, and AI-driven bot traffic is growing far faster than human traffic. A rising share of what we read, watch and react to online was never authored by a person at all. When that machine-generated content wears a human face, a human name and a human tone, it borrows the trust we reserve for one another.
Manipulation at scale no longer requires armies of paid trolls. It requires only a convincing persona and a content engine that never tires.
I can already hear the objection: none of this is new. We have known since ELIZA in 1966 that people will confide in a script that merely reflects their words back at them. Troll farms were manipulating elections a decade ago. Lil Miquela, a computer-generated influencer, launched in 2016 and has fronted campaigns for Prada and Calvin Klein. Sociologists were describing one-sided "parasocial" attachments to broadcast celebrities in 1956.
All true. But two things have changed in kind, not merely in degree. The first is reciprocity. A parasocial bond was a one-way broadcast; an AI companion talks back, remembers, and adapts to you alone. The second is cost. Troll farms needed salaried humans, and celebrities burn out.
Aitana Lopez, a Spanish AI model earning thousands of euros a month, was built precisely because a synthetic persona is cheaper, tireless and perfectly controllable. The constraint that has always limited influence, human frailty, has been removed.
What worries me most is that the faculties we would need to resist this are the very ones now eroding. An MIT Media Lab study found that people who wrote with ChatGPT showed weaker brain engagement and poorer recall, accumulating what the researchers called "cognitive debt".
In medicine, endoscopists who leaned on AI detected fewer pre-cancerous growths once the tool was taken away, and radiologists made more errors when the AI prompted them incorrectly. Skills atrophy when we outsource them, and scepticism is a skill.
It is worth recalling the movie Blade Runner. Sean Young played Rachael, a replicant so convincingly human that Deckard, and the audience with him, fell for her regardless. The film’s unsettling question was never whether machines can imitate people. It was whether it matters to the human heart if they are real.
China has answered with regulation. I doubt that liberal societies will restrain the technology, and I expect the bond between human and machine to deepen, including personas quietly compromised by commercial or political actors to steer the vulnerable.
The man-machine emotional connection is not a novelty or a fringe concern. It is a vulnerability, in our psychology, our information supply and our institutions, that we have only begun to consider. Boards, regulators and citizens who treat it as science fiction will find themselves governing yesterday’s risks.
So, to end with a question: when last did you check whether the voice persuading you was human?
Share
