At a parish meeting I was at a couple of years ago, the previous treasurer proposed recording the discussion to produce accurate minutes. The reaction to that was so hostile that she had to put her phone away. People really didn't want their every remark available for uncontrolled dissemination. Which brings us to agentic AI.
Agentic AI systems have a problem. People acquire social competence through apprenticeship inside families, friendships and institutions. This is how you learn who may say what to whom, when a remark is tentative and when the literal instruction is not the real objective. Embarrassment, loyalty and exclusion repress conduct which may be technically possible but which is actually socially outrageous.
Up to now the LLMs absorbed the textual residue of human judgement without participating in the social life which produced it. Agentic task performance today misses purpose, overlooks significance, continues blindly when a competent person would stop and ask. The real world is an open system and today's AI systems lack both context and norms.
No wonder the frontier labs are invested in smart glasses: give the AI eyes, ears and persistent memory!
Technically sensible; socially disastrous. The glasses are neither an ordinary tool, explicitly invoked, nor a recognised participant in their own right. Rather, they insert an invisible third party into every encounter. The wearer in principle gets some benefit, hopefully, while everybody else has little idea whether the device is recording, identifying or remembering - and there is no control.
But, you say, perhaps young people may normalise the glasses. Amongst themselves perhaps, with a new social etiquette. But adoption by wearers is still not consent from bystanders like older folk.
Perhaps, you say, a humanoid robot would at least be a legible social fact. It could be introduced, given a role, asked to leave, turn away or forget when the discussion gets delicate - but that's not going to happen, is it? Or not easily.
People are constrained by dependence, reputation, emotional and physical vulnerability. Artificial agents have none of these constraints by default. An AI may simulate tact and conscience but viscerally we won't believe it. Its memories belong to a corporation; its loyalties are as flexible as its updates.
Knowing the norms is not enough: you have to be bound by them.
Long-term field trials are obviously necessary: compare the same AI in phones, glasses, speakers and robots; investigate bystander consent, candour, avoidance and trust.
You know, I don't think this revolution is going to be particularly quick.
Perhaps you saw the recent BBC comedy Ann Droid? Caught the zeitgeist to a degree. Son rents domestic help&companion Robot for his active elderly mother. Robot (that and others) often misrepresent and respond inappropriately to the various social scenes that occur.
ReplyDeleteNevertheless various people wrote into the Radio Times saying that they would buy one of these Robots, when available.