There is a technology more terrifying than artificial general intelligence.
The family WhatsApp group.
Five adults are trying to arrange a weekend away.
Someone can do the 12th but not the 13th. Someone wants Spain. Someone won't fly before 10am. Somebody has found an Airbnb which is apparently perfect except that it's three hours from the airport.
Forty-seven messages later:
So... what dates can everyone actually do?
Human civilisation has put people on the Moon.
We remain surprisingly bad at arranging dinner.
This matters more than it sounds.
Because if the idea I explored in Part One, Artificial Personal Intelligence, ever becomes real, the important moment may not be when my AI becomes clever enough to represent me.
It may be when my AI meets yours.
And they start talking.
Not because the machines have become conscious. Not because they've formed some secret electronic society.
Simply because they can coordinate things that currently require humans to spend an extraordinary amount of time finding each other, exchanging information, reconciling preferences and deciding what to do.
That sounds like administration.
It might turn out to be something much bigger.
The holiday problem
Start harmlessly.
Five friends want to go away together.
Instead of everybody publishing their diaries, budgets and preferences into yet another app, their personal AIs negotiate.
Mine might reveal:
Available 12 to 16 September.
Maximum spend £700.
Prefer somewhere warm.
Avoid very early flights unless the saving is substantial.
Another person's AI knows that its owner needs step-free accommodation. It doesn't need to explain why.
Another knows that £700 is absolutely not £750, however enthusiastically somebody describes the hotel.
The agents exchange only what they need.
Then mine comes back:
Porto works for all five people. Option A is £643 each and meets everyone's stated preferences. Option B is £588 but requires a 06:20 flight. Three people have already authorised me to reject flights before 08:00.
Nothing revolutionary has happened.
We've just eliminated 47 WhatsApp messages.
Although frankly that may be enough.
Now change the numbers.
Not five people.
Fifty.
Five thousand.
Five million.
Suddenly this looks less like holiday planning and more like a new form of organisation.
We underestimate how expensive organisation is
One of the least glamorous forces shaping society is coordination cost.
It takes effort to discover that other people want the same thing.
Then you have to find them.
Communicate.
Agree what you actually mean.
Work through differences.
Collect information.
Divide up the work.
Arrange meetings.
Raise money.
Chase the person who volunteered to do something three weeks ago.
Eventually someone creates a spreadsheet.
At that point civilisation has officially begun.
Companies exist partly because coordination is difficult.
So do councils, charities, trade unions, clubs, political parties and residents' associations.
They convert lots of people vaguely wanting something into something actually happening.
And that requires machinery.
Committees.
Officers.
Membership lists.
Constitutions.
Minutes.
Meetings about the meetings.
This isn't necessarily bureaucracy gone mad.
It's the infrastructure required to organise humans.
Artificial Personal Intelligence could slash the price of that infrastructure.
The group that doesn't exist
Imagine a council is considering a change in your neighbourhood.
Today perhaps you notice the consultation.
Perhaps you don't.
Someone posts it on Facebook.
The residents' association gets interested.
Someone knocks on doors.
Three people end up writing most of the response.
Twenty-seven people sign it.
Several hundred others find out what happened after the decision has already been taken.
Now imagine your personal AI spots the proposal because you've asked it to watch for things likely to matter to you.
It doesn't immediately bother you.
First it works out whether this really intersects with your preferences.
Then it discovers that 1,847 other personal agents have reached a similar conclusion for their owners.
Yesterday, there was no group.
No chair.
No secretary.
No mailing list.
The agents compare concerns.
Perhaps 1,200 people care mainly about traffic.
Four hundred are worried about accessibility.
Three hundred object to losing trees.
Some belong to more than one group.
Then something more interesting happens.
The agents discover that the apparent disagreement isn't quite what it seemed.
Preserve six trees.
Alter the junction.
Add a pedestrian crossing.
Support rises dramatically.
The group hasn't simply organised opposition.
It has searched for agreement.
That may be the more interesting capability.
We normally think collective action means finding people who already agree with us.
AI might help discover what we could agree upon.
The plumbing is beginning to appear
We're nowhere near the society I'm describing.
But some of the plumbing exists.
Agent-to-agent protocols are being developed so independently built AI systems can discover one another, exchange information and coordinate tasks.
MCP does something related, giving AI applications a standard way to connect to tools, data and workflows.
Researchers are already examining multi-principal systems where different agents represent different humans or organisations and have to negotiate between competing interests.
In other words, the family holiday problem is becoming a computer-science problem.
I find that strangely reassuring.
Now give the group buying power
Suppose 40,000 people are looking for a new electricity tariff.
Today they enter the market separately.
Their personal agents might instead discover that they have sufficiently similar requirements and temporarily combine their demand.
We represent 40,000 authenticated customers potentially willing to switch supplier for a twelve-month tariff meeting these conditions. What will you offer?
Nobody has set up a company.
Nobody is chair.
There may not even be a permanent group.
It forms on Monday.
Negotiates on Tuesday.
Buys on Wednesday.
Disappears on Thursday.
The following week some of exactly the same people might form another temporary coalition around broadband.
Or insurance.
Or solar panels.
Or care services.
Buying clubs, mutuals and co-operatives aren't new.
What would be new is making formation almost frictionless.
A temporary institution.
A swarm with a constitution.
Work gets interesting too
Now imagine employees.
Your AI understands your employment contract.
Other employees have their own.
The organisation announces a restructuring.
Today people individually try to understand what it means.
Some contact a union.
Some don't belong to one.
Some know employment law.
Most don't.
Personal agents could identify shared consequences almost immediately.
312 employees appear to be affected by the same proposed contractual change.
187 have authorised anonymous comparison of the relevant clauses.
146 would like independent advice.
Specialist legal advice would cost £2,400, or £16.44 each if shared between those interested.
That doesn't make trade unions obsolete.
Quite the opposite.
Trusted institutions may become more valuable because personal agents need expertise, legitimacy and somewhere for responsibility to land.
But it could make collective organisation possible before a formal organisation exists.
That is different.
Of course, my AI may simply disagree with yours
There is a dangerously cheerful version of all this in which millions of helpful personal agents discover elegant compromises.
Humans have not traditionally been quite so accommodating.
Why would our machines be?
There are twenty concert tickets.
Fifty thousand agents want them.
There are ten charging points.
Fifty cars need one.
A planning decision creates genuine winners and losers.
Better coordination doesn't remove scarcity.
It may make competition much more efficient.
That could be worse.
This is where I think one of the deeper questions about agentic AI appears.
We spend a lot of time asking:
How do we make an AI better at achieving its objective?
We may eventually need to ask:
What happens when millions of very successful AIs pursue incompatible objectives at the same time?
That's a familiar problem.
We call it society.
Institutions probably aren't going anywhere
There is another seductive version of this future where AI dissolves bureaucracy.
No committees.
No councils.
No unions.
No companies.
Just efficient personal agents coordinating everything beautifully.
I don't buy it.
Institutions do more than process information.
They provide legitimacy.
They hold money.
They accept liability.
They resolve disputes.
They survive after individuals leave.
They give us somewhere to put responsibility.
If 5,000 personal agents negotiate the maintenance of a community park and something goes badly wrong, somebody still has to answer:
Who was responsible?
Artificial Personal Intelligence may not abolish institutions.
It may simply change when we need them.
A group could remain fluid while discovering shared interests.
Become more formal when money is involved.
More formal again when contracts are signed.
Eventually it may become something we'd recognise as an organisation.
AI doesn't remove governance.
It may allow organisation to emerge after the need is discovered instead of before it.
And then the organisations get AIs too
So far I've assumed that my personal AI represents me.
But organisations will have them too.
Of course they will.
The supermarket has an agent.
The employer has an agent.
The insurer has an agent.
The political campaign has an agent.
The advertising industry has several thousand.
Suddenly my AI isn't just coordinating with yours.
It's operating in an environment full of systems actively trying to sell, persuade, recruit and organise us.
Perhaps an advert reaches my personal AI and gets the response:
This message is emotionally persuasive but contains no new information relevant to Martin's stated purchasing criteria.
Somewhere, an advertising executive quietly shudders.
Or perhaps the seller's AI becomes extremely good at discovering precisely which arguments my AI will accept.
We've not abolished persuasion.
We've mechanised the negotiation around it.
And this is where Part One comes back.
If my agent belongs to me, at least I can ask whose interests it represents.
If my supposedly personal AI is actually supplied by the same platform that hosts the seller, provides the advertising and runs the marketplace...
well.
We have seen versions of that before.
The organisational superpower
The personal computer gave individuals access to computation once available mainly to institutions.
The Web gave individuals publishing and communication power.
Artificial Personal Intelligence might give us something else.
Organisational power.
The ability to discover people with compatible interests.
Compare constraints without exposing everything about ourselves.
Pool buying power.
Commission expertise.
Divide work.
Negotiate.
Form something temporary and dissolve it when the job is done.
Perhaps we're asking the wrong question when we obsess over which jobs AI will automate.
A more interesting one might be:
Which institutions exist partly because humans are expensive to coordinate?
Some will remain essential.
Some may become vastly more powerful.
Some may discover that people can organise around them instead of through them.
And entirely new forms of organisation may become possible simply because assembling them becomes cheap.
Which brings us to the bit I've deliberately avoided.
Politics.
A company facing a £100 million regulatory cost has every reason to organise, employ specialists and lobby government.
Ten million citizens each losing £10 generally don't.
The interests are worth exactly the same amount.
One side is organised.
The other is ten million people getting on with their lives.
Now give those ten million people personal AIs capable of noticing that they share an interest.
What happens when the group can form before anyone has consciously decided to join it?
What happens to lobbying?
What happens to political parties?
And eventually, what happens to voting?
That's where Artificial Personal Intelligence stops being just a technology story.
And where Part Three gets rather dangerous.
Comments
Post a Comment