Some see the beginnings of another intelligence.
Some see a glorified autocomplete.
Some see an existential threat.
Some see an office tool.
I suspect the truth is both less dramatic and more interesting.
Perhaps AI’s great trick isn’t that it thinks.
Perhaps it simply helps us avoid ignorance at speed.
That sounds rather underwhelming compared with artificial general intelligence, machine consciousness and the imminent replacement of humanity. But the more I use these systems, the more I think it may be the more important idea.
Ask an AI about something you know extremely well and the magic often disappears.
You read the answer and think: yes, fair enough. Mostly right. A bit obvious. Occasionally wrong.
Sometimes very confidently wrong.
But then ask it about something just outside your own field.
Suddenly it becomes much more interesting.
You might start with change management and find yourself moving through organisational learning, absorptive capacity, Senge, team mental models, psychological safety, project governance and systems thinking.
None of those ideas is new.
What changes is the cost of putting them in the same room.
And I’m beginning to think that may be the real revolution.
An enormous lessons-learned machine
Organisations have always been terrible at lessons learned.
We dutifully hold a session at the end of the project.
We identify what worked.
We identify what didn’t.
Someone writes it down.
Then it enters the great corporate archaeological record, ready to be rediscovered shortly after the next project has made exactly the same mistakes.
I’ve increasingly come to see LLMs as an accidental answer to this problem.
Not because they contain a neat database of lessons learned. They don’t.
But because they have been trained across enormous quantities of human writing describing what people have tried, what failed, what succeeded, what good practice might look like and what academics have subsequently decided to call it.
Ask the right question and what comes back is often something like:
This looks rather like things humans have encountered before. Here are the patterns.
That isn’t intelligence in the human sense.
It may not need to be.
It’s an extraordinarily fast approximation of organisational memory.
And because it works across disciplines, it can sometimes do something our actual organisational memory struggles with.
It can synthesise.
A problem that looks like a project management issue might also be an organisational psychology problem.
A technology implementation might actually be an absorptive capacity problem.
A governance problem might partly be a team mental models problem.
The AI doesn’t necessarily understand any of these things.
But it can put them next to each other quickly enough for us to notice.
That distinction feels important.
It may raise the floor before it raises the ceiling
There’s also something slightly less glamorous going on.
AI may help normalise reasonable practice.
Perhaps even good practice, provided the people using it have enough training to know the difference.
An experienced project manager doesn’t need ChatGPT to explain what a risk register is.
Someone running their first significant project might.
An experienced change practitioner may already know to think about stakeholder readiness, organisational capacity and behavioural reinforcement.
Someone who has accidentally become responsible for a change programme might not.
The technology gives that person access to a reasonable approximation of what experienced practitioners have learned.
That doesn’t magically make them an expert.
But it does make ignorance slightly less defensible.
This may prove to be one of AI’s biggest effects.
We spend a great deal of time asking whether AI can surpass the very best humans.
Perhaps we should spend more time asking what happens when millions of ordinary decisions are pulled upwards towards competent practice.
That could be far more disruptive.
Not because AI creates brilliance everywhere.
Because it makes avoidable stupidity more difficult.
Which brings us to hallucinations
Of course, there’s an obvious problem with treating AI as some enormous repository of human practice.
It makes things up.
Worse, it makes things up using exactly the same confident grammar it uses when it is correct.
The paper Hallucination Stations: On Some Basic Limitations of Transformer-Based Language Models pushes this further. Its argument is that there are fundamental computational limits to what transformer-based systems can reliably perform and verify, particularly as tasks become more complex.
I find that quite reassuring.
Not because hallucinations are good.
Because it moves us away from the strange assumption that sufficiently fluent language must eventually turn into a mind.
For virtually the whole of human history, language has been pretty good evidence that there is somebody home.
A sentence implies a speaker.
An argument implies somebody making it.
A joke implies somebody understood why it was funny.
We now have a technology that breaks that ancient connection.
It can produce the outward appearance of thought without giving us any particular reason to believe that the thing producing it is thinking in a meaningful human sense.
That is psychologically difficult.
We instinctively infer a mind behind the words.
Perhaps we need to learn not to.
The real danger may therefore be less that AI hallucinates and more that we hallucinate the intelligence behind the AI.
We see fluent output and supply the consciousness ourselves.
And then there’s creativity
This is where my own argument gets untidy.
I’m still not convinced AI is creative in the sense that humans are creative.
Yes, it produces novel things.
But novelty isn’t necessarily creativity.
There’s randomness built into the generation process. Given a sufficiently enormous possibility space, occasionally the dice will land somewhere interesting.
You could argue that this is creativity.
I’m not entirely convinced.
A kaleidoscope can generate a pattern that has never existed before. We don’t normally imagine that the kaleidoscope has had an artistic breakthrough.
Yet I also have to be careful here because human creativity is itself heavily recombinant.
We borrow.
We imitate.
We misremember.
We combine things that previously sat apart.
The Beatles didn’t invent music.
Picasso didn’t invent painting.
Einstein didn’t invent mathematics or physics.
Creativity often appears when existing things are combined in a way that somebody hadn’t previously considered.
And AI is extremely good at combination.
So perhaps the distinction isn’t between creative and uncreative.
Perhaps it is between creative output and creative agency.
An AI can generate something unexpected.
It can combine ideas in a way I hadn’t considered.
It can even produce something I find beautiful.
But it doesn’t appear to care.
It doesn’t wake up annoyed by yesterday’s answer.
It doesn’t become obsessed with an idea.
It doesn’t decide everybody else is wrong and spend ten years trying to prove it.
It doesn’t have the strange mixture of vanity, curiosity, insecurity, boredom and bloody-mindedness that seems to power an awful lot of human creativity.
It generates possibilities.
We provide the wanting.
And I suspect that difference matters much more than whether it can pass a creativity test.
Artificially cheap synthesis
So perhaps we’ve been naming the wrong thing.
The transformative capability isn’t necessarily artificial intelligence.
It might be artificially cheap synthesis.
Until very recently, joining several bodies of knowledge together was expensive.
First you had to know they existed.
Then you had to find them.
Then you had to read enough to understand them.
Then, if you were lucky, you spotted the connection.
That could take years.
An LLM can sometimes move across those boundaries in seconds.
Not reliably.
Not perfectly.
And certainly not with the authority we sometimes project onto it.
But cheaply enough that asking the question becomes almost free.
That changes something.
The expert doesn’t need to remember everything.
The novice doesn’t have to begin entirely from zero.
The project manager can ask what another discipline might say about the problem.
The change practitioner can bring organisational theory into a live delivery conversation without disappearing into a library for a fortnight.
The policymaker can ask what analogous situations have taught other organisations.
The value isn’t necessarily in the answer.
It is in reducing the distance between one piece of human knowledge and another.
Perhaps we should stop asking whether it thinks
This is why I’m becoming less interested in the argument about whether AI thinks.
Maybe it doesn’t.
I currently see no compelling reason to believe that an LLM thinks in any meaningful way.
But perhaps that question has been allowed to dominate because it makes for a better story.
A machine that thinks is Frankenstein.
HAL 9000.
Skynet.
The computer we actually have may be stranger.
Something without understanding that can nevertheless help a human understand more.
Something without expertise that can surface expertise.
Something without creativity, perhaps, that can be extremely useful during a creative process.
Something that knows nothing and yet makes an astonishing amount of what humanity knows available almost instantly.
That isn’t really a new mind.
It is something closer to a new interface to accumulated human knowledge.
And suddenly the disagreement around AI makes more sense.
If you expect intelligence, you notice everything it cannot do.
If you expect a search engine, you’re astonished by what it can do.
If you treat it as an oracle, it is dangerous.
If you treat it as a rapid, fallible synthesis engine, it becomes rather useful.
Maybe that is enough.
We don’t need machines to become gods for this technology to matter.
They may simply make it much harder for us to say:
I didn’t know anyone had thought about that before.
AI probably won’t abolish ignorance.
But it may make remaining ignorant of what humanity already knows increasingly optional.
And that could turn out to be quite a revolution.
Comments
Post a Comment