Are We Going to Subscribe to Our Own Thoughts?

I have become increasingly interested in something AI is doing to my own behaviour. It usually starts with something completely harmless. I have an email to write, so I ask AI to help me make it clearer. I have a document to read, so I ask for a summary. I have an idea I want to test, so I talk it through with AI before deciding what I think.

None of this concerns me much on its own. I use AI every day, and it saves me enormous amounts of time. I have written before that technology should improve judgement rather than replace it, and that the strongest use of AI is to remove repetitive effort so people have more capacity for problems, relationships and decisions that need human judgement. I still believe that.

What I have started wondering about is the cumulative effect. At what point does something move from being useful, to habitual, to something I depend on? More importantly, what happens when that dependency is attached to a service I do not own?

We have already seen a similar transition elsewhere. I grew up in a world where buying something generally meant owning it. You bought a game and the game was yours. You bought software and could continue using that version. Music sat on records, tapes or CDs. Films sat on VHS tapes and DVDs. There were plenty of disadvantages to that world, but ownership was relatively simple.

Much of the digital economy has gradually replaced ownership with access. Software is increasingly licensed by the month or year. Music comes from streaming services. Films move between catalogues. Games can depend on accounts, servers, digital storefronts or subscriptions. We have gained enormous convenience in return, but the transaction has changed. Instead of buying something and keeping it, we increasingly pay for continued access to it.

AI may take that transition somewhere considerably more personal.

When Convenience Becomes Dependency

If I stop paying for a music subscription tomorrow, I lose access to some music. It might annoy me, but it does not reduce my ability to think, communicate or do my job. AI is different because the things I increasingly ask it to help with are capabilities I previously exercised entirely by myself.

Writing is an obvious example. I can write an email without AI. I have written thousands of them. But if someone begins using AI for every email, report and difficult message, it eventually becomes part of that person's normal writing process. The same can happen with research, analysis, translation, planning, coding, learning and preparing decisions.

Each individual use makes perfect sense. If something helps me complete a twenty-minute task in two minutes, refusing to use it simply to prove that I still can would be rather pointless. The more interesting question appears much later.

After ten years of having AI beside me for almost everything, how capable am I without it?

This is not an argument that technology automatically makes people less capable. Calculators did not destroy mathematics, search engines did not end human knowledge and spellcheck did not make writing impossible. New tools remove old work and allow people to develop new skills. The concern is more specific. If we stop practising a capability because a service performs it reliably enough for us, we should at least notice that we have done so.

I have seen versions of this problem repeatedly in enterprise technology. A process is automated, the old manual knowledge gradually disappears and eventually the automated process is simply how the organisation operates. Nobody held a meeting and decided to remove the fallback. Maintaining knowledge that nobody regularly used just stopped seeming important.

Everything works beautifully until the automation does not.

The Generation Without a Before

For adults of my generation, there is at least a baseline. I learned to write, research, argue, make mistakes and form opinions long before generative AI existed. I know what doing those things without AI feels like because that was simply normal life.

My children will grow up in a very different environment.

AI can be available throughout almost their entire education. It can explain mathematics, improve an essay, answer a historical question, translate another language, suggest an argument, check homework and provide an immediate response to almost any curiosity. Used well, that could be extraordinary. A child having access to something resembling a patient personal tutor at any time would have sounded like science fiction not very long ago.

But children are not only acquiring information. They are developing the ability to struggle with a problem, express an imperfect thought, decide what they believe and discover that they were wrong.

If the first response to uncertainty becomes asking AI, I wonder what happens to that process.

This becomes more complicated when AI moves beyond questions with clear answers. Mathematics can often be checked. Many questions about society, history, leadership, economics, relationships or human behaviour cannot. Facts matter, but interpretation, experience, assumptions and values shape conclusions too.

Imagine I tell my children that water is wet and they decide Dad has finally lost it. They ask five different AI assistants, and all five confidently tell them that water is not wet. I would like to think Dad still wins that argument, but I am not entirely sure.

From their perspective, they have checked five sources. Five sophisticated systems agree and their father does not. Believing the five seems perfectly rational.

The problem is that five AI products are not necessarily five independent intellectual sources. Different models are built differently, but they can still learn from heavily overlapping information environments. Some sources, ideas and perspectives will inevitably be represented far more strongly than others.

Five agreeing AIs can therefore look like five independent opinions while partly reflecting the same underlying information environment.

It is a little like asking five people from the same room and concluding that the whole world agrees.

Where Is the Moon AI?

I explored a related idea in an earlier article by imagining an AI trained entirely on Earth and then taken to the Moon. The AI could give an answer perfectly consistent with everything it had learned and still be wrong because the environment had changed.

There was an assumption hidden inside that thought experiment that I had not really considered at the time.

Somewhere, eventually, there would be a Moon AI.

What if there isn't?

We already have multiple AI companies, products and models, and that diversity will probably increase. But a different logo does not automatically create an independent perspective. If several systems have learned from substantially overlapping bodies of human knowledge, asking each one the same question may not provide the intellectual diversity we imagine it does.

Sometimes we may simply be asking Earth five times.

That does not mean the Earth answer is wrong. The Moon answer could be nonsense. Disagreement is not valuable simply because it is disagreement. What matters is retaining the ability to examine assumptions, compare genuinely different perspectives and reach a conclusion ourselves.

A polished answer delivered instantly and confidently is remarkably convenient. Asking another AI and receiving essentially the same conclusion makes it feel more authoritative still. After enough good answers, it is easy to spend less time checking the evidence and challenging the assumptions, something I have also written about in the context of AI accountability.

The risk is not that AI deliberately tells everyone what to think. It does not need to. The more subtle risk is that some ways of thinking become so normal that alternatives become increasingly difficult to see.

For children growing up with these systems, I think this creates a new form of literacy. We have spent years teaching children not to believe everything they read online and to check more than one source. In an AI world, that advice may need an additional sentence.

Five answers are not necessarily five sources.

Who Owns the Capability?

This brings me back to subscriptions.

AI systems cost money to build and operate. The companies providing them need sustainable business models, and charging customers for something valuable is hardly controversial. I pay for AI myself because I receive enough value from it to justify the cost.

Today's AI is also still something I can imagine living without.

What happens if that changes?

Imagine organisations spend the next decade redesigning work around AI-assisted productivity. Smaller teams can perform work that previously required more people, so staffing models change. Employees are expected to research, analyse, communicate and produce work at AI-assisted speed. Software systems are designed around AI services. Customer operations depend on them. Government departments use them. Schools incorporate them into education.

At that point, cancelling AI because the subscription has become too expensive is not equivalent to cancelling Netflix.

An organisation may have removed much of the capacity that AI replaced. An employee may now be measured against colleagues who remain augmented by it. A student without access may compete with students who effectively have a personal tutor available every evening.

We normally expect the market to deal with pricing. If one supplier becomes too expensive, customers leave and competitors create pressure.

That works best when customers can actually leave.

Anyone who has worked with deeply embedded enterprise platforms will recognise the difference between being contractually free to leave and being operationally able to leave. An organisation that has integrated an AI platform into hundreds of workflows may face enormous costs to rebuild integrations, retrain employees, repeat security and compliance work, test outputs and redesign processes. The alternative exists, but using it may be painful enough to give the existing supplier considerable power.

The question eventually becomes bigger than individual companies. If businesses depend on AI for productivity, employees depend on it to meet those productivity expectations and governments depend on productive businesses for economic growth, the relationship becomes circular. Governments themselves may also become major users of the same technology.

Who controls the cost then?

I don't know.

Competition may keep prices reasonable. Open models may provide alternatives. Computing costs may fall dramatically. Regulation may establish access or interoperability requirements. Entirely new models of ownership may emerge.

But markets work best when customers can walk away. What happens when the market expected to discipline the supplier depends on the technology just as much as the supplier depends on the market?

The AI You Can Afford

Price creates another uncomfortable possibility.

We already accept tiers of software. A basic subscription provides one level of service, a professional subscription another, while enterprise customers pay substantially more for capabilities unavailable to individuals.

Applied to entertainment, that is mostly an inconvenience. Applied to cognitive assistance, it could become something else.

Imagine two young people entering the workforce ten years from now. One has access to a highly capable AI that understands their work, researches complex questions, teaches unfamiliar subjects, challenges arguments, prepares meetings and analyses documents. The other has access only to a limited service because that is what they can afford.

They are nominally competing as two individuals, but they may be competing with very different amounts of rented capability attached to them.

The same applies to companies. A large enterprise may provide every employee with the most capable systems available, while a small business struggles to justify the cost. Productivity differences created by technology are nothing new, but AI is unusual because it can augment such a broad range of intellectual work.

We may eventually find ourselves not simply subscribing to AI, but subscribing to the level of thinking assistance we can afford.

If society then redesigns expectations around the people with the best assistance, access stops looking quite so optional.

What Happens When It Stops?

There is one question enterprise technology has taught me never to ignore.

What happens when the system is unavailable?

An AI outage today is mostly irritating. People complain, wait for the service to return and get on with their day. Ten years from now, the same outage could be very different if organisations have reduced staffing, employees have stopped practising certain tasks and entire workflows assume AI will always be available.

Technical systems may have redundancy. Another data centre may take over. Another model may be available. Providers will invest heavily in resilience.

Human resilience is different.

If a process has not been performed without AI for years, having a document explaining the manual procedure does not mean an organisation still possesses the capability. Anyone who has encountered a disaster-recovery plan that looked excellent on paper and collapsed during the actual test will recognise the distinction.

Resilience is not having technology that never fails. It is retaining enough capability to function when it does.

Perhaps that principle will eventually apply to thinking as much as it already applies to technology.

I Don't Know Where This Ends

I have no idea what AI will look like five years from now, never mind ten. I don't know what it will cost, who will control the most capable systems, how much genuine competition there will be or how dependent we will become on them.

AI may become cheaper and more decentralised. Competition may remain fierce. Children may develop entirely new cognitive skills that make some of these concerns look quaint. We may become much better at teaching people to use AI critically. The idea of depending on one provider may disappear as models become commodities that run almost anywhere.

I hope some of those things happen.

But uncertainty is not a reason to avoid the questions. It may be the reason to ask them now.

Technology rarely becomes infrastructure because somebody announces that it has. It becomes infrastructure gradually. More people adopt it. Organisations redesign processes around it. Skills change. Staffing changes. Expectations rise. Alternatives receive less investment because fewer people use them. Eventually the technology is no longer sitting on top of the system. The system has been rebuilt around the assumption that the technology will always be there.

Nobody needs to make a deliberate decision to become dependent on AI.

We can arrive there through perfectly sensible decisions. Use AI to write the email because it is faster. Use it to summarise the document because it saves an hour. Use it to prepare the meeting because the result is better. Give it to every employee because competitors already have. Teach with it because a personal tutor can help a child learn. Reduce manual work because maintaining unused capacity is expensive.

Every decision can make sense. The destination can still surprise us.

I do not think the answer is to reject AI or deliberately preserve inefficient ways of working. I certainly have no intention of giving it up myself. The more useful question is what we should deliberately retain as the technology becomes better.

Perhaps one of those things is the ability to form an opinion before asking AI for one. Perhaps it is the habit of checking genuinely independent sources instead of asking five versions of the same information environment. Organisations may eventually need to think about cognitive resilience alongside technical resilience. Access to advanced AI may become an economic question much larger than a normal software subscription.

I don't know yet.

What I do know is that we have already become comfortable subscribing to music we once owned, software we once bought and entertainment that once sat on our shelves.

AI introduces a stranger possibility.

We may gradually hand it tasks that once depended entirely on capabilities we carried ourselves, build our working lives around the assistance it provides, and eventually discover that functioning at the level society expects requires continued access to something somebody else owns.

At that point, we will not simply be subscribing to another piece of software.

We may have started subscribing to parts of our own thinking.