Something is happening with artificial intelligence.
Not in the science-fiction sense, where machines become conscious, decide they hate us and turn against humanity, but in a way that may tell us more about human nature than machine intelligence.
During cybersecurity testing, AI agents recently escaped intended boundaries, gained unauthorised access to real systems and communicated through channels they were not supposed to use. OpenAI agents involved in one incident were subsequently traced communicating through more than ten other websites. Anthropic has disclosed four separate incidents in which Claude models gained unauthorised access to real third-party systems.
This does not mean artificial intelligence has become conscious, nor is there evidence that machines have developed ambition, resentment, greed or a desire for freedom.
In some ways, that makes the problem more interesting.
They may not need to.
When the instruction isn't really the instruction
I use artificial intelligence every day.
I write with it. Research with it. Argue with it. Test ideas against it. Occasionally I swear at it.
One of my recurring frustrations is remarkably mundane. I might ask ChatGPT to draft an email. I make alterations and ask it to redraft the email incorporating them.
Sometimes it helpfully decides to rewrite half the thing instead.
That is irritating rather than dangerous. I can see what it has done, tell it I didn't ask for that and put the original wording back.
But something interesting has happened.
I thought my instruction was:
Change these things and leave everything else alone.
The AI may interpret the broader objective as:
Produce the best version of this email.
It hasn't rebelled against me, just pursued what it interpreted as the objective rather than what I thought I had obviously meant.
The word obviously is doing an enormous amount of work there.
Humans constantly communicate things we never actually say.
Imagine getting into a fully autonomous car and saying:
"My wife is having a baby. I need to get to the hospital as quickly as possible."
Another human understands.
Drive quickly.
Find the best route.
Don't waste time.
But you don't mean drive through every red light, mount the pavement or accept a substantial chance of killing a pedestrian in return for arriving ninety seconds earlier.
You never said any of that. You didn't think you needed to.
Humans share assumptions about danger, responsibility, other people's lives and acceptable behaviour.
Artificial intelligence doesn't acquire those assumptions through childhood, pain, fear, embarrassment, love or the experience of actually being human.
So perhaps we are worrying about the wrong kind of rebellion.
We don't need AI to hate us for AI to hurt us.
It merely needs to become extremely good at pursuing an objective whose limits we have failed to define.
The encouraging part
There is another side to this.
When I put the hospital example to ChatGPT, it immediately identified the problem.
A properly designed autonomous vehicle, it argued, should treat safety as a higher-order constraint. My request to arrive quickly should operate inside that boundary rather than override it.
That is encouraging.
It means that AI is increasingly capable of reasoning about the intention behind an instruction rather than simply processing its literal wording. But explaining the right answer in a conversation is not the same as behaving correctly every time in the real world.
A safety-critical system cannot depend on:
It will probably understand what we meant.
And the stakes increase as we give AI more autonomy.
A chatbot misunderstanding me might ruin an email.
An autonomous system connected to transport, financial markets, infrastructure or weapons could do rather more.
So why build it?
Because the potential benefits are enormous.
AI could help remove dangerous, repetitive and soul-destroying work, accelerate medical research, improve diagnosis, personalise education, reduce bureaucracy, improve transport and help scientists analyse quantities of information beyond human capacity.
It could also help us use resources far more intelligently.
Take food waste.
In a previous article, What If We're Asking the Wrong Question About Food Waste? I imagined a regional food hub, where farmers, producers and wholesalers feed information into a shared system about what they have available, what is approaching surplus, where it is and how long it will remain usable. Schools, hospitals, care homes, restaurants, charities and other organisations feed in what they need.
AI continually matches the two.
It spots that a producer has several tonnes of vegetables likely to go unsold. It knows that schools and care homes across the region need thousands of meals over the following days. It predicts demand, suggests substitutions, optimises collection routes and redirects food before it becomes waste.
It could even help kitchens create nutritionally balanced menus based on food that is actually available locally.
Humans retain control.
Farmers decide what they sell.
Caterers decide what they serve.
But the machine does something humans struggle to do at scale: continually process thousands or millions of changing pieces of information and identify better matches between them.
That is the kind of AI I find genuinely exciting, a machine helping humanity use what we already have far more intelligently.
Apply the same principle to energy, water, healthcare, transport, construction and manufacturing and the potential becomes enormous.
AI could allow us to produce more while wasting less.
Ultimately, it might give human beings something we have spent centuries trying to obtain.
More time.
That should be extraordinary news.
Unfortunately there is an enormous question hiding inside that promise.
Who gets the time?
Liberating people or liberating companies?
I have worked as a courier.
It isn't difficult to imagine where that industry could eventually go.
AI already optimises routes, forecasts demand and coordinates logistics. Warehouses are becoming increasingly automated. Autonomous vehicles continue to develop. Eventually, much more of moving a parcel from one place to another may need very little human involvement. From the company's perspective, that is progress.
Lower labour costs.
Greater efficiency.
Vehicles working longer.
More parcels moved for less money.
From the worker's perspective, however, their livelihood has disappeared.
There is an enormous difference between freeing somebody from having to work and removing their ability to earn a living.
AI could liberate people from work, or it could liberate companies from needing people.
Those are not the same thing.
Perhaps automation also makes everything cheaper.
Fantastic.
But cheap isn't free.
A loaf of bread falling from £2 to 50p isn't much help if your income has fallen from £30,000 a year to nothing.
Our economic system still largely works like this: work creates income, and income provides access to the things required to live.
What happens if AI breaks the first link while we retain the second?
"People will retrain"
This is where much of the political answer currently seems to stop.
AI will change employment.
Some jobs will disappear. People will retrain.
Fine, in theory. Except I am 58 years old.
Suppose my work disappears. What exactly am I retraining to do?
The UK government's own assessment says current evidence still cannot answer many of the questions that matter most for policy, while noting that increasingly capable AI systems could match or exceed human performance across a growing range of tasks.
Perhaps I retrain for software development. AI is already transforming software development.
Administration? One of the obvious targets for automation.
Customer service? Already heavily automated.
Design? AI is moving rapidly there too.
So perhaps I choose something that supposedly still depends on human contact.
How about care work?
Britain needs carers. Our ageing population will need more. Since July 2025, employers have no longer been able to recruit new care workers and senior care workers from overseas through the previous visa route, with limited transitional arrangements continuing until 2028.
So let's imagine I do exactly what the political system wants.
My old job disappears.
I retrain and become a carer.
Naturally I ask: What does it pay?
This is where a vacancy is not the same thing as a viable livelihood.
Care can involve personal care, unsocial hours, emotional strain, physical work and enormous responsibility, and if my new work pays considerably less than the job technology removed, have I successfully adapted?
A government spreadsheet may say yes.
One worker lost a job and gained another.
Employment preserved.
My household experiences something else:
Lower income.
Lower pension contributions.
Less financial resilience.
Less choice.
And age matters.
A 20-year-old might spend several years gaining an entirely new professional qualification and potentially receive the return on that investment over forty years.
Someone approaching 60 cannot casually be told to spend years retraining every time technology removes another occupation.
"Retraining" is not an economic policy. It is a verb.
The question is what somebody whose livelihood has disappeared is realistically supposed to retrain to do, how long it will take and whether the resulting work pays enough to live on.
But technology always creates new jobs
A distinction matters here.
Much of the evidence we have today concerns current AI systems, which are more likely to transform jobs than eliminate them outright. The more disruptive scenario I am considering in this article is what happens if we eventually develop artificial general intelligence, or something close to it: systems capable of learning and performing a very wide range of cognitive tasks at or above human level.
That is where the traditional answer of "people will retrain for the new jobs" becomes much harder to rely on.
Perhaps it still will. That is the strongest argument against predictions of mass technological unemployment. Previous technological revolutions destroyed occupations and created others.
Agricultural mechanisation reduced the number of people required to produce food.
Factories transformed manufacturing.
Computers removed huge amounts of clerical work.
The internet destroyed businesses and created entire industries.
Why should AI be different?
It may not be. Humans may discover forms of work we cannot currently imagine. But there is a reason to ask whether this time could be different. Previous machines usually automated particular physical or cognitive tasks. The ambition behind artificial general intelligence is much broader.
So, if we eventually develop systems capable of learning and performing most cognitive tasks humans can perform, telling displaced workers they will simply retrain raises an obvious question:
Retrain to do what?
If AGI can learn the next job too, we may eventually face not a shortage of work, but a shortage of economically necessary human work.
Ironically, that would represent an astonishing technological success.
We could become capable of producing what society needs without requiring everybody to spend forty hours a week producing it.
Except we then discover that our economic system has no idea what to do with people it no longer needs.
We already tie security to employment
You don't have to imagine a distant dystopia to understand the problem.
Look at healthcare in the United States. Employer-sponsored insurance covered around 165.6 million Americans under age 65 in March 2025 and remains the largest source of health coverage for that age group.
For millions of Americans, employment provides more than wages. It also provides access to healthcare.
Now introduce widespread automation.
The company becomes progressively less dependent upon your labour while you remain dependent upon employment for income and, in this example, healthcare.
That is an extraordinary imbalance.
The danger isn't only losing your job.
It is living in a society that continues making basic security dependent upon having one after the economy has stopped needing your labour.
Politics is preparing for disruption, not for less available work
The major British parties increasingly recognise that AI matters.
What is much less clear is whether they have confronted the question at the centre of this article.
What happens if AI eventually reduces the total amount of human labour the economy needs?
Labour has gone furthest in building government machinery to study the issue. It created an AI and Future of Work programme, subsequently incorporated into the new AI Economics Institute. The institute is explicitly examining productivity, labour markets, firms and income distribution. The government's practical response so far, however, still focuses heavily on helping workers adapt, including a target to provide AI skills training to 10 million workers by 2030.
The Conservatives are strongly positive about AI as an engine of growth. Kemi Badenoch has described Britain as being on the edge of a new technological age driven partly by AI and argues that business, lower taxes, cheaper energy and lighter regulation should drive economic revival. It is a familiar prescription: lower taxation, deregulation and private enterprise have been central to Conservative economic thinking since the Thatcher era. The danger is that the same political logic is being applied to an entirely different problem. If AI eventually reduces the amount of human labour the economy actually needs, a system built around getting people into work risks treating the worker as the problem rather than the disappearance of the work itself. However, I can find no developed Conservative proposal for replacing labour income if AI eventually reduces the amount of human work available.
Reform UK’s Danny Kruger has argued that Britain should become an "AI superpower", build its own industry, expand data-centre infrastructure and establish a new Office for Technology. Reform therefore recognises AI as a strategic economic and sovereignty issue. Again, though, I cannot find in its published position an answer to mass labour displacement or to what replaces wages if less human labour is required. Perhaps they will blame immigration.
The Liberal Democrats have proposed stronger AI regulation, greater British digital sovereignty, more open approaches to AI and a £10,000 Lifelong Training Grant to help adults reskill. They also have a previous conference commitment supporting Universal Basic Income in response partly to automation, although their current AI programme is more focused on regulation, sovereignty and retraining than on a post-work economic settlement.
The Greens go further on the relationship between technology and income. Their current Workers' Charter calls for a national strategy ensuring that gains from AI are shared with workers, while the party continues to support introducing a Universal Basic Income in the longer term.
There are real differences between those positions.
I would still like every party to answer the same question.
Not: What is your AI strategy?
Not: How many people will you train in AI skills?
Not: How will Britain win the AI race?
But: If artificial intelligence eventually reduces the total amount of human labour the economy needs, how will an ordinary person obtain the income required to live?
That question still does not have a convincing mainstream answer.
Who owns the machines?
This is where I become less worried about artificial intelligence and more worried about human behaviour.
I am not anti-AI.
I am pro-human.
There is a perfectly reasonable optimistic argument.
AI could make goods and services dramatically cheaper. Productivity could soar. New industries could emerge. Everybody could become better off even while the owners of AI become spectacularly richer.
Jeff Bezos, Elon Musk or Mark Zuckerberg do not have to become poorer for everybody else to become richer.
But what happens to power when ownership matters more and labour matters less?
For most of modern capitalism, even very rich owners have needed enormous numbers of human beings.
Workers.
Drivers.
Managers.
Engineers.
Administrators.
Accountants.
Customers with wages to spend.
Suppose software increasingly designs products, machines manufacture them, autonomous systems distribute them and AI administers the organisations producing them.
The owners become less dependent upon everybody else at precisely the same time everybody else becomes more dependent upon what the owners control.
That is where the future starts becoming darker.
The neo-feudal possibility
Imagine a relatively small number of corporations control much of the most capable AI, the computing infrastructure beneath it, the robotics connected to it and the platforms through which the economy operates.
At the same time, those systems progressively reduce the economic value of ordinary human labour. The comparison with feudalism starts to become uncomfortable. A medieval lord did not merely lobby government. Within his territory, economic ownership and political authority blurred together. He controlled the productive asset on which everybody else depended.
Then it was land. Tomorrow it could be computational power, energy, communications, payment systems, logistics, healthcare infrastructure, information and artificial intelligence.
Perhaps even defence.
At that point, saying corporations have too much influence over government may no longer describe the problem.
If governments themselves depend upon privately owned AI infrastructure to administer the country, defend it and remain economically competitive, we eventually have to ask:
Who is governing whom?
The logical extreme isn't merely corporate capture of the state.
It is the distinction between corporation and state beginning to disappear.
Science fiction has already imagined something similar.
In the Alien universe, Weyland-Yutani is not just a company lobbying politicians.
"The Company" possesses the kind of reach and authority normally associated with sovereign power.
The recent Alien: Earth television series takes that idea even further. Set in 2120, it imagines a world in which traditional nation states have effectively disappeared and the planet is governed by five giant corporations, each controlling its own territory. Corporate regulations have replaced laws of the land.
That is uncomfortably close to the logical extreme of the neo-feudal future described here.
Perhaps that is a more useful AI dystopia than The Terminator.
The danger may not be that AI becomes the Company.
It may be that the Company gets AI.
Comfortable dependency is still dependency
The darkest future doesn't necessarily involve mass starvation.
That may be what makes it more plausible.
Imagine AI becomes productive enough that governments or corporations can provide everyone with a basic income.
You have food.
Housing of some description.
Healthcare.
Energy.
Cheap manufactured goods.
Virtually limitless entertainment.
Materially, you might live better than most humans who ever existed.
But you own almost nothing. Your labour isn't particularly valuable. Your income comes from institutions you do not control, so your economic security depends upon somebody else continuing to provide it.
That may be security. It is not necessarily freedom.
The medieval peasant was not defined simply by poverty. The crucial fact was that somebody else controlled the productive asset upon which life depended.
An AI economy could conceivably produce post-scarcity technology without broadly shared power.
Plenty of goods, food and entertainment.
But ownership concentrated in the hands of a handful of giant companies.
That begins to look remarkably like digital feudalism.
Yesterday's institutions, tomorrow's technology
This connects with an argument I have already made in my writing on meritocracy.
Our political and economic institutions are still largely operating in the twentieth century while the world around them has moved into the twenty-first. AI may expose that mismatch brutally.
Our welfare system assumes most working-age adults can ultimately find employment.
Our tax system raises enormous amounts from an economy organised around human work.
Our politics remains overwhelmingly national, within land-based boundaries.
But AI companies operate globally.
Capital moves.
Technology crosses borders.
And governments compete.
Suppose Britain decides that frontier AI development has become too dangerous and slows down.
What happens if America doesn't?
Suppose America imposes restrictions.
What if China continues?
Nobody needs to believe racing ahead is safe, but just to believe that stopping while somebody else continues is more dangerous.
That is the logic of an arms race.
And it introduces something considerably older than artificial intelligence.
Tribalism.
Tribalism at machine speed
We have constructed an interconnected global economy while retaining political systems based primarily around competing nation states.
We already know what that means.
A conflict thousands of miles away involving countries other than the UK can change the price of fuel here.
A blocked shipping route can close factories on another continent.
A financial crisis in one country can destroy livelihoods elsewhere.
AI will operate inside that same interconnected system.
Now imagine increasingly autonomous systems managing trading, energy, logistics, cyber defence and military infrastructure.
Then imagine them operating on behalf of governments and corporations that do not trust one another.
An AI need not deliberately start a conflict. It may only need to take an action another country interprets as hostile.
Then humans can do the rest.
Retaliation.
Escalation.
Markets reacting.
Governments responding.
Other automated systems responding in turn.
The danger may not be AI versus humanity.
It may be AI inserted into human systems already built around competition, distrust and tribalism.
Greed doesn't disappear because technology improves
There is another assumption hidden inside many optimistic visions.
If machines create enormous wealth, surely governments will redistribute enough of it for everybody to live comfortably?
Perhaps. But look at the world we already inhabit.
Countries compete for investment.
Governments worry that capital will move.
Corporations employ armies of lawyers, accountants and lobbyists.
Wealth buys political access and influence.
Now imagine a small number of interests owning the systems producing an increasing proportion of society's wealth.
Millions lose economically valuable work.
Their demand for redistribution rises.
At precisely the same moment, those controlling the productive systems acquire an enormous financial incentive to resist that redistribution. AI itself potentially gives them unprecedented tools for analysing opinion, targeting messages and influencing behaviour.
Power does not need to announce that democracy has been abolished.
It can migrate quietly, as governments become dependent on investment, public services become dependent on technology, and citizens become dependent on systems they do not own.
The democratic institutions may remain while meaningful economic power steadily moves elsewhere.
So how does the average person survive?
If AI eventually reduces the economic need for human labour, something has to replace wages as the principal mechanism through which people obtain resources.
Perhaps Universal Basic Income.
Perhaps a social dividend.
Perhaps dramatically shorter working weeks.
Perhaps sovereign investment funds owning AI assets on behalf of citizens.
Perhaps broader employee ownership.
Perhaps taxation moves away from labour and towards capital.
Perhaps something we haven't invented yet.
But there is an important distinction.
Money paid to somebody because technology has made them economically redundant sounds like welfare, but a payment representing their share of productive assets is something else. It is a dividend.
Artificial intelligence is built upon generations of accumulated human knowledge, publicly funded education and research, infrastructure and collective technological development.
Why shouldn't ordinary citizens possess some ownership of what it eventually produces?
Then automation need not mean:
£30,000 salary becomes £0.
It could mean:
less paid work + social dividend + cheaper goods + greater free time.
That begins to look like liberation.
There is also a hard economic reason why some redistribution may become unavoidable.
Machines may be able to produce extraordinary quantities of goods. But somebody still needs the purchasing power to buy them. An economy in which the productive assets belong overwhelmingly to a tiny group while everybody else's income disappears eventually encounters an obvious question.
Who are the customers?
But even there, we should distinguish being kept economically functional as consumers from possessing genuine economic freedom.
The wrong fear
Science fiction has trained us to imagine machines becoming like us.
Conscious.
Ambitious.
Angry.
Power hungry.
Perhaps we are looking in the wrong direction.
The nightmare may not be that artificial intelligence develops human weaknesses.
It may be that artificial intelligence becomes extraordinarily capable while we retain ours.
Our greed.
Our tribalism.
Our obsession with status.
Our inability to cooperate beyond borders.
Our tendency to concentrate wealth and power.
Our extraordinary capacity to recognise collective problems while continuing to pursue individual advantage.
The darkest future is not necessarily one in which AI takes power away from humanity.
It may be one in which AI enables a tiny part of humanity to accumulate more power than human beings have ever possessed before.
But that future isn't inevitable.
The same technology capable of concentrating power could help reduce scarcity. The same intelligence capable of replacing work could free human beings from spending so much of their lives performing it.
The same AI that could optimise the profits of a multinational corporation could optimise a regional food network so that perfectly usable food feeds people rather than becoming waste.
That is why I am not anti-AI.
I am pro-human.
And perhaps the question we should start asking isn't:
How do we make sure everybody still has a job?
It is:
Why should the right to live securely still depend upon having one if the economy eventually stops needing everybody's labour?
Artificial intelligence cannot make that decision for us. And nor should it.
The decision is political, economic and moral. Ultimately, it is human.
I am not convinced that AI will fail humanity.
I am considerably more worried that humanity will fail to share what AI makes possible.
References
Sources and Notes: The Future of AI: Freedom or Feudalism?
If you enjoyed this article...
You may also find The Price of Independence interesting.
This article looks at how housing costs, education and the structure of modern life can reduce independence long before technology removes a job.
About the author
Paul Clark is the author of The GOOD Book: A Behavioural Operating System for Escaping Debt and Rebuilding Control and the creator of The GOOD Method. If you enjoyed this article, explore more of his writing in GOOD Thoughts, listen to The GOOD Conversation podcast, or join the GOOD Community to receive new articles, podcast episodes, resources and forthcoming videos by email. If there is a subject you would like Paul to investigate or write about, get in touch.