AI is a new era of the cyber-revolution, itself a new era of the Industrial Revolution.

At Nazaré, a great undersea canyon channels the energy of the Atlantic into walls of water 100 feet high. Surfers launch themselves onto these walls and try to hold control at the edge of uncontrol, reading every variation in the great power behind them. Looking at the AI wave - and the broader technological sequence it belongs to - we are all like those surfers at Nazaré. The question is whether we can understand this power well enough to ride the wave towards human self-realisation, or will we spend the coming decades mobilising all our efforts of rescue?

Each wave of modernity feels like the most powerful we have ever faced. But instead of trying to start afresh, attempting to anticipate precisely how unfinished revolutions will unfurl, it is worth revisiting how it felt on a past crest and what we can learn from that.

This essay looks at how contemporaries saw the shift from steam to electricity 150 years ago, what that tells us about our possible futures, and how we could respond. 

Here, I explore three particular reactions: those of Zola, Kropotkin and Wells.

Humanity was irrevocably marked by industrialisation in the age of coal and steam. Marx makes the point analytically in Chapter 15 of Capital: “In handicrafts and manufacture, the workman makes use of a tool, in the factory, the machine makes use of him; [the factory] confiscates every atom of freedom, both in bodily and intellectual activity.” One of the most powerful novelised descriptions comes in Zola’s Germinal, where the coal mine, Le Voreux (‘the voracious one’), kept dry by the incessant operation of its steam-powered pumps, devours its workers and eventually crushes the hopes of Etienne and Catherine for a domestic life together. Zola wrote Germinal in 1885, three years after Edison had installed the world’s first central electrical power station at Holborn Viaduct in London. In 1901, Zola wrote his own reply to it. Travail presents electricity as the techno-utopian solution to the ills germinating out of Le Voreux: it opens at the end of a strike at a great steel-works called l’Abîme (‘the abyss’), starting where Germinal ended. Electricity transforms the town. In a utopian vision of communal plenty, the saintly owner-inventor looks forward to a fully electrified future of abundance and simple enjoyment. Mechanical power - once an instrument of oppression under steam - becomes part of the human commons, as free as air and water seemed back then. This is the original statement of “too cheap to meter”, when power is so abundant that billing usage isn’t even worth the trouble.

Two years before Travail, the anarchist thinker Peter Kropotkin directly observed how electricity was changing Europe’s industry in Fields, Factories and Workshops. He saw hydro-electricity transforming production into small, self-governed units. In Oyonnax, France - then the global centre of comb-making - a hydro plant allowed “300 workers to leave the small workshops to work in their houses”. Where Zola saw a utopia of abundance, Kropotkin realised that electricity was technologically permissive: it allowed human-scale associative organisation based on shared power.

Not everyone who felt the coming wave was quite as lit up about it. In 1899, H.G.Wells published When the Sleeper Wakes - a monopolistic dystopia in which economies of scale in power production have put all social functions into the hands of the faceless ‘Trust’. The population is pacified by Babble Machines, public phonographic devices braying news, sensation and “viral content” into the streets. They are given pretend work by “the Labour Company”, whose purpose is simply to manufacture jobs. For Wells, new clean and silent power leads to palliative blandness in the service of concentrated technocratic power.

How does the arrival of electricity illuminate us now, on the crest of AI? First, the lesson of humility. Electricity became a Rorschach test, revealing the observer as much as the technology: by the time of Travail, the Dreyfus affair had been won, the Republic seemed secure, and perhaps Zola suffered from willed optimism. Yet the results were still practically influential. The movement for “gas and water socialism”, combined with Liberal-Conservative municipalism, united local councils in turning private energy monopolies into locally controlled public utilities. This footnote rings loud today - given Andy Burnham’s hopes for Manchesterism - Kropotkin highlighted Manchester’s municipal electricity supply as vital to the future of its cloth industry.

The second lesson: no one, utopian or dystopian, mourned the dark satanic mills of Le Voreux or l’Abîme. That is where we now stand. The steam age of the cyber-revolution has broken and that wave is passing. Web 2.0 is in full disruption, and we are surely hopeful for an end to its particular harms: the dark patterns of influence and addiction, the teen mental-health damage, behavioural exploitation for private monopoly gain and mass distribution of fraud. The rabbit-holes that seem all to lead, as if by a pervasive gravitational pull, to deep hatreds of women, Jews and black people. 

My favourite account of the AI investment euphoria starts from the fact that the great platforms of Web 2.0 built their profits on attention-extraction.They came to intermediate our purchases and our social lives, and levied an advertising charge on almost all consumer goods in the economy, whether or not a particular consumer was influenced by an ad. Agentic and generative AI threatens that model: if an agent works on your behalf and buys for you, then the agent is the intermediary, not the search engine or the social network, and it is optimising for you, “its human”. As the leaked Google memo bemoaned, platforms “have no moat”, no durable barrier, that is, to protect those profits from competitors, or better still for us, from the agent-empowered consumer and citizen themselves. So the Web 2.0 platforms are spending all their current profits, and borrowed money besides, in the hope that such a barrier can be built rapidly. They feel they are fighting for their lives. For this at least we should be glad: the coming period is one of Schumpeter’s famed “gales of creative destruction”, in which new technology sweeps away incumbents. But, unless we act with great foresight, we might only allow them to be replaced with  a new rostrum of dominant firms. Schumpeter’s meteorological metaphor understates the agency we in fact have over the shape of the wave and how we ride it.

If Germinal is the novel of the steam age at its bitter maturity, the novel of our own passing age is Ben Lerner’s Transcription: a deep exploration of how communications technologies transform humanity, from radio to attention-harvesting smartphones today. The screens of Transcription are like Zola’s mine, Le Voreux: “even if the light is blue it is a black hole for the eyes”. And Lerner, like Zola, allows a single, ambiguous ray of hope. A child who will not eat is coaxed back to food and to life through the trance-state of unboxing videos. Here, humanity re-asserts its ability to make meaning in the face of the commodification of attention.

But Transcription is a book of the pandemic years, silent on the post-2022 fact of Turing-test-passing intelligences in our midst today. It digests the wave that is ending, not the one that is building. Still, what Lerner insists on is a constant: technologies of communication spread like epidemics, running their damage and delivering their wonders to society, leaving a permanently transformed humanity to face the next wave.

So what of these artificial minds? Zola, Wells and Kropotkin offer three lenses here too. Like Wells, we can zoom in on all that is most exploitative in the technology and extrapolate. Like Zola, we can lock onto the end of the horror and paint an idyllic future. Like Kropotkin, we can observe where the genuine spring shoots are and let them feed our optimism of the will. My sympathies are with the third; but let’s put flesh on each.

Wellsian AI

Translating Wells’ vision to AI is not hard. First, the AI labs discover their protective moat. Because the underlying models will quickly become commodities, competing down to cost, the real barrier to entry is data. Personal data creates a familiar and hyper-convenient environment for each user. Then new data generated by our interactions constantly feeds further scale.

Next comes social control. Hyper-personalised AI agents - whose early sycophancy we are already seeing - will be deployed first for the bottom line (advertising dressed up as advice), and then for control proper. The Wellsian Trust (what a wonderfully misapplied term) delivers the opiate of endless friendship entertainment, with harsher measures waiting in the wings if pacification fails. 

Finally add the geopolitical layer. Silicon Valley controls Washington, with Europe as a feudal possession..China runs its own version of total control, and great-power competition ensures that no bottom-up public control of AI is possible. This is the world implied by the interview between Musk and Zanny Minton Beddoes. The old Musk who wanted a pause now thinks the tiger is out of the bag; hang on to its tail and enjoy the ride, though perhaps you need to be a trillionaire with a rocket to Mars to do so.

Zolaesque AI

What about the Zola reaction? Focus on the end of the horror and paint what follows: human self-realisation for all. Here, Zola mirrors William Morris’ utopian ideal “All work which would be irksome to do by hand is done by immensely improved machinery; and in all work which it is a pleasure to do by hand machinery is done without.” The vision has real attractions. I think of a piece by John Berger contrasting the 2011 London riots with a summer scene on a lakeshore: families enjoying the sun, the water and the company, children queuing for ice creams. Heaven on earth, perhaps tantalisingly close in a world of full automation. 

Nor is this optimism limited to escaping drudgery. When an AI model recently helped mathematicians disprove a conjecture of Erdős, Tim Gowers, one of today’s great mathematicians, welcomed the AI as a tool, like algebra, like pencil and paper, for pushing the boundaries of knowledge. 

But in that very same week, so-called Grantagate exploded: a literary fiction prize, it seems, may have been awarded to a story which many thought had been written by machine. The event fuelled the growing neo-Luddism. The trouble with the Zola reaction is its pure optimism: the inevitability of our rosy post-work future seems to evade worries about the nitty-gritty of how society shapes it. When Musk treats the prospect of universal abundance as a technical guarantee, he manages to ignore who actually holds the power (including his own grip on the US administration). Like Pharaoh redoubling the Hebrew slaves’ workloads when Moses asked for rest, today’s tech elites will not surrender control voluntarily. What will force these  bro-Pharaohs to listen when the call comes to “let the people free”. 

Kropotkin AI

So what are the green shoots that might lead us down a decentralised, Kropotkin-style path instead?

Vanishing friction 

The first is that Web 2.0 platform power may really be undermined by AI agents. That power arose from transaction costs - the friction required to do something online. Platforms made laborious tasks simple for us, and so we handed our data over to them. The transaction costs of using agents have fallen to almost zero, and with them the fundamental source of platform power. This is Kropotkin’s divisibility of horsepower reborn: just as small electric motors liberated 19th-century comb-makers to leave central factories and work productively at home, personal AI agents free individuals from relying on giant tech monopolies.  

Open-source AI

The second green shoot is that open-weight models (whose parameters are published for anyone to download, run and improve) show that there seem to be no intrinsic and unavoidable economies of scale in learning. Domination is not inevitable. Projects like OpenClaw, an open-source personal agent that runs on your own machine, became one of the fastest-growing software projects in history, with users organising themselves into local meetups from Los Angeles to Dubai. However much Big Tech might will a moat, the technology favours decentralisation. 

The tacit knowledge debate

And the third green shoot relates to the socialist calculation debate. AI appears to counter the long-standing Hayekian result: central planning fails because no one authority could ever collect millions of people’s unwritten, messy personal preferences (look at the LLM-generated synthetic focus groups Electric Twin are creating) or the ‘tacit knowledge’ required for production. 

Yet this need not become a carrier of Wellsian gigantism. A Kropotkinian reading is equally available: if knowledge can be encapsulated and preferences understood without prices and markets that come to be dominated by the data-and-attention-oligopolies, any grouping of us can become the production units we want to be. What Kropotkin saw in the divisibility of horsepower, we now have in the reproducibility and shareability of tacit knowledge.

The comb-makers of Oyonnax needed the waterfall, but they also needed their spirit of association, their societies, cooperatives, mutuals and unions that civilised the steam-driven world. Citizens and communities will need their interests protected in cyberspace in much the same way. We are at a brief moment of transition where the future of AI remains genuinely open. Whether we get the Wells future or something better will not be settled by the inherent characteristics of the technology, but by how fast a decentralised, commons-based alternative matures.

For all the green shoots Kropotkin noted, those observations alone were not enough to usher in a human-scale, decentralised modernity in which power - in all senses - was spread evenly. So what does this mean concretely for the British government? I see three broad strands of action.

1. Contain platform power

The platforms, both the old Web 2.0 behemoths and the AI-native pair of Anthropic and OpenAI, have “dominance or bust” stock-market valuations. They need to create barriers to effective competition if they are to pay back investors for their AI turn. The first job of government policy is to make sure that such barriers are not created artificially. If there turn out to be economies of scale in intelligence, so be it. But if we can enjoy the upside of artificial minds without the downside of dominance, let us do so. 

The policies to achieve this are relatively well-understood from the previous era of trying to diminish Web 2.0 platform power:

• Insist on data portability so that each person’s history with a model or with agents does not itself become a source of lock-in, a barrier to entry or an advantage in training the next generation of model. We should require that all interactions and history be available under user control via an API in real time from the start.
• Insist that humans have the right to direct an AI-agent working on their behalf to any web end-point intended for them. This evens the playing field between AI labs deploying intelligence to lock us in and our own agents countering those efforts.
• Insist that any organisation using data on the basis of consent have a duty of care to the data subject, in order to stop the highly under-handed and harmful data profiling practices so common on the Web 2.0 platforms.

These three policies will severely constrain AI labs’ abilities to repeat the platform-dominance playbook. But they are not sufficient on their own, because there are elements of natural monopoly in cyberspace that require public action, in order to avoid private monopolisation.

2. Provide the digital public infrastructure tools that allow good decentralised solutions to emerge

Platforms provide real value, especially in their early days. Who does not use Google, Apple or Microsoft to sign in to any number of services online, for convenience and for a basic sense of security? This value has allowed a handful of firms to become the default identity providers while also, conveniently, adding to their trove of personally linkable data. 

Identity will become an even more important service in the agent economy. How will a service provider know whether an agent really has been authorised by their bona fide human to perform some transaction on their behalf? The agent will need an identity and a set of permissions, and that identity will need to be tied back to the human. This infrastructure of trust needs to be provided in an interoperable and standards-based way in order to stop it becoming a vector for monopolisation. 

Other public goods that will become central to a well-functioning agent ecology, like fraud monitoring and even social discovery, will also need standards and vigilance if they are not to be leveraged into power over us. This is the kind of digital infrastructure that countries like India, through the Aadhaar identity system and the Beckn ecommerce protocol, have developed and deployed as public goods. We should learn, and borrow, from them.

3. Maintain a capacity to decide our future as a sovereign, democratic country

Britain is meeting this transformation from the periphery, which is a new position for us. We were the workshop of the world and therefore the centre of the first Industrial Revolution (a position our institutions and our culture sometimes still seem to assume). Today the Bay Area and Beijing are the brainshops of the world, and with it, the centre has moved. Using what power we have left will take some political and cultural adjustment, because the temptation is either to overstate our leverage or to conclude that we have none at all.

We have rather more than none. We write our own data law, our own publishing, copyright and advertising laws; we regulate closely several of the markets in which AI stands to make the most money - financial and legal advice among them. The task is to use those instruments deliberately to: 

• set the terms on which models may train on British creative work and British public records 
• insist that public-interest AI, in health, in education, in local government, is procured so that the resulting data and models stay under public control rather than becoming the property of a vendor
• build enough domestic capability to ensure we are a customer with alternatives rather than a captive one. We should use our data sovereignty, because our data is the one input to the AI economy that is unambiguously ours to govern, and without which no platform can serve our needs.

The other half of sovereignty is democratic resilience. We must recognise the degree to which our basic processes of democracy became open to manipulation under Web 2.0: we should have protected ourselves earlier against Putin’s information warfare, Musk’s troll-stoking right-populism and even the accidental wreckers in the form of eyeball-chasing Macedonian teenagers. 

There is some hope that agentic media will not have identical vulnerabilities: creating a model is an expensive business, whereas creating a tweet is not. Even Musk has found it hard to make Grok unwoke out-of-the-box without heavy-handed (and hilarious) interventions. But that expense is itself a vulnerability, since it puts the supply of models within reach of well-funded actors and of nobody else. 

Lerner warns us about the voices of angels and devils that filled the airwaves of the 1930s; public policy needs to become more McLuhan-esque. If the medium is the message, and the message makes society, then society needs to regulate the medium if it is to shape its own destiny. 

This is a time for furious paddling into the AI wave, truly a giant this time. It does not need to be a goggle-eyed ride with Musk and the other Pharaoh-bros. Berger’s simple summer lakeshore scene is in sight too, but only if we get public policy right.