Soundtrack to this post.
Working Class Movement Library catalogue, via Wikimedia Commons. Public domain, copyright expired.
Three years ago I put up a post about a Buckminster Fuller quote, the one about inventing jobs because of a false idea that everybody has to be employed at some kind of drudgery. I said then that “automation is taking away people’s work” reads to me as an argument for automation rather than against it. Who actually wants to work? If work was the thing we wanted to spend our lives doing, they would not need to pay us to turn up. We work because we need to eat and to have somewhere to live.
I still think that. What has changed is that I now use AI most days, and the political question has stopped being abstract. So I want to set out why I think the left, and the greens in particular, are on the wrong side of this, and what it would take for me to be wrong.
First, a caveat about the framing. It is tidier to say the left is against AI and the right is for it than it is true. Plenty of the right hates it too, for reasons about national labour markets, or Silicon Valley, or the politics they believe is baked into the models. And some of the loudest enthusiasts would describe themselves as being on the left. What is true is that almost every objection I find persuasive comes from the left, and almost nobody on the left is making the case for.
The objection that nearly changed my mind
Keynes wrote in 1930 that his grandchildren would be working fifteen hours a week by about now. Output per hour is several times what it was then. The fifteen hours never arrived. The gains were not distributed, they were captured, and everything else follows from that.
There is a harder version of this. Every reduction in working time we actually enjoy, the weekend, the eight hour day, holidays, sick pay, was won by organised workers with leverage, and that leverage came from being difficult to replace. AI is precisely a technology for making people easier to replace, including the people whose scarcity used to give them a bargaining position. So the machine that could pay for the leisure is also the machine that dismantles our ability to demand it.
I do not have a rebuttal to that. I think it is correct. What I dispute is the conclusion. If the problem is that the gains get captured, the answer is to contest the ownership, not to oppose the machine. Opposing the machine is the one strategy that has never worked, and it leaves the ownership question untouched.
I also want to correct myself. In an earlier post I wrote that data is the primary resource of the AI era, and that reclaiming our data was the first step. I was wrong about that, or at least I was a year late. The models are already trained. Withholding my files from now on does not untrain anything. The binding constraints are compute, capital and electricity, and none of those are in citizen hands or heading there. Hyperscaler capital spending this year is over six hundred billion dollars, around ninety per cent of the accelerator spend goes to one company, and roughly nine tenths of leading edge chip production happens at one foundry in Taiwan. That is a narrower chokepoint than oil ever was. Self hosting my files while renting my thinking is not the answer I thought it was.
The carbon
This is the objection I take most seriously, partly because I made it myself in that same post. Datacentre electricity was around 415 terawatt hours in 2024 and the IEA expects it to roughly double by 2030, reaching about 945 terawatt hours, which is a little more than Japan uses today. Emissions in the sector are rising while every other sector is supposed to be falling. Those are the facts and no amount of enthusiasm dissolves them.
The reply I hear most often is that the carbon spent now buys carbon saved later, through better materials, better grids, better science. The IEA has actually modelled this, and existing applications could deliver reductions several times larger than datacentre emissions by 2035. But the same report says there is currently no momentum that would deliver those reductions, and that without deliberate conditions the effect could be marginal. So it is not a fact about the technology, it is a promise about policy, and we have all heard promises of that shape before from people who wanted to keep drilling.
There is a better argument available here and it is sitting in Scotland. Last year turning Scottish wind down cost around £350 million, and the replacement generation, mostly gas, cost over a billion. Of the ten terawatt hours curtailed across Great Britain, northern Scotland alone accounted for 8.8 of them, meaning only 61 per cent of the wind power that region could have generated reached the grid. We generate clean power we cannot move and then pay twice for the failure. Compute that can be sited near constrained generation and switched off when the wind drops is a genuinely good use of that. Publicly owned compute soaking up wasted Scottish wind is a better answer than anything I have heard from the industry.
So here is my limit, since a green position without one is just enthusiasm wearing a green coat. Compute buildout should not exceed the clean generation added to the same grid over the same period. If datacentre demand keeps outrunning clean supply, then the carbon is not being invested, it is being borrowed with no plan to repay it, and I would oppose it.
The training data
The models were built out of other people’s work without consent or payment. That is true and I am not going to pretend otherwise. The UK government’s preferred fix was a data mining exception with an opt out, which would have put the burden on every illustrator and session musician in the country to go round asking not to be scraped. The consultation drew over eleven thousand responses, a Lords committee called unlicensed training on creative work a clear and present danger to the sector, and in March the government confirmed it no longer has a preferred option and is not proceeding with any of the four proposals for now. That is not a settled victory, because the opt out is parked rather than buried. But it is a win, and it was won by organised labour, which rather proves the point about leverage above.
My position is that this was expropriation for the common good. That is a respectable argument on the left with a long history behind it, and it is the argument I want to make. But it only works on one condition, which is that the thing expropriated ends up commonly owned. Taking from freelance illustrators to enrich shareholders is not the common good. It is just taking. Aaron Swartz went to war against paywalls around knowledge, and the same argument cuts against a private lab enclosing the commons in weights it owns and renting the access back. If I am going to use him, I have to follow him all the way, and that lands on open weights and public models rather than on a defence of the companies as they are.
I should also concede the part that hurts my original argument most. The first work this technology took was not the drudgery. It was illustration, session music, translation, copywriting. Those are the jobs people do for less money than they could earn elsewhere because they would rather do that than anything else. My whole case rests on work being something we would drop if we could afford to. The people hit first are the exception to it.
What it has given me, and to whom
Part of my job is getting utilities reconnected to empty properties so they can be prepared for someone to live in. A chunk of that is chasing suppliers, waiting on hold, and writing the same firm letter for the fifth time. I used to lose hours to a single complaint email. Now it takes minutes, and it says exactly what needs saying to get the thing moved.
The person who benefits from that is not me. It is whoever is waiting for that property. Every day off a void is a day sooner that a family gets a home. That is a small, concrete, material good delivered to precisely the people the left says it speaks for, and I trust it more than any argument about post scarcity leisure.
I would not extend that to the parts of public service where the human contact is the service itself. A housing officer’s time with a tenant is not overhead wrapped around the work, it is the work. And the pattern to watch is the one where cuts hollow out a service, automation fills the hole, and the staffing never comes back. Fewer than half the public trust councils or central government to use this stuff responsibly, and on the evidence of the last twenty years of government IT they have earned that scepticism.
Where I cannot answer
If the frontier models went dark tomorrow I would be stuck. Not inconvenienced, stuck. I have built things I could not now rebuild unaided, on infrastructure I do not own, from a supplier who can change the terms whenever they like. I moved my files off Big Tech and onto my own server and called it a small act of ownership, and I meant it, and I did it while quietly becoming a tenant again somewhere I was not looking.
That is the honest state of it. I am not making the case for AI from a position of independence. I am making it from inside the dependency, which is exactly why I want the ownership question settled rather than left to five companies and a foundry.
On the slop question, I have written before that if a work moves me the origin matters less than the responsibility around it. Who is paid. Who is harmed. Who is accountable. I still think that is the right test, and it is also better than the arguments usually made in its place, because people have always been perfectly capable of producing rubbish without assistance. What is new is not the average quality, it is the volume, and volume is a pollution problem rather than a taste problem. The sinks are the places we go to find things out. None of them were built for junk at no cost. The part that worries me most is not the writing at all, it is that when anything can be fabricated the default response to evidence becomes doubt, and deniability is a resource that accrues to whoever already has the lawyers.
Which gives me my red lines, and I did not have them written down until I went looking. No automated decision that takes away someone’s liberty, their immigration status, or the money they live on. No system where the person it lands on cannot reach a named human with the authority to overturn it. And nothing deployed on the back of work whose makers were neither paid nor represented in the decision to deploy it. That third one I am on the wrong side of every day I open one of these tools, and the only thing that fixes it is the models ending up commonly owned rather than rented, which is why I keep arriving back at the same demand.
Those hold whoever owns the model, including if the owner is the state, because public ownership was no protection at all in the Dutch benefits scandal or in Robodebt.
The ask
Public ownership settles who captures the gains. It does not settle what the thing is allowed to do to you, so the programme needs both halves.
Inference should be regulated as a utility rather than sold as a product. Compute should be publicly owned at national or European scale, sited where the power is being wasted and switched off when it is scarce. Models above a certain capability, and anything trained on public material, should have open weights so they can be run by people who do not own datacentres. And some decisions should be off the table for everyone, whoever owns the machine.
None of that is anti-AI. All of it is the sort of thing only the left has ever been any good at demanding. The right will not regulate this because it does not want to, and the tech industry will not do it voluntarily because nobody ever has. Leaving the question to whoever owns the GPUs is not caution. It is a decision, and it is being made without us.
I would rather we made it. Preferably before I have to explain to someone waiting on a housing list why their new home is delayed by a policy debate about whether the machine should have been allowed to send the email.
