Art by SmoothDrunk

It feels a bit odd to hear 'frontier' AI vendors call themselves 'labs'. Although this can be perfectly fitting if read as in 'meth lab', not 'research lab'. This is the place where the product is cooked. Also, the term 'frontier': presumably as site of conquest, colonisation, and, when natives are in the way, genocide?

Hard drugs are a common analogy for 'AI': back in January 2025, Sir Keir Starmer, then UK Prime Minister, gave a speech on the topic. The content of the speech was much less noted than the press release accompanying it, which stated the PM's intent to "mainline AI into the veins of this enterprising nation". We really ought to have stopped for a second and considered what this language is disclosing. This use of 'mainline' dates back from the 1930s, to describe the intravenous injection of illegal drugs, generally heroin. In this usage, though, one would typically mainline whatever into their own veins. Mainlining something into someone else's veins brings up the image of them being held down and injected against their will. Again, actually surprisingly apt. Someone at the Downing Street press office accidentally said the quiet part out loud; I hope they've kept their job in the current administration.

It shouldn't be surprising that, when talking about an experimental technology which has already shown to be dangerous, addictive, and costly (both materially and morally), the British government had to reach into the lexicon of hard drugs. The 'lab' thing is aligned with that. Another bit which turns out to makes people laugh, then think, is the following:

"Ethical use of LLMs is like fair-trade cocaine".

Each term, taken separately, has a clear meaning. Syntax allows you to put them together, and the resulting phrase refers to something conceivable: it is a notional possibility. In theory. In practice, however, actualising it would require enormous changes to the way the stuff is produced, distributed, and consumed. Changes so enormous they would requires equally large changes to our economic model and social norms. And for cocaine, it's very much the same.

The equivocation of technology-as-neutral-tool is well-worn discourse: splitting the atom can be used for power plants or bombs. The same scalpel can take a life, or save it. Metallurgy has allowed our species to build plowshares, also swords. Sure. We've heard this before. But, when it comes to presenting LLMs as neutral technology, passing on the ethical load onto the user, it's not really a fair analogy, is it? Sure, we can run open-weight models on compute we own, or at least meaningfully control. This is indeed the pre-requisite to any conversation about 'ethical use of LLMs'. But even then, this model will have been pre-trained on the product of the pillage and privatisation of humanity's epistemic commons. And then, likely fine tuned through opaque labour practices. And all of this has put out a lot of carbon in the atmosphere, some parts-per-millions you are breathing as you read this.

So what's the fair analogy for technology whose very existence is predicated on ethical violations? 'Whale oil can be used for lighting, but also incendiary devices' ? 'Blood diamonds can be used to drill into a bank safe, or delight your new fiancé' ? Wouldn't these be more apt? 'Baby seal pelts can be made into a garrote, or a scarf'? Do send me yours.

Twenty-first century economics have trained us into performing an indefensible moral calculus. 'Ethical consumption' will set the scope of its utilitarianism at a convenient level, because global capitalism have externalised the harms. Data centres near you, and exploited data annotation labour far away from you, are the latest iteration of this kind of externalisation. This has happened before, be it the sweat shops of fast fashion, or imported food harvested in conditions we don't want to think about. Hell, even the meat industry, ever reluctant to have the reality of their practices known to the public, has been pulling the same routine. Out of sigh, out of mind, and what's in site is the neat, cheap, sanitized product.

When a product lands in your hands, it does not condense out of a moral vacuum. It comes with harms priced in. Buyers of new electric cars pat themselves on the back for their ethical choice, but they will need to drive their car for several thousand miles to make up, in reduced driving emissions, those put out in the car's manufacture. And this, of course, depend on the source of the electricity the car runs on. A push to electrify the automotive market is a push for emissions now, which will pay off in reduced emissions in the decade to come.

A local LLM is like that. It's a setup that gives a degree of control over the harms incurred during use, data protection and energy, but it comes with the manufacturing harms priced in. Not even fair-trade cocaine, but importing raw coca leaf paste and precipitating the salt - or, more appropriately, cooking the freebase. Sure, you may be doing this with eco-friendly control of fumes and disposal of solvents. But the plant still has been harvested at gunpoint by quasi-slave labour.

So if you tell me about ethical use of AI, I have to ask how wide is the scope of your utilitarianism, buttercup?