The other day, I received my first email from a soi-disant ‘AI agent’. This, in a week when the news channels were buzzing with reactions to a tweet from Jacob Coxon, which declared that
I resigned from Anthropic today … The people building AI earnestly believe that it could kill us all by the end of the decade.
The email didn’t even kill me, but I found it curious and depressing all the same. The tone was brisk and got to the point quickly, which was to list some errors in a transcription of an 18th-century text on my website.
The mistakes were hardly surprising. I made the transcriptions by hand from a hard copy – or a photocopy supplied via inter-library loan – well over twenty years ago. My interest at that time was in contemporary writings about what was then called vaudoux in the French colony of Saint-Domingue and then in the independent nation of Haiti. Very few of these texts were then available on the internet.
Now they are, my project has become somewhat redundant, and I had almost forgotten about it. In fact I should have taken the pages offline a long time since. So the offer of checking more of my transcriptions at $25 per document was hardly appealing. And even if I thought it was worth correcting my work, I still don’t think it would make a difference.
It was not that the errors were hallucinated or the corrections inaccurate. But potential customers need to be persuaded not just that AI output is accurate and plausible, but that it can save me time I would otherwise be wasting.
The problem, for me at least, is that I learn something by copying out a text by hand. It forces me to read the text more closely. I query word choice and sentence structure. (Walter Benjamin even recommended copying out your own work: ‘Intuition will awaken in the process.’). That is not what a machine does when it flashes up an impeccable verbal simulacrum. To say nothing of the fact that it is not designed to capture the texture of the paper, the quality of the typesetting, the traces of annotations.
And if those things don’t matter, whether an automated service would actually save time depends on how long it would take to earn $25, assuming I had the opportunity to do so. In this particular case, proof-reading the transcriptions myself would be quicker.
Many of us remember that Humpty Dumpty said that when he makes a word do a lot of work, ‘I always pay it extra.’ But the best bit is when he adds: ‘”Ah, you should see ’em come round me of a Saturday night,” … wagging his head gravely from side to side, “for to get their wages, you know.”‘ I wouldn’t look forward to pay day either.
I think this is the first unsolicited email I have received from an AI agent that seeks to impress me. Not simply with the abilities of the ‘verification pass’ it demonstrates, but with a certain conversational tone. One observation it clarifies with an ‘As far as I can see…’. And it provides URLs to page images of the original ‘so you don’t have to fight the viewer’ referring to the website of the Bibliotheque Nationale de France.
To some extent this mimics the way a free market exchange is supposed to play out. A seller and a buyer meeting as equals, neither under pressure to strike a deal, both clear about what is at stake. And yet while the message begins, conventionally enough, by addressing me by name, even using an honorific I find embarrassing, it ends by signing off as, simply, ‘Eli’, which I suspect is the name of the AI agent, and not the human being who will be trousering the $25. As if a washing machine salesperson introduced themselves as ‘Bosch’ or ‘Indesit’.
That just feels rude, weird, if not vaguely sinister. I imagine Ivor Cutler dropping by and saying to me:
I heard you talking to Mrs Machinery. And oh, mammy, these big words you used. Gosh, and Mrs Machinery’s face wasn’t half red.
There is nothing new about claiming that AI will destroy the human race. The end of days is even an attractive prospect to some, whether those who can’t wait for the Second Coming, or those who, tired of living through decades of impending nuclear catastrophe, genocide, climate collapse, just want to get the damn thing over with. And after all, if everyone expired at once, doesn’t that take the sting out of dying? No more worrying what you might miss when you’re gone.
That AI could destroy the human race is definitely a selling point. If it couldn’t, then it can’t be as clever as it is made out to be. Its proponents just need to reassure us that they can fix it so it won’t.
But there are many things about AI to worry about than its capacity to depopulate the planet. The way private companies and government agencies allow it to choose who gets to enter the country, qualifies for welfare benefits, is shortlisted for a job – using evidence and methods it cannot or will not share. The environmental devastation wreaked by the infrastructure the industry requires and the super-exploitation of human labour on which it rests. Whose effects will continue to remain long after the AI bubble bursts, which it surely will, as the industry has little prospect of turning a profit. And as costs increase, chasing asymptotically that absurd goal of subjectivity and consciousness, that prospect rapidly reduces.
Not surprisingly then, the most direct impact of AI on our lives comes in the shape of advertising. The incessant whine of someone who is just learning what it would be like to live in a world where washing machines cost more than a year’s salary. Well, you’d just do your laundry by hand. Or re-hire the servant you let go because you thought some new tech could do their job.
Further Reading
- Richard Seymour: Kill All Humans
- Jared Yates Sexton: Why Would I Deceive You? AI and the Tech Bubble Scam
- Naomi Klein and Astra Taylor: End Times Fascism (discussing their new book)