Word-guessing programmes like ChatGPT are not going to take over the world. But when this bubble pops, there’s a lot to be lost
When people talk about the dangers of AI, they often fret about whether the software agents that litter our inboxes and wreck our search engines will one day take all our jobs. Those are the reasonable worriers!
The unreasonable AI worriers insist that the word-guessing programmes will some day awaken, become God and turn us all into paper clips. This is like worrying that, if we keep breeding our horses to run faster, eventually our prize mares will start foaling locomotives.
What people don’t worry about is the fact that 35 per cent of the S&P 500 is tied up in seven firms, six of which have yet to make a penny’s profit on AI. (Nvidia, the seventh one, is profitable, because it’s the company the other six have given all their investors’ money to.)
Advertisement
AI then is the greatest money-wasting venture the human race has ever embarked upon. At $1.4tn in capital expenditure and counting, this industry – with an annual global turnover of less than $60bn – could become the most toxic investment bubble in history. As Stein’s Law has it, “anything that can’t go on forever eventually stops”.
How have we arrived at this juncture? The narrative push to convince investors that tech monopolists could continue to grow has passed through three distinct phases. In the beginning, tech companies insisted they could devour one another’s markets. Google+ would eat Facebook; Facebook’s “pivot to video” would eat YouTube.
This led to the next phase of this narrative project: claiming that your company is about to conquer a non-existent market, such as Web3, cryptocurrencies, non-fungible tokens (NFTs), or (God help us all), “the metaverse”.
The final phase of the growth narrative was AI. Like the other narrative projects, it involves fanciful claims about the vast market opportunity represented by a new kind of software.
But AI is different: first, because the software is extremely technically interesting; and second, because the markets have coughed up vastly more money for AI than they ever splashed out on those earlier bubbles.
Facebook spent “only” $60bn on the metaverse (plus the cost of reprinting all their headed paper with “Meta” across the top). But Meta’s AI bill for the past three years sits around $150bn, and they claim they’ll double that total this year.
To understand why the AI bubble has attracted so much more capital than, say, the metaverse or NFTs, we have to examine the ideology of AI as it is sold: AI dangles the possibility of a world without people.
For government officials, AI promises the bypassing of the permanent civil service, thus allowing policy to be enacted by a single visionary minister and their army of software agents.
Bosses, meanwhile, hope that AI might liberate them from the haunting knowledge that, if they don’t show up for work, everything hums along fine, whereas if all the workers stay home, the business shuts down immediately. Bosses want to believe they’re in the driver’s seat, but they fear that they’re strapped into the back seat with a toy steering wheel.
AI, though, might let them wire that toy steering wheel directly into the firm’s drive train. Once the boss replaces his workers with AI, he need only tap out a prompt and watch as the product or service is conjured directly into existence, without it having to pass through the hands of workers, who might tell him what a stupid idea it is.
This is an attractive proposition. Who among us has not lamented that hell is other people? Your AI boyfriend or girlfriend can give you dating without dates, just as Mark Zuckerberg hopes to use AI to deliver social media without socialising. It will just be you and your army of chatbots spending every hour God sends “maximising your engagement” with his platforms.
I’m not saying AI is not real. It is. A decade ago, computer scientists tried applying existing statistical methods in novel ways, and made a substantial breakthrough in machine learning. These techniques turned out to be highly scalable. When we simply apply the same techniques with increasing intensity, we got better and better results.
But these results have since plateaued, and anything that can’t go on forever eventually stops. Some of these results lend themselves to being turned into products. But in the absence of the bubble, we would call these products “plug-ins”.
My first word processor was a programme that came printed in the pages of a magazine, which I had to retype into my computer. In the 45 years since then, my word processors have got new features all the time. Many of these features are useful – and some of them are awfully stupid.
Now we have a new word-generation plug-in (ChatGPT, Claude etc). If you like that plug-in, good for you (please don’t send me its output, though… ugh). The existence of a new plug-in, even an exciting one, does not militate against feeding all the writers into a wood chipper, nor does it herald the imminent emergence of a new machine god that might enslave the human race.
And that plug-in isn’t worth $1.4tn. We can tell, because no one is willing to pay the true cost of the “tokens” the AI giants sell. As these giants get ready for their flotation and attempt to clean up their balance sheets by raising prices, their largest enterprise customers are imposing strict limits on AI usage.
Twenty-five years ago, I lived through the dotcom bubble. Many people point to that bubble and insist that even though that market frothed with silly firms, the underlying ideas were sound, and a bet on the web was a good one. It is argued that this means AI must be a good bet, too. This is an obvious fallacy: the fact that one thing stopped losing money and became profitable doesn’t imply that losing money is itself an indicator of long-term growth.
Unlike AI, the web had brilliant unit economics. Adding a user to the web made the web more profitable. Every new AI user makes the AI sector lose more money. Each use of the web made the web more profitable. Every time you prompt an AI, the company supplying it loses money. Every generation of the web was more profitable. Every new generation of AI loses more money than the last one.
AI then is a normal technology, a grab bag of plug-ins that different people will find useful to different degrees. But it is also an abnormal bubble, vastly larger and more dangerous than the tech bubbles that preceded it.
AI isn’t going to do your job, though an AI salesman might well convince your boss to fire you and replace you with an AI that can’t do your job. And AI might still destroy your livelihood, when 35 per cent of the stock market collapses overnight. (Anything that can’t go on forever eventually stops.)
Cory Doctorow is the author of Enshittification: Why Everything Suddenly Got Worse and What To Do About It. His new book, The Reverse Centaur’s Guide to Life After AI: How to Think About Artificial Intelligence Before It’s Too Late, is published by Verso on June 23
