Humans are not perfectly vigilant

And that’s bad news for AI.

Cory Doctorow
7 min readApr 1, 2024
Children playing on a climber. The colors of the climber and the foliage behind them has been oversaturated and shifted, making it surreal. The kids’ heads have been replaced with the red eye of HAL 9000 from Kubrick’s ‘2001: A Space Odyssey.’ Standing in the foreground at rigid attention is a man in short-sleeved military garb, wearing aviator shades. Image: Cryteria (modified) CC BY 3.0

I’m on tour with my new, nationally bestselling novel The Bezzle! Catch me in Boston (Apr 11) with Randall “XKCD” Munroe, Providence (Apr 12) and beyond!

Here’s a fun AI story: a security researcher noticed that large companies’ AI-authored source-code repeatedly referenced a nonexistent library (an AI “hallucination”), so he created a (defanged) malicious library with that name and uploaded it, and thousands of developers automatically downloaded and incorporated it as they compiled the code:

These “hallucinations” are a stubbornly persistent feature of large language models, because these models only give the illusion of understanding; in reality, they are just sophisticated forms of autocomplete, drawing on huge databases to make shrewd (but reliably fallible) guesses about which word comes next:

Guessing the next word without understanding the meaning of the resulting sentence makes unsupervised LLMs unsuitable for high-stakes tasks. The whole AI bubble is based on convincing investors that one or more of the following is true:

  1. There are low-stakes, high-value tasks that will recoup the massive costs of AI training and operation;
  2. There are high-stakes, high-value tasks that can be made cheaper by adding an AI to a human operator;
  3. Adding more training data to an AI will make it stop hallucinating, so that it can take over high-stakes, high-value tasks without a “human in the loop.”

These are dubious propositions. There’s a universe of low-stakes, low-value tasks — political disinformation, spam, fraud, academic cheating, nonconsensual porn, dialog for video-game NPCs — but none of them seem likely to generate enough revenue for AI companies to justify the billions spent on models, nor the trillions in valuation attributed to AI companies:

The proposition that increasing training data will decrease hallucinations is hotly contested among AI practitioners. I confess that I don’t know enough about AI to evaluate opposing sides’ claims, but even if you stipulate that adding lots of human-generated training data will make the software a better guesser, there’s a serious problem. All those low-value, low-stakes applications are flooding the internet with botshit. After all, the one thing AI is unarguably very good at is producing bullshit at scale. As the web becomes an anaerobic lagoon for botshit, the quantum of human-generated “content” in any internet core sample is dwindling to homeopathic levels:

This means that adding another order of magnitude more training data to AI won’t just add massive computational expense — the data will be many orders of magnitude more expensive to acquire, even without factoring in the additional liability arising from new legal theories about scraping:

That leaves us with “humans in the loop” — the idea that an AI’s business model is selling software to businesses that will pair it with human operators who will closely scrutinize the code’s guesses. There’s a version of this that sounds plausible — the one in which the human operator is in charge, and the AI acts as an eternally vigilant “sanity check” on the human’s activities.

For example, my car has a system that notices when I activate my blinker while there’s another car in my blind-spot. I’m pretty consistent about checking my blind spot, but I’m also a fallible human and there’ve been a couple times where the alert saved me from making a potentially dangerous maneuver. As disciplined as I am, I’m also sometimes forgetful about turning off lights, or waking up in time for work, or remembering someone’s phone number (or birthday). I like having an automated system that does the robotically perfect trick of never forgetting something important.

There’s a name for this in automation circles: a “centaur.” I’m the human head, and I’ve fused with a powerful robot body that supports me, doing things that humans are innately bad at.

That’s the good kind of automation, and we all benefit from it. But it only takes a small twist to turn this good automation into a nightmare. I’m speaking here of the reverse-centaur: automation in which the computer is in charge, bossing a human around so it can get its job done. Think of Amazon warehouse workers, who wear haptic bracelets and are continuously observed by AI cameras as autonomous shelves shuttle in front of them and demand that they pick and pack items at a pace that destroys their bodies and drives them mad:

Automation centaurs are great: they relieve humans of drudgework and let them focus on the creative and satisfying parts of their jobs. That’s how AI-assisted coding is pitched: rather than looking up tricky syntax and other tedious programming tasks, an AI “co-pilot” is billed as freeing up its human “pilot” to focus on the creative puzzle-solving that makes coding so satisfying.

But an hallucinating AI is a terrible co-pilot. It’s just good enough to get the job done much of the time, but it also sneakily inserts booby-traps that are statistically guaranteed to look as plausible as the good code (that’s what a next-word-guessing program does: guesses the statistically most likely word).

This turns AI-”assisted” coders into reverse centaurs. The AI can churn out code at superhuman speed, and you, the human in the loop, must maintain perfect vigilance and attention as you review that code, spotting the cleverly disguised hooks for malicious code that the AI can’t be prevented from inserting into its code. As “Lena” writes, “code review [is] difficult relative to writing new code”:

Why is that? “Passively reading someone else’s code just doesn’t engage my brain in the same way. It’s harder to do properly”:

There’s a name for this phenomenon: “automation blindness.” Humans are just not equipped for eternal vigilance. We get good at spotting patterns that occur frequently — so good that we miss the anomalies. That’s why TSA agents are so good at spotting harmless shampoo bottles on X-rays, even as they miss nearly every gun and bomb that a red team smuggles through their checkpoints:

“Lena”’s thread points out that this is as true for AI-assisted driving as it is for AI-assisted coding: “self-driving cars replace the experience of driving with the experience of being a driving instructor”:

In other words, they turn you into a reverse-centaur. Whereas my blind-spot double-checking robot allows me to make maneuvers at human speed and points out the things I’ve missed, a “supervised” self-driving car makes maneuvers at a computer’s frantic pace, and demands that its human supervisor tirelessly and perfectly assesses each of those maneuvers. No wonder Cruise’s murderous “self-driving” taxis replaced each low-waged driver with 1.5 high-waged technical robot supervisors:

AI radiology programs are said to be able to spot cancerous masses that human radiologists miss. A centaur-based AI-assisted radiology program would keep the same number of radiologists in the field, but they would get less done: every time they assessed an X-ray, the AI would give them a second opinion. If the human and the AI disagreed, the human would go back and re-assess the X-ray. We’d get better radiology, at a higher price (the price of the AI software, plus the additional hours the radiologist would work).

But back to making the AI bubble pay off: for AI to pay off, the human in the loop has to reduce the costs of the business buying an AI. No one who invests in an AI company believes that their returns will come from business customers to agree to increase their costs. The AI can’t do your job, but the AI salesman can convince your boss to fire you and replace you with an AI anyway — that pitch is the most successful form of AI disinformation in the world.

An AI that “hallucinates” bad advice to fliers can’t replace human customer service reps, but airlines are firing reps and replacing them with chatbots:

An AI that “hallucinates” bad legal advice to New Yorkers can’t replace city services, but Mayor Adams still tells New Yorkers to get their legal advice from his chatbots:

The only reason bosses want to buy robots is to fire humans and lower their costs. That’s why “AI art” is such a pisser. There are plenty of harmless ways to automate art production with software — everything from a “healing brush” in Photoshop to deepfake tools that let a video-editor alter the eye-lines of all the extras in a scene to shift the focus. A graphic novelist who models a room in The Sims and then moves the camera around to get traceable geometry for different angles is a centaur — they are genuinely offloading some finicky drudgework onto a robot that is perfectly attentive and vigilant.

But the pitch from “AI art” companies is “fire your graphic artists and replace them with botshit.” They’re pitching a world where the robots get to do all the creative stuff (badly) and humans have to work at robotic pace, with robotic vigilance, in order to catch the mistakes that the robots make at superhuman speed.

Reverse centaurism is brutal. That’s not news: Charlie Chaplin documented the problems of reverse centaurs nearly 100 years ago:

As ever, the problem with a gadget isn’t what it does: it’s who it does it for and who it does it to. There are plenty of benefits from being a centaur — lots of ways that automation can help workers. But the only path to AI profitability lies in reverse centaurs, automation that turns the human in the loop into the crumple-zone for a robot:

If you’d like an essay-formatted version of this post to read or share, here’s a link to it on, my surveillance-free, ad-free, tracker-free blog: