There are big forces limiting AI’s reach.
Kevin Roose’s new article on AI in the New York Times, covers some of the possible dangers and weirdnesses of emerging AI technologies. One of the notable differences from most work in this genre is that in addition to interviews and book research, Roose did some admirable tinkering and self-education about AI implementations such as DALL-E. Roose’s learning is critical to the article’s excellent conclusion: that for popular media to meaningfully discuss AI technologies, journalists need to become more educated and thoughtful about such technologies and stop relying on tired sci-fi tropes. AI is no longer part of science fiction, it’s a real and impactful part of everyday life.
That’s why I think it’s a shame that the article itself falls victim to some tired tech-journalism tropes! I’m a professor of computer science who teaches machine learning at the world’s most selective and innovative university. I also run the school’s AI Lab, a collaboration with AI incubator Deepcore for students who are trying to launch AI-based startups. Before becoming a professor I worked at IBM research where I worked on AI for Industry and AI for education. My Ph.D. is in computational neuroscience. When I read articles about AI, I’m intensely aware of the sci-fi tropes that Roose descries. But I also feel that such journalists fall victim to their own penchant for narrative.
AI research is highly quantitative and non-narrative. Any attempt to “write a story” about AI will necessarily fail to describe some aspects of it. You see this Roose’s own article. He quotes Ajeya Cotra who studies AI risk at Open Philanthropy:
“A.I. systems can go from adorable and useless toys to very powerful products in a surprisingly short period of time,” Ms. Cotra told me. “People should take more seriously that A.I. could change things soon, and that could be really scary.”
Ms. Cotra is correct about this — but why should this be surprising? You can turn on a computer by flipping a switch. It’s because in stories, big changes are usually the result of big efforts. Something feels weirdly a-causal about AI suddenly showing up and taking over huge sectors of the economy.
Ironically, for AI researchers, these big changes are the result of a big effort, and they’re arriving more or less right on time! Much of the AI boom of the last 10 years relies on algorithms described in the early 1980s. When those algorithms were first described, researchers at the time also estimated how much computing power they would require and suggested that it would be about a century before sufficiently powerful computers became available. That 100 year estimate was updated twice due to two unforeseen upgrades — one was an algorithm called fast-greedy stochastic gradient descent, and the other was the proliferation of powerful graphics cards for playing video games — these two upgrades cut the waiting time by about 2/3rds. So the exciting applications we see today are arriving almost exactly as predicted — what’s unique about the story is that it took a forty year break for computers to get faster!
Journalists tend to focus on what is happening right now, and this hurts their ability to understand forces that unfold over decades or centuries. Most AI Ph.D.’s however, spent 10 years banging their head against a single algorithm. With all due respect to Roose, a few weeks messing around with DALL-E is not going to result in sufficient enlightenment to write him deep, thoughtful AI articles. I’m glad that he’s doing it, but hey, let’s be realistic.
But fortunately for humanity, there are a lot of slow, big forces that are preventing AI from being hugely disruptive to human civilization as well. And since education is my job, here are four of them.
- Elegant embodiment.
The first force is the most technical and science-fictiony: AIs don’t have bodies.
The first order impact of this is pretty straightforward: It’s very hard for a robot to kill you by moving pixels around on a screen. It’s also pretty hard for it to take your job by nudging some numbers in a database.
But the secondary effects of this are even more profound. Humans are body-first organisms. Evolution equipped us with an elaborately engineered body, and then slapped in a brain to handle complexity. This means that the engineering difficulties of building a self-assembling, self-repairing, soft and conforming, many-degrees-of-freedom robot with tons of sensors and motors were solved before the problem of intelligent behavior was even addressed.
Currently engineers are a long, long way from working skin. A long way from sensors or self-repairing myelin. But even if these decades of necessary advancements to physical body technology magically took place tomorrow — we wouldn’t be able to build AI systems that could control such a body.
Most of our current algorithms rely on huge amounts of example data. The collective databases of humanity (mostly servers on the internet) store tons of text, images, and numerical data — and so tasks related to manipulating text, images, and spreadsheets are good candidates for training a machine learning system. But there is virtually no data about the movement of physical bodies. What little we have tends to come from systems like LIDAR point clouds that detect the outsides of bodies moving in space, rather than highly instrumented embedded sensors of the sort that are useful for being a body moving in space. A colleague of mine once got a very expensive, very sophisticated humanoid robot body as part of a grant and the entire first year of work was just trying to figure out how to keep it from ripping itself apart because it had no pain sensors.
This means that no matter how good the image processing of current AIs, they can do nothing with this information. The gap here is far larger than is generally appreciated. There is nothing in the task of visually identifying a pen on a table that tells you that you can pick up that pen from that table. “Picking up” information comes from being embedded in a body that can do things like picking up. Absent a body with motors visual tasks tell you absolutely nothing about motor tasks and most of the information won’t transfer.
This means that progress in physical space will tend to be slow and incremental. A good example of this is self-driving vehicles. Driving is (usually) an exceedingly simple task that machines are well suited for. But progress has been extremely slow and advances have come largely from better engineering of the car (e.g., better sensors) than the AI. We should expect similarly incremental work for most large physical objects.
It also probably shouldn’t be a surprise that op-ed writers — who make media and exist on the internet — overestimate the impact of AI technologies that primarily make online media.
The second force holding back the robots is lawyers.
One of my favorite examples here is court stenographers. A court stenographer is the most automatable job one could possibly imagine. Your role is to simply record what is said in court. This job has been replaceable since the phonograph. Even if one includes the role of transcription, this job is now easily performable (except in a few edge cases with accents), by speech-to-text systems. In fact, most court stenographers *do* record court proceedings and run it through a speech-to-text program. The elaborate use of a steno machine is increasingly an act of kabuki theatre.
And yet, court stenographers are so in demand that New York state will literally pay for you to become one! It’s a decent middle-class job with good benefits. That’s because a court stenographer’s real job is not to write down what happens — it’s to go to jail in the event of an emergency.
The role of a court stenographer is a legal fiction: in an advocacy based legal system, neither the prosecution or defense can be expected to keep accurate records — that would be in conflict with their clients’ interests. The judge also can’t be the record-owner: they are meant to be a neutral party who decides fairly. If the judge isn’t fair, we surely can’t rely on their own records to prove it. Thus there needs to be a fourth disinterested party in the room: hence the court stenographer.
This person needs to be responsible for the record. The mechanics of what they actually do don’t really matter, only that they are the owner of the record. And what this ownership means is that it is the stenographer who is liable for any accidents. They are fined, they go to jail, and they care about their life, wealth, limb and reputation.
Robot’s don’t have stakes, and so liability is meaningless to them. They can’t die, don’t own anything, don’t care about jail time and so on. Much of our society is predicated on the notion that death and punishment are to be avoided. Machines do not avoid these things.
Again, this is a major impediment to self-driving cars. Trucking is an obvious industry for such vehicles. But currently most trucking corporations are small firms with 10–30 trucks. This is largely for insurance reasons. Big fleets are more exposed to long-tail risks like running over someone’s kid, or loosing and entire container of iPhones. As trucking corporations grow past a certain size they tend to be liable for catastrophic accidents that result in bankruptcies. To defend against this, companies often hire independent driver contractors who carry their own insurance.
No technical advancement will ever fundamentally alter these legal and economic dynamics. I can easily imagine a future where trucks drive themselves, but a human still rides around in each one to ensure that there is someone to die, be impoverished, or go to jail if something goes wrong.
Unlike truck drivers, op-ed rarely carry liability insurance to protect their readers in the event of mischief.
The final force impeding the rise of the robots is friendship. Not in the sense of hugging and playing pretend, but in the sense of social trust. Most jobs involve social relationships beyond the specific task requirements these roles require. The trust and interdependence engendered by these social bonds are not transferable — they are specific to this individual relationship and interaction. A mother who builds a bond with a child is never immediately substitutable for another mother because the social bond is specific to those two individuals. Few roles are as specifically socially bonded as parents. Most roles involve an amalgam of trust-laden social bonds, cognitive tasks, embodied actions, and legal liability. A nurse is a great example of a job that hits all of these notes.
It’s not that AIs can’t form social bonds with humans, they certainly can — see the recent hootenanny at google over the supposedly sentient chat bot — but these social bonds are also un-substitutable. If you socially bond with your phone’s chat bot a la Joachim Phoenix in Her, you don’t want a different chat bot. Your bond is specifically about the experiences you had with your chat bot, and a different chat bot would not share these same experiences with you. Even if you loaded the memories of your prior chat bot into a new one, you would simply interpret the new chat bot as being the old one because it could refer to the same prior experiences — in exactly the same way you interpret yourself as the same person as the one who went to bed yesterday despite the obvious gap in conscious experience.
And so in the venue of social bonds, AI’s face a massive competitive disadvantage relative to humans — they are weird robots who don’t have a lot of experiential common ground with humans — and even if they improve drastically, they can still only be as good at social relationships as a human, because it’s not the skill at social relationships that matter it’s simply the fact of the social relationships. My son screamed at me a lot when he was a baby, but I love those negative experiences far more than most of the charming, socially adroit interactions I’ve had with adults.
Social trust is built in other ways as well. Scientists for example put their reputations on the line by challenging the larger community to falsify their claims. They make bets of honor. AI’s have gotten into this game recently as well, with implementations like AlphaFold making predictions about the shape of expressed proteins. Far from worrying about losing their jobs, scientists felt more secure: because now they have to investigate these millions and millions of predictions. X-Ray crystallographers now have thousands of years of work ahead of them — and this work has to be carried out in real, physically embodied spaces to see if AlphaFold’s predictions comport with physical reality.
But op-ed writers? I have very little social stakes in my relationships with them, and they don’t seem to lose much authority when they’re wrong.
AI is an existential threat for media makers
It makes complete sense for tech writers to adopt an apocalyptic tone when discussing AI. Their content creation takes place entirely in the realm of words and images. They do not indemnify readers against bad outcomes. Unlike with original art, where fans have specific (para-)social relationships with artists, journalists and tech writers are mostly interchangeable cogs. Do you, dear reader, care who is writing this article? It is a brief puff of anxiety provoking prognostication about robots that you are looking at so that you can signal to others that you are intelligent and well informed.
Recent breakthroughs in AI are tremendously exciting and they can do a huge amount of work currently done by humans. However, it’s important to remember that the narratives of robots coming for our jobs are amplified by the fact that the robots are coming much more quickly and directly for the jobs of people who write articles online. GPT and DALL-E are existential threats to tech bloggers. They are of negligible danger, and minor help to the larger, grass-touching economy.