I am always looking for metaphors to explain what AI looks like to me when people ask, because I have been writing about it for going on 35 years, since the days when it looked like genetic algorithms and expert systems were the thing. These days I have been settling on traffic, just because we all understand it, having sat with the inconsistencies in it all. And universally, we are quite clear that “we” are not part of the problem. It’s the other drivers, the driving registry, the potholes, or the roads. Lately it’s those self-driving cars. But of course, it’s never us.

I am likely part of the problem, even though I 'know' I am a good driver. And it does not hurt that, in America at least, those blemishes on your driving record and with the insurance companies sort of reset after a time. I am a good driver. I know this because my driving record is clean, and sometimes I even drive calmly and smoothly enough that one might hold a martini glass and not spill a drop on a long journey.

That said, I also recognize that others might disagree. My wife certainly thinks I’m a horrible driver. She’s always informing me about this, that, and the other thing, particularly on these English roads where the rules of driving are so different. In America, we don’t have roundabouts, and every road is a full lane in either direction. It’s not like you have to play this game of chicken at high speed down a narrow road packed with cars, where one of you has to commit to sitting off to the side and negotiating. No, in America, you drive.

I bring up that martini glass because years ago, when I was maybe 19 or so, I was driving an older woman around. I had just made a U-turn on Market Street in San Francisco. I had made that U-turn quickly because we had missed something, and she said, “George, I want you to imagine I’m holding a martini glass. I want you to drive such that I don’t spill a single drop: slow turns, slow acceleration, and not just slow braking, but planning for the braking.” Could I brake such that she never actually felt that rush of deceleration, or, conversely, accelerate smoothly?

I didn’t know it then, but she had handed me the secret to make sense of what AI has become, even though it was not a thing yet. The whole trouble with driving, whether it’s a car or the signals we share (that’s the AI equivalent), is that you can’t feel your own driving. You feel the road, the other cars, the clock. You don’t feel the jerk of your own hand on the wheel the way you feel a hand someone lays on your shoulder. The martini glass fixed that. It sat there on the imaginary tray between us and reported back on the one thing I was blind to, which was what I was actually doing.

I would posit that 99% of all humans, except the odd chauffeur, would fail the martini test catastrophically on 100% of their journeys. They don’t plan braking, they don’t pace acceleration, and they don’t turn with grace and elegance. We always seem to be in a hurry, spun up by other drivers, or maybe we’re just not paying attention. That’s pretty much how we treat the signaling layer, which is really what AI is all about when you unpack the physics, networks, data centers, algorithms and social media platforms that run on them. The tell is when you catch yourself yabbering on about some polarizing post you read in the news or social media. Don’t get me wrong. Fast acceleration and braking can be fun. You just might spill a few drops along the way.

In fact, I would assert that you too are a perfect driver, despite whatever evidence might indicate otherwise, and everyone else is a horrible driver. If you’re of that frame of mind, you might categorize them. I’ve heard men talk about women drivers being too distracted, and women talk about men drivers being too testosterone-riddled. In America, they talk about Chinese drivers. The AI equivalent is all those other people we like to talk about, which, depending on your bent, might show up as the woke, MAGA, or your least favorite politician. Those signals, which seem to be getting louder by the year and more quickly, are the AI equivalent of a traffic jam with no apparent cause, no accident up ahead holding up the flow. It is, of course, other people’s fault because the signaling layer is hard to unpack when our less-than-helpful thoughts are part of the problem. Like me, you are probably a perfect thinker.

The physicist David Bohm, in his little book On Dialogue, gives us an interesting turn of phrase for describing this very same phenomenon:

It’s they who are thinking all that, and I am thinking right.

He calls this a mistake that seems small on the surface, but since we are all doing it in a gazillion different ways, and now faster thanks to more, better, faster algorithms and networks, it scales up. What he asserts we should examine is how thoughts that are really ours have been quietly filed under someone else’s name. Whatever is wrong in your thinking gets hidden in the one who is doing the looking:

Because that is the safest place to hide them. Hide them in the looker, and the looker will never find them.

The bad driver is always in the next car or the next meme. That’s not just an observation about traffic, memes or all of the other things we point to when we talk about AI. It’s an observation about where we keep the parts of ourselves we’d rather not observe. Now, for me, I know I’m a perfect driver because some days I actually do drive smoothly, as if my passenger had a martini glass. I’ve got the presence of mind not to spin them up with tales of how horrible all the other drivers are. If they want to look, they can see it for themselves. They don’t need me yabbering on about it.

That’s why I think my wife’s a bad driver. She’s always spinning me up about the other drivers, and she gets spun up too. I remember just a few days ago she was chasing someone’s bumper at 80 mph so as not to let another horrible driver somehow get in front. This seemed really important to her. For one, I felt a little fear that we were moving so fast. Second, I knew it went against the recommended guidelines spelled out in driver training books in both the U.S. and the UK. To actually be safe according to those standards, which pretty much no one follows, you’d have to leave a few hundred feet to come to a full stop if the car in front of you has an issue.

Bohm describes a woman who had a stroke in her sleep and woke to find herself under attack. Someone was hitting her in the dark. People came running, turned on the light, and discovered she was hitting herself. Her motor nerves, and the muscles they drove, still worked. But her connection to the sensory nerves had blown out in the stroke. So she experienced her own hand as an assailant that was outside of herself. And the more she defended herself, the worse the attack got. It was like chasing her own bumper at 80 miles an hour.

Now, after an experience like that U-turn, which was definitely probably not illegal, I might have learned my lesson. But of course not. Like you, I am human and constantly missing my destination. And most of the time those U-turns don’t result in any serious consequences. Except those two times someone else hit me, and it was their fault, like the off-duty police officer who ran into me while he was talking on his phone before that was illegal. Guess who was cited as causing the accident by the other police officer who showed up?

Now, I bring this up because we all imagine that humans are somehow better drivers than AIs. We have a bias against them. They’re different. Every time some AI autopilot does some bone-headed thing, like interfering with police responders, we hear about it. The latest one I saw is that they don’t know how to get rid of passengers who fall asleep in the back because they’re tired or drunk, or who are having a heart attack, or, in two incidents, having a baby. No, they don’t know how to do any of those things, but I’ll bet you that a lot more of them might drive such that you could hold a martini glass and not spill a drop. Good luck trying to get an Uber driver to do that. They seem to be in the biggest hurry. Is it their fault, or is it those algorithms spinning them up in the process of earning their keep?

I lost nearly a quarter of a star for calling out an Uber driver for his interesting driving style. I was no longer a five-star Uber passenger, but merely a 4.76 after that fraught journey. I think it’s like that with a lot of things. We have this self-image, and it’s impossible to hear information that says otherwise.

Here’s another example. I was suffering from hip issues for about seven years, trying every alternative thing to get better: acupuncture, physio, Pilates, Feldenkrais. Really, there’d be pain, and I’d be walking along, but I didn’t limp because I’m a healthy person. The data said otherwise. I noticed that one day, watching one of those 360° cameras reassembling my walk, and I’m like, 'Ouch'. Later, my mom commented on it. Oh goodness, don’t get Mom worried about my health unless she’s paying for the surgery, in which case it’s helpful. She said, 'Oh, you have a limp', and I pushed back. Now, I was walking eight miles a day at the time, and I was proud of those eight mile walks, but I was having none of that limping business, even though her words and, worse, her look of concern said otherwise.

Shortly before my hip replacement, I read a book by a hip surgeon recounting his personal experience of showing up on the other side of the knife. He observes that when we ask people if they limp, 90% of the time they say no. Of course their partner often has a different opinion. That’s a funny thing. We have this felt sense that we’re good drivers. We don’t hallucinate. But really, we humans do hallucinate or conflate, and the blessing is that sometimes other people or even AI might call us out or point us in directions or facts we might not have found otherwise, or never even thought to look for.

Right now, I’m sitting in my room looking at my so-called blackout blinds, which aren’t blackout blinds. Why aren’t they blackout blinds? I spent half an hour telling the first guy from the company that I wanted blackout blinds, and we picked them. Then he measured my room, but those measurements weren't good enough. They had to send another guy to come in and measure my room. Then he said, 'By the way, the blind you ordered is not on offer anymore. You’ve got to pick a new set'.

He said, 'Here’s a set that’s like the one you had', and he pulled one up, and it looked white. I was happy, but they’re not blackout blinds, so who hallucinated there? Was it the first guy, the second guy, or this artificial organization that didn’t close the gap between the half hour I spent explaining how I really wanted blackout blinds and what I actually got? They’re nice. Still, they’re not the blinds I asked for, that I spent money on, that I waited a month for. Oh, by the way, they were going to deliver them on July 4th, and they didn’t show up until July 18th.

'Who hallucinated there?' is exactly the right question, and it doesn’t have a human answer, because no single human made the mistake. Bohm would call what happened incoherence. My intention and the result didn’t agree, and the signaling between the stewards of the company’s brand and me quietly conspired against those actual blackout blinds. That’s the artificial organization frame of it all that the Cambridge political scientist David Runciman talks about in The Handover: How We Gave Control of Our Lives to Corporations, States and AIs. He argues we have been building artificial minds for far longer than the latest crop of AI, and have been quietly part of them without really noticing.

This is why I find the current panic about AI hallucination so quietly funny. These latest incarnations of AI-generated content and lawyers censured for fake cases are still new and exotic. Air Canada even argued that its chatbot, which made up a bereavement refund policy, was a separate legal entity responsible for its own actions. The tribunal wasn’t having it. It’s never us. But this ungroundedness, call it hallucination or incoherence, has been going on for a while. 

A couple of years ago, a UK parking company sent me a penalty notice after its camera somehow read my exact number plate off a blurry photo of a different car, while mine was parked five miles away. After five hours of research and letter writing, they thankfully acknowledged their mistake. Runciman says the truly consequential handover already happened, centuries ago, when we let these artificial agents make decisions on our behalf. The newest artificial minds are simply stepping into a world already run by the older ones, just a little faster and at scale because of all the automation that clouds and newer models enable.

Now, back to those autonomous cars for a minute, because I think they have something to say about how we have been doing AI. Back in 2017, some researchers put an autonomous car on a track in Tucson, Arizona, with twenty humans to study the potential of pacing humans. On their own, and in this controlled environment, the humans kept doing the stop-and-go thing we are used to on real roads, which leads to all those traffic jams with no apparent cause. After adding that one autonomous car, the waves went away, and fuel consumption for all the cars dropped by as much as 40 percent. They estimated that programming just 5% of the cars to drive like that might smooth things out for everyone else. The gist is that driving with that virtual martini glass might chill out everyone behind you and not just your passenger.

I think a lot of companies have been driving AI like the humans on that track. They floor it, lay people off because AI is going to do their jobs, and then slam on the brakes and quietly hire people back or declare a so-called AI pause. Speaking of driving, even Uber CEO Dara Khosrowshahi  is talking about the need to slow down to get things right:

We see sometimes the results of trying to go too fast and some of these AI companies with data centers, they were pushing through, you could argue, too quickly and there’s been a huge public blowback against it.

The frontier Big AI equivalent might look like throwing all of those super smart models at sorting out all of the challenges with building more sustainable data centers that will power all of this without a huge public blowback, addressing the trust gap, and being a lot more methodical about weaving judgment into all the places where the LLM stuff mucks things up. You can’t floor it on any of that, unless you like protests and confrontation. You have to plan your braking, the way that older woman asked me to on Market Street all those years ago.

So yes, it’s certainly hard to notice that we humans have been incoherent and ungrounded individually and collectively, with lots of interesting cover stories, for a while. It’s just more noticeable now that the signaling layer is evolving faster. The challenge is that, individually or collectively, it’s always painful to sit with the possibility that there might be some gaps in our image of ourselves, of other humans, and of these artificial organizations that most certainly don’t hallucinate, not like those new-fangled AIs do. The bad driver is always in the next car. The limp is always someone else’s. And the strange gift of these 'AI upgrades' might be that they are offering us new mirrors, a new martini glass we did not ask for, so that for once we might feel our own hand on the wheel.