Marathon day. An early train into London, then an unfamiliar journey across a race-disrupted city from Paddington to Blackheath, all in good time for the start of the race. I was nervous, of course, but was cheered by the sight of another bib-wearing runner — more experienced at marathons, less familiar with London.
Me: “How do you plan to get to the start line?”
He: “I’ve asked ChatGPT. It says Elizabeth Line to Liverpool Street, then the train to Blackheath.”
That didn’t sound right. Was there a train from Liverpool Street to Blackheath? Google Maps and Citymapper suggested getting to Blackheath from Charing Cross or Waterloo.
Me: “Are you sure? I’d suggest the Circle or Bakerloo to Charing Cross.”
He frowned for a moment and pulled out his phone. “No, ChatGPT says that ‘The Circle Line is not a good choice on marathon day. It will be too crowded. There are too many stops and too many steps. It’s a route for tourists, not for runners.’”
I checked Google Maps. Sure enough, there is no train from Liverpool Street to Blackheath. ChatGPT’s recommendation would leave him stranded, trying to catch a bus over the marathon route, then trying to get on to the train from Charing Cross at a busy London Bridge. I told him that sounded like a bad idea. He frowned again and typed another query into his phone. “Oh, you’re right. ChatGPT says, ‘Correction: take the Elizabeth Line straight to London Bridge.’”
Me: “The Elizabeth Line doesn’t go to London Bridge.”
You’ve heard tales of artificial intelligence hallucinations before, but it’s not the AI that fascinates here: it’s the human.
The route-finding algorithm on Google Maps is a minor miracle. It will solve a complex optimisation problem across multiple modes of transport, taking into account real-time congestion or delays, and it’s been available on smartphones and browsers for years. It is a proven, practical example of AI in action. So on marathon day, when the stakes are high and the clock is ticking, why would anyone turn instead to a fancy word-guessing machine such as ChatGPT?
Perhaps it’s that ChatGPT seems so human. It served up an uncanny impersonation of a friendly and knowledgeable local guide. The Circle Line? Pfft, it’s fine for tourists but you’re a marathon runner: think about all those steps! (It’s true, the creaky old Circle Line does have steps.)
Part of the bot patter reminded me of clickbait ads: INSURANCE COMPANIES HATE THIS LOOPHOLE! ChatGPT wasn’t just giving a route, but giving a rationale, even explaining why we shouldn’t listen to the lamestream advice of Google Maps. This is the approach of a confidence trickster.
In the introduction to her book The Confidence Game, psychologist Maria Konnikova explains: “The true con artist doesn’t force us to do anything: he makes us complicit in our own undoing . . . we believe because we want to.” One difference between the con artist and the large language model (LLM) is that the con artist knows the truth and is trying to conceal it. One similarity between the con artist and the LLM is that both of them have perfected seeming plausible.
A recent paper in Nature finds that when LLMs are trained to be warm and friendly, they also produce dramatically less accurate answers, “promoting conspiracy theories, providing inaccurate factual information and offering incorrect medical advice”.
That sounds bad. I’d suggest that the reality is worse: the sycophantic AI not only produces mistakes, it persuades us to believe them.
In 1950 Alan Turing, the mathematician and visionary of the computer age, famously proposed an “imitation game” in which a human judge would communicate through a teleprompter with a human and a computer. The computer’s job was to imitate human conversation convincingly enough to persuade the judge.
Turing’s test remains intriguing, but there is a longstanding difficulty: the fallibility of the judge. A primitive 1960s chatbot, Eliza, responded like a parody of a therapist (“How does that make you feel?” “Why do you feel sad?” “Please go on.”). People lapped it up; it’s nice to feel listened to. A 1980s chatbot, MGonz, just fired off insults and was perfectly plausible, partly because insults are simple to deliver and mostly because they prompt rage rather than reflection in the human recipient. And Robert Epstein, an expert in the Turing Test, has written entertainingly about how he was fooled into a four-month correspondence with a sexy Russian lady who was, in fact, a 2006-era chatbot. None of these bots had a thousandth of the sophistication of a modern LLM, but they didn’t need it: when humans are sad, angry or amorous, we aren’t very sophisticated judges, either.
We are all going to find ourselves in strange variations of the Turing Test in years to come, and I wonder if we are up to it. And not just us, but those with power over us. As Cory Doctorow, author of Enshittification, is fond of observing: you won’t be replaced because an AI can do your job, you’ll be replaced because an AI salesman convinces your boss that it can. If my journey to the marathon start line is any guide, that salesman will have an easy job.
The capabilities of modern AI are impressive. But what determines whether we use it is not the capability, but the impressiveness. They are correlated but they are not the same thing. There’s a tale about the French poet Jacques Prévert seeing a fellow begging for change on the streets of Venice with a sign that read “Blind man without a pension”.
Prévert stopped to chat to him; not many people were moved to contribute, and Prévert offered to write a new sign.
The next day, he returned to find the man overjoyed. “It’s incredible; I’ve never received so much money in my life.”
Prévert had written: “Spring is coming, but I won’t see it.”
The new sign contained no news — in fact, it was less informative than the old. But it told a story. Google Maps was the first sign: it told me where to get my train. ChatGPT was the second sign: it told my companion not just where to go, but how to feel about taking such a clever route.
I left him at Paddington, urging him not to try to take the non-existent Elizabeth Line train to London Bridge. I am not sure I was as convincing as ChatGPT.
I ran the London Marathon in support of the Teenage Cancer Trust – not too late to make a donation.
Written for and first published in the Financial Times on 6 May 2026.
After this story was published Google's spokesperson reached out and asked us to publish a slightly different version of that statement. The new statement no longer stated that "it's critical that we maintain humans in the loop."
In a carnival game, you roll seven ordinary dice and then arrange them to form a 7-digit number.
If your number is a multiple of 2, you’ll win £2.
If your number is a multiple of 3, you’ll win £3.
If your number is a multiple of 4, you’ll win £4.
If your number is a multiple of 5, you’ll win £5.
If your number is a multiple of 6, you’ll win £6.
If your number is a multiple of 7, you’ll win £7.
The catch is that you have to announce the prize you’re attempting before you roll the dice. Which prize should you pick?
At first it seems that the £2 prize must be best. If even one of the seven dice produces an even number, you can put that at the end of string and fulfill the condition. This will happen 99.2 percent of the time.
Surprisingly, though, choosing 7 has an even higher success rate, 99.997 percent! “In fact, almost all numbers can be rearranged to make a multiple of 7,” writes James Grime. “But finding the multiple of 7 is the tricky part.” See the paper below for a strategy that will win the jackpot nearly every time.
American roads have become increasingly unsafe in recent years. There is disagreement about both the causes and solutions to this ongoing problem. I am not going to pretend to solve this issue, but I do offer one explanation: the 16% of Dangerous Drivers Dilemma.
These in the 16% are the drivers who you see darting in and out of traffic, weaving at dangerous speeds on the highway. They are also the drivers who are constantly fiddling with and checking their phones. Every stoplight is a chance for them to get a hit on their addiction—perhaps not as dangerous, but certainly annoying delays.
This 16% concept encompasses drivers who are old, young, addicted, angry, and generally incompetent. In this article, I walk through the numbers and consider ways to grapple with this dilemma.
A simple illustration of the 16% of unsafe drivers.
This 16% number I landed on has been derived from the concept of the normal distribution. Plot any type of skill or trait and you will find a bunched-up portion of humans in the middle, with a spread of variation emanating out from the average, on both extremes. This is the normal distribution.
The normal distribution is the spread of scores or measures within a population. There will naturally be a majority of scores in the middle at the average, spreading out with fewer responses the further towards the extremes. Of course, there can be subset population variations.
With the normal distribution, half of the observations will be below the mean and the other half above the mean. One standard deviation from the mean in either direction is 34.13% of the observations. Within two standard deviations is 13.59%. From three, it’s just 2.14%, fourth is .13%, and anything beyond these extremes is miniscule. I get 16% from the percentages beyond one standard deviation: 13.6%+2.1%+0.1% = 15.8%. I round this up to 16% just for ease of the moniker.
The normal distribution pops up throughout everyday life. The most well-known application of the normal distribution likely comes from IQ or other forms of testing, which I recognize is controversial (where most people relate these principles to a Bell Curve). Nonetheless, these tests are meant to be normally distributed. But even throwing this one out, there are other examples to illustrate normality of human capabilities.
The normal distribution works for basic demographic traits such as height or weight. But also works for skills, such as basketball or other sports. Only about 3.4% of high school basketball players make it to play in college at the NCAA level, all according to data from the NCAA. Of those who made it to college, only about 1.2% make it to a top pro league like the NBA. College basketball players are in the 95th percentile, while NBA players are in the 99th percentile on the normal distribution of basketball skills.
Approximated Normal Distribution of Men’s Height in the US, via Michael Minn.
These same breakdowns work with driving. Drivers for NASCAR are the same extreme end of the distribution for driving as their NBA counterparts in basketball. Like any skill, most people will sit near the average of the driving distribution. Yet, there will inevitably be a group at the end extreme end, opposite of NASCAR, that make up the 16% of dangerous drivers.
The driving test gives us one proxy for this concept. Data shows that 78.8% of American drivers pass the skills version of the test (the driving portion), only a few shades less than the 16% mark in terms of failure rates. Chalk some of that up to test nerves. Some may take multiple times to pass, too.
While we do not have great data on driving tests from the US, the UK provides national passing rates. In the UK, roughly 49% of drivers pass the practical driving test on the first try, aligning perfectly with the normal distribution. Drivers who pass the test in the second or third time account for almost 37%, which again aligns closely with the normal distribution between the mean and 1 standard deviation. Finally, a little over 14% needed four, five, six, or more attempts to pass. In fact, some drivers need 30 or 40 attempts to pass. Those in the extreme end are all the 16%.
Another support for my 16% conception comes from the field of traffic engineering. When deciding what speed limit a street should have, traffic engineers do a study to see how fast drivers typically drive in the area. The engineers then average out all of the scores and set the limit at the 85th percentile of all driving speeds.
This means that speed limits in the US are explicitly tied to drivers at around basically one standard deviation from the norm. The practice tracks with the 16% conception. Traffic engineers quite literally believe in this range of human capabilities to successfully navigate roadways.
The process for setting speed limits in the US via USDOT.
In recent years, the profession of traffic engineering has been heavily criticized for too much adherence to car-centric environments. Strong Towns, in particular, has taken aim at these tactics. The organization’s founder, , says that the 85th percentile rule can be useful for high-speed streets, but it is not a one-size-fits-all solution:
A road is different than a street, which is a platform for building wealth within a community. On streets, where there is vastly more complexity, the 85th percentile speed is almost always misapplied. Especially where a forgiving design approach is used for street design, drivers will not perceive the complexity of the environment. When engineers widen lanes, create recovery areas, make curves more gentle, and use other design features developed for highways, they give the driver the incorrect signal that the environment is simple, that they can relax and not be hyper-vigilant. Speeds go up and tragedy is the inevitable result.
Despite the critique, it shows that there is a logic to the 85th percent rule; it is just taken to the extreme in most of our built environments. This is not a test of driving ability, as I am suggesting with the 16% concept, but rather just of speed. It does, though, give us a practical proxy that is used in the real world. I’m sure traffic engineers would love to have some real driving test for their speed studies.
Insurance companies know all about the 16% already. They price in the fact that not every driver is the same as the other. It is industry standard to charge different rates for various driver types. Very young drivers get charged the most, then the rest of us pay some kind of standard, and there are a range of ways to reduce that with good driving record.
Insurance companies do this kind of pricing because they are not in the business of losing money. On the contrary, they are quite profitable, and part of that business success is recognizing the 16% of bad drivers. In general, they track with age as a proxy, with young drivers paying the most by far. There is a curvilinear relationship to age, though, as it goes up, insurance rates go down. This is true until about 60, then rates start ticking back up.
It is a crude measure of the normal distribution, yet still illustrates the concept tied to real-world money. Follow the money! There is a reason why young men in particular pay the most: they are more aggressive, irresponsible, and stupid behind the wheel. The insurance industry accounts for this disparity in driving skill.
According to Matt Timmons, author of State of Auto Insurance in 2025, insurers will offer sharp discounts to perceived good drivers.“This includes a 10-15% discount for drivers who take a defensive driving course, a 10% discount for safe drivers who’ve been accident-free for over 5 years and a 20% discount for drivers with low mileage use,” he said.
The insurance industry is very lucrative. They are not into losing money. Even they can see the 16% Dangerous Drivers Dilemma.
The 16% label is not a Scarlet Letter. Instead, we age in and out of being in the 16%. Most very young drivers are terrible drivers by dint of experience. There is simply a learning curve for learning to drive. But there is also an immaturity factor. Teenagers think they can live forever and nothing can hurt them, with tragic consequences.
Confession time: I was in the 16% when I was a teenage driver. So were most of my friends. We were all idiots. And my own high school experiences saw multiple vehicular deaths that still weigh on me to this day. After a few years, young drivers age more into the middle part of the distribution to safer and reasonable driving.
Luckily, young drivers usually age out of this irresponsible stage, which is what happened to me. However, on the other end, drivers can age back into the 16%. As we get older, we lose our faculties. Our eyesight gets worse, our reaction time slows, and our mobility diminishes. Driving skills regress. Data from AAA show that young drivers are the most dangerous, but senior drivers are more unsafe than middle ages.
This degradation of physical abilities is why a lot of senior drivers drive very slowly. Driving slowly is certainly safer than the maniac teenage driver, but mixing slow with fast speeds into traffic causes its own problems. Again, it goes back to the 85th rule. Likewise, even at slower speeds, vehicles can be deadly if the driver loses control of the 5,000 pound SUV.
But there are even more issues with seniors losing their ability to drive, yet continuing to drive out of necessity. There has been a slew of tragic deaths recently due to senior drivers losing control of their vehicles and plowing into innocent bystanders. This is becoming a growing issue as the Boomers enter these age ranges.
Many cities have noticed that there is a small group of people who have an enormous amount of traffic violations. In fact, much of crime operates in the same way. Remove a small number of extreme end criminals, and a greater total percentage of crime will decrease, argued in the aptly titled paper “1% of the population accountable for 63% of all violent crime convictions.”
These are called “super speeders.” They make up a good portion of the 16%, and likely the least sympathetic. Often, this is not about physical ability, like rookie drivers or aging seniors, but just a total disregard for anyone else outside of their own car. States have been grappling with these super speeders across the country.
For instance, in New York, those labeled as “super speeders” average roughly 179 speeding-in-a-school-zone tickets, which equals a traffic violation every other day of the year. Georgia and Maryland recently passed "Stop Super Speeders" bills. Virginia and D.C. did something similar, too. The extreme end of the normal distribution for driving causes so much traffic violence and damage that states must make special efforts to grapple with this subset.
Automatic ticketing cameras are another tactic against super speeders. A pilot program in Oakland installed 35 speed cameras at 18 locations. The results showed that 64.4% of drivers who received a warning did not get another ticket. But 30% of drivers received two to five tickets, 3.7% six to 10, and 11 or more at around 1%.
These types of cameras do work to a degree. “San Francisco’s speed cameras slashed excessive speeding from 25% of traffic to just 2% to 6% within a year,” according to the LA Times. But in New York City, the automatic ticketing has a lower fine than if written by a cop, hindering some of the sting to super speeders.
Ticket accumulation need not be considered the best predictor of crash likelihood, but around 1 in every 8 persons (12. 66%) who accumulate more than four tickets within an observed two-year period are later involved in a CAS-recorded crash, where they are the driver at fault. This represents a large concentration of risk compared to the overall base estimate of 1.22% incidence over three years. This risk is clearly identifiable in a particular portion of the driving population because records are kept for every ticket.
While speed is the most important factor, these 16% go beyond just one measure. There is a general flippant demeanor to dangerously driving a 5,000-pound vehicle. In Maryland, one driver had over 900 unpaid parking tickets.
We could lop off some serious dangers if we just took these people off the road. Our laws are usually are forgiving to license suspension, even when they have some kind of horrible incident. And I don’t use accident here. Whenever someone like this kills or maims another person, it is not accidental; it follows a pattern of risky behaviors.
There are some aspects of the 16% dilemma that have been exacerbated in contemporary America. Our stroads do not help anyone, whether in the 16% or not. One of the key problems is the phones (I know, the root of so many modern ills).
I didn’t drive for almost 10 years, from roughly 2009 to 2019. This is the exact time period of the proliferation of the smartphone. The differences in driving then to now are stark. Everyone is on their phone. I am dumbfounded by just how often I see a fellow driver looking down at their phone. Given how dangerous it is, there should be more shame in it.
But people do not care, they simply look down at their lap, then snap back up like awoken from a delirium. In a traffic line, these phone drivers cause delays by limiting how many people can make it through a light. When they do snap up, they try to jam on the gas to catch back. This just creates more unsafe situations, all for a screen addiction that we have normalized.
On the extreme end, there are a lot of 16%ers who are simply driving around on their phones. It is stunningly dangerous. Yet, you will see multiple people doing on any commute. Next time you are driving, try to count how many you see on their phones. It should be about 1/16 drivers or so with their head looking down.
Another problem is that American cars have ballooned to ridiculous sizes. Bigger cars are simply more dangerous to everyone not inside of them. There are just realities of driving that relate to physical space. A person of smaller stature will have a more limited view in certain larger types of vehicles. They will be a danger to other people on the road because they have lower vision. There is no way to get around that other than simply driving a smaller car.
Height limits on certain types of cars would make sense. But there would be some kind of civil rights appeal. This is a very American problem, perceived personal freedom is a higher value than consideration to broader community, even if the behavior is risky or detrimental. See one example from Florida recently:
The Delirium of Dangerous Driver
As a society, Americans have been resistant to removing the 16% from out roads. Our overinflated, auto-centric design has limited our choices to everyone having to drive. Walking, biking, or taking public transportation is deemed too cruel and unusual for Americans. at recently published an article detailing just why it is so difficult to stop super speeders. In short: we aren’t allowed to.
There are various laws that limit punitive measures. If someone doesn’t even hurt someone, just got caught driving in a dangerous manner, there is even more leeway. In one egregious case from Pennsylvania, an off-duty cop hit a family in a crosswalk because he was texting. He faced no real consequences, not even a suspended license, according to local news coverage. Stories like this can be found across the county.
There are arguments that speed governors are dangerous or too inconvenient. The problem with these arguments is that most of us cannot understand speed. “Quite simply, the faster you are going already, the less time you save by going 10mph faster still,” says Rory Sutherland for The Spectator. “Accelerate from 20-30 mph, and you save ten minutes on a ten-mile journey. Accelerate from 70-80 mph, and you save just over a minute.”
The 16% just make things worse with a lack of object permanence, jamming on the gas only to have to slam on the brakes because of cars in front of them. This creates more traffic because they end up bunching up with other cars when weaving in and out of traffic. We are too lenient on this subset of driver. They make everything worse. Yet even when states try to crack down, there are limitations. New York’s news Super Speeder Bill has been hailed as admirable, but it only installs governors for those with 16 tickets in a year. 16! That’s over one per month.
of has been critical of the unwillingness to punish these super speeders. He tweeted recently of the new New York law: “Don’t ever forget this story. Leading up to the crash, the killer had:”
“21 speeding tickets in 2 years”
“6 red light camera violations in 6 months”
“70+ other violations in 2 years”
He closed, “She was still free to drive at a murderous speed.” I agree. We must no longer accept the scourge of the 16% on the rest of us. Do we really believe we can stop maniac driving like this by asking nicely? Of course not.
Despite the obvious problem, there is not much political will in the country to grapple with super speeders, reckless drivers, or the 16% in general. Democrats tend to lean away from more law enforcement due to equity concerns, while Republicans have become distrustful of more government control, even if traffic enforcement. Changing laws and adding more enforcement is just not always feasible electorally.
This is where the Strong Towns approach can come in to address the problem. In this conception, street design should dictate the comfort of driving. If we want drivers to slow down, we need to make lanes narrower, add bollards or barriers (real ones, not plastic flexipoles), cover with overhanging trees, and other general traffic-calming measures. I have advocated for these things myself (see my video essay on the topic),
One street redesign example from San Diego, California.
What the Strong Towns approach does is take safety out of the hands of waffling politicians. In a properly designed street, a 16% driver will constantly show themselves by damaging their own vehicle. We want them to hit bollards and scrape concrete walls. This outcome is certainly better than flesh and blood of innocent pedestrians.
A driver who cannot successfully navigate a fast food parking lot without smashing their own truck they cannot be trusted to not hit a child. The 16% always reveal themselves.
My guess is that anyone who gets mad at this 16% concept is in fact in the 16%. They likely cannot stay off their phone, have heated road rage, or just simply struggle with the basics of driving. No matter, getting bad drivers off the road should not elicit hate. It should elicit cheers. And since we cannot always do that, we can help protect the rest of us from the 16% of dangerous drivers through smarter street design.
Thanks for reading College Towns. You can subscribe for free and without a Substack account (just an email).
My Patreon, featuring my full interview with Richard McClintock! It’s entirely free, forever: https://www.patreon.com/c/rabbitholevideo Thanks to the Patreon GOATs: Melanie Lin, Andy, Mathew Vicknair, Jay Hsin, Joshua Loduca, Casper, Alexander Hyunh, Robin Lowe, John Early, Eric, Barbara Abraul
*– Chapters –* 0:00 Why is this strange text everywhere? 2:09 The professor’s discovery 3:53 45 BCE to 1500 CE. Or not. 7:03 1914 - The Rackham Book 10:37 1987 - From print to digital 15:05 1966 - The Letraset hunt 20:12 An emailed answer… 22:42 Rewriting the mistake 25:25 The end :)
Log in
