The death of deliberate friction

In 1978, CBS correspondent Robert Schakne submitted a Freedom of Information Act request for the criminal records of the Medico brothers, who had been winning defense contracts through a congressman later convicted of taking bribes. When the FBI denied his request, Schakne sued, arguing that every arrest and conviction of the Medico brothers was already sitting in the public records folders of various county courthouses, so the compiled version couldn’t possibly be private. 

The Supreme Court disagreed unanimously. In his majority opinion, Justice John Paul Stevens argued that while technically public, the friction of gathering hundreds of records scattered across county archives and local police stations was high enough to provide privacy to the accused. The deliberate friction of compiling the records served to provide practical obscurity, foundational to the system’s design.

Until very recently, this friction was embedded in systems everywhere, from calling to make restaurant reservations to filing insurance claims. But today, Schakne could presumably task Instinct or Muse with calling and emailing every bureaucratic keeper of the Medico brothers’ records, compiling the documents at the cost of a few dollars in tokens. Tasks like this have only gotten more trivial as AI assistants have become more capable and prevalent. We're seeing them consistently complete complex work with ease and proactively handle challenging or, perhaps more importantly, annoying tasks. Institutions rely on hassle costing people their time to keep them from doing things, but agents don't mind the hassle. What happens when deliberate friction erodes across all institutions?

One of the first tasks I used Instinct for was arguing with United Airlines about an inexplicable surcharge. I shudder at the thought of spending hours listening to terrible hold music and debating with an offshored customer service rep to get my refund. Given the surcharge was just a few dollars, I’d probably accept the injustice and move on. But given that the cost of having Instinct argue indefinitely on my behalf is free, why not? After two weeks of the bot being at work (and me doing absolutely nothing), United issued the refund as a goodwill adjustment, a retail term for when the support agent gives in because the cost of the contact ($5-10 of labor per call in a typical center) dwarfs the amount in dispute.  

Getting a refund used to cost an hour on hold and a fair amount of frustration, but now it costs almost nothing. And when transaction costs drop, people transact more. Many systems can't absorb that surge, because they were implicitly relying on friction to keep demand in check. The hassle was a load balancer, a core part of the system design.

Many systems are at risk of breaking due to the disappearance of deliberate friction. Resy is being flooded with a wave of reservation bots and popular San Francisco restaurants are booking out within seconds of tables being released. The NIH recently capped grant applications at six per person per year, after scientists used AI to submit 40 applications in a single round. The number of self represented plaintiffs in federal lawsuits has increased from 23,000 to 41,000 in three years. 

So far the bandaid solutions have been defensive, capping how many transactions anyone can make, and forcing users to prove they're human. But most systems need not be adversarial; neither the diner nor the restaurant benefits from the arms race of bots and bot detection. Instead of an auction based on speed or labor, systems will allocate access by some other criteria, a deposit, a lottery, a reputation score, or a novel protocol. Black Friday doorbusters used to be settled by whoever lined up and pushed through the doors first. The stampedes were resolved after Walmart implemented a one hour guarantee where anyone in the store during the window got the discount and the item shipped to them later. That's what a cooperative system looks like: both sides want the transaction to happen and just need a fair way to decide who goes first.

But cooperation only works when both sides want the transaction to happen. In a lot of systems, one side actively discourages the transaction but never built a system to regulate it because the sheer effort involved was a sufficient barrier. Justice Stevens’ doctrine of practical obscurity rested on the belief that the records were protected because nobody would do the work to gather them, but that protection is gone now that the work is free.

Practical obscurity is a primary mode of defense for a shocking proportion of our institutions. For example, vulnerability research in cybersecurity is not gated by secrecy. The code of most of the software that runs the world is public on GitHub and the bug classes to exploit are documented in textbooks. But finding a bug took a highly skilled engineer weeks of sifting through a codebase, and there were maybe a few thousand of those people alive. 

It’s only been a few months since that stopped being true. Now anyone can point an agent at a repository, let it run overnight and wake up to a list of exploitable bugs. Security teams at every large corporation have been running fire drills adopting AI cybersecurity tools to keep pace with the threat of AI vulnerability discovery. In September 2026 alone, Microsoft fixed a record breaking 974 vulnerabilities, almost as many as in all of 2025. We have handed machine guns to both the attacker and the defender. 

The systems that used friction as a queue will now need to redesign the queue and the systems that used it as a gate now have to build one for the first time. Facing the new market reality, I’m excited to see systems evolve to find a more efficient, positive sum equilibrium.