Friction Was the Feature

There was a time when applying for a job meant choosing a handful of roles, tailoring a resume, and writing a real cover letter. The effort was a nuisance, but it quietly enforced focus. If you were going to burn a Saturday on an application, you probably cared about the job.

Today, a candidate armed with an LLM can parse dozens of job postings, lift phrasing from each, and generate a set of keyword-optimized cover letters in no time. They can auto-tailor their resume to each posting. They can submit 30 applications in one sitting.

This is better, right?

Not for anyone, actually. Applications soar; recruiters drown. So we bolt on more automation: applicant tracking systems, resume parsers, AI interview schedulers. We convince ourselves we’ve built a better machine, but we haven’t redesigned the only machine that matters: the system matching the right people to the right work.

As one recent college graduate told WSJ after months of unanswered applications: “I was putting a lot of work into getting myself into the job market, and the ladder had been pulled out from under me. Nobody was there to really answer what I could do to improve… It was just radio silence.”

We automated the production of artifacts but haven’t fixed judgment. The result is a marketplace of immaculate verbiage with very little meaning, of noise without signal.

Everyone looks more “efficient.” Very little about it feels effective.

Recruiting is just one place where friction used to be the feature. When the marginal cost of creating words falls toward zero, any system that uses those words as a proxy for quality begins to fail.

This essay is about how that failure is spreading.

In 1865, economist William Jevons noticed something strange about coal. The Watt steam engine was much more efficient than earlier designs, so you might expect coal consumption to fall. Instead, it exploded. Making coal more efficient made coal more useful; more uses meant more total coal burned.

This pattern, now called Jevons’ Paradox, shows up whenever you dramatically cut the cost of something important. Instead of saving resources, efficiency gains often unlock new demand that overwhelms the savings.

Enter knowledge work. For the last few decades, the main scarce resource has been human time and attention. We used to ration that time with friction: you only wrote a long memo, or crafted a careful email, or applied for a job, when it was worth the effort.

Generative models have blown a hole in that logic. When the marginal cost of ideation, drafting, and refactoring words drops toward zero, we don’t get fewer, better artifacts. We get more of everything: more job applications, more emails, more grant proposals, more slide decks, more status updates. Talk becomes cheap.

Economists Anaïs Galdin and Jesse Silbert just gave us a detailed look at what this does to a real labor market. Using data from Freelancer.com, they studied how AI-assisted writing changed job applications on the platform. Before AI tools, a stronger, more customized proposal was a meaningful signal: a one-standard-deviation increase in their “signal” measure was as attractive to employers as cutting your bid by roughly $26.

But after Freelancer added an AI proposal tool, employers were flooded with high-signal proposals. The same increase in signal now only buys the equivalent of a $15 price cut (40% less), and employers’ responsiveness to writing quality keeps falling: workers in the top ability quintile are hired 19% less often, while those in the bottom quintile are hired 14% more often [1]. AI broke the system of judgment.

The dynamic shows up far beyond labor platforms. It helps to borrow a line from economist Charles Goodhart: “When a measure becomes a target, it ceases to be a good measure.”

For a long time, we treated certain artifacts as measures of effort or quality. A thoughtful cover letter. A carefully written reference on behalf of a colleague. A multi-page strategy doc. They were hard enough to produce that their existence told you something about the person behind them.

Once AI arrived, those artifacts quietly turned into targets. We told people to personalize outreach, respond to more emails, and write more detailed documents. LLMs optimize for those metrics beautifully. But as they do, the metrics stop tracking the thing we actually care about: competence, sincerity, fit, progress.

This is what I call “Goodhart’s inbox”. We’re surrounded by messages and docs that perfectly satisfy the surface criteria we asked for. But the more we optimize for these outputs, the less focused we are on outcomes: whether anyone is actually understanding each other or moving work forward.

Where is this doing the most damage?

Product reviews and recommendations: A decade ago, shopping online meant reading a messy mix of obviously human reviews. Now AI makes it cheap to generate synthetic ones. Regulators are scrambling to keep up: in 2024 the U.S. Federal Trade Commission finalized a rule banning the sale or purchase of fake reviews and testimonials, including AI-generated ones that misrepresent real experience. Amazon says it proactively blocked over 275 million suspected fake reviews in 2024 alone (begging the question of how many they missed). The volume of text goes up; our trust in that text goes down. The fix won’t be “more reviews,” but better proofs: verified purchases, photo or video evidence, identity checks, and ranking systems that weight verifiable behavior more heavily than elegant prose.

Cold outreach: I get a steady stream of cold requests for time, often from MBA students and early-career folks trying to break into product management. In the past, if a note was well written, referenced something we had in common, and made a clear ask, I said yes. Now many messages feel more polished but interchangeable: a neat story about their journey, a tailored compliment, a reference to a detail of my profile. Effort used to be a signal. But now that legible effort disappears, resulting in occasional inbound notes like this:

School admissions: Essays and personal statements were always imperfect, but they gave committees a rough sense of a student’s voice and process. Now almost every applicant can enlist a model to help brainstorm, outline, and refine a reflective, well-structured essay that checks every box. At the same time, volume keeps climbing: in the 2024–25 cycle, the Common App reports nearly 1.5 million distinct first-year applicants submitting more than 10 million applications to its member institutions, an increase of 8% from the prior 2023-2024 cycle. Committees respond with more automation and blunter filters: AI screens, heavier reliance on test scores and school prestige. Written recommendations are on the same path; a glowing letter no longer guarantees the writer cared enough to dedicate an hour for that candidate.

Warranty and return claims: Models make it easy to draft perfect warranty claims, complete with policy citations and the right tone. That feels empowering when your product is defective. At scale, it’s a nightmare. The National Retail Federation estimates that consumers will return about $890 billion of merchandise in 2024, around 16.9% of annual sales and more than double the $309 billion and 8.1% rate in 2019. Retailers are responding with stricter policies and new fees: surveys suggest roughly two-thirds now charge for at least some returns, and “return insurance” is becoming common. Once again, we have more words and more cases, but often a more frustrating path to actual resolution.

The loss of friction hurts systems on four levels:

1. Worse matching. When artifacts lose their signaling power, systems get dumber at pairing people and opportunities. In the Freelancer.com study, cheaper writing and noisier signals reduce the chances that high-ability workers are hired and make the whole market less meritocratic.

2. Time. We are writing and reading more than ever, but we aren’t necessarily moving faster. Every choice is instead wrapped in a thicker layer of prose: longer emails, longer proposals, more “option-exploring” documents. This chart from the same Freelancer paper says it all:

   

3. System complexity. Generative AI makes it trivial to produce highly detailed contracts, policies, and user agreements. Legislators can ask an assistant to add another clause to a bill. Companies can ask for another page of legalese in their terms and conditions. The practical check of human comprehension falls away, because no one person needs to hold the whole thing in their head anymore. We risk creating systems of law and policy that are technically precise yet socially illegible, and it becomes much harder to question them or hold anyone accountable for how they work in practice.

4. Fairness. When the legibility of effort disappears, institutions often fall back to cruder filters: brand-name schools, elite employers, personal networks. The people AI was supposed to empower—the scrappy, the non-traditional, the outsiders—can end up worse off because their hard-won effort no longer shows up as above-and-beyond.

The answer is not to ban AI. No one liked cover letters anyway, and we’re not reverting to a world where drafting a page of text costs an afternoon. The question is how to redesign our systems so that AI boosts signal instead of burying it.

Proofs over prose: If a model can manufacture the artifact instantly, that artifact is a weak signal. We should ask for things that are harder to fake. In hiring, that might mean retiring the cover letter and moving to paid trial projects or structured live exercises, graded against a transparent rubric. In admissions, it could mean time-boxed essay prompts in a proctored environment or video responses. For customer service, it might be as simple as requiring a short video of the defect with a timestamp, or a few specific diagnostic steps performed on camera.

Add good friction: We can reintroduce small, smart costs that make spam harder and commitment legible. Imagine a job board like Tinder’s “superlike” feature that limits you to a few “super applications” per month, each requiring bespoke answers or a short exercise. Or a cold-outreach app where, if the recipient accepts your request, you automatically donate $10 to a charity they’ve preselected. (Gated unsuccessfully attempted this in 2022, but that was before ChatGPT.) The premise is the same: friction that rewards commitment and punishes spam.

Use AI to compress, not inflate: Right now, a lot of AI adoption looks like putting a jet engine on the old workflow: we still do weekly status memos and grant proposals; they’re just twice as long. A better pattern is to use AI to shrink the prose layer and enlarge the decision layer. Let it generate options and counterarguments, then spend human time choosing and committing. Let it draft the research plan, then spend your energy in actual conversation with users. And finally, in an efficiency-driven world of synthetic prose, there’s more value than ever in the most human of things: showing up in person, building real relationships, treating people with kindness.

Regulators and platforms will keep playing whack-a-mole at the edges: banning fake reviews, tightening return policies, tweaking ranking algorithms. Those are necessary, but they can’t fix systems that are still built for the pre-AI era.

That part is on us: leaders, educators, hiring managers, and anyone who runs a process where words have become easy.

The friction was never the point. The signal was. If we move the effort to the proof that matters, and keep our targets pointed at outcomes rather than outputs, AI doesn’t have to bury us in noise. It can help us hear the music again.

— John

A draft of this post was edited with GPT-5. I’ll leave the reader to decide whether that’s ironic.