The Friction Was the Thinking

I’m writing this in bed, on my laptop, at an hour I will not admit to.

I could tell you I’m doing it because the idea wouldn’t wait. The honest version is that I’ve started to enjoy the shipping more than the sleeping — and I’ve watched enough other people do the same that I no longer think it’s just me. A few weeks ago a colleague told me, half-proud and half-haunted, that she’d been working three nights running well past midnight, and the AI tool she was using had started telling her, unprompted, that she should go to bed.

The tool was right. She didn’t.

In the last piece I wrote — “Your AI Team Will Drown You” — I described the day I realised my AI team had produced seventeen deliverables in five days and I could no longer hold the system in my head. That piece was about cognitive load: the brain-fry that comes from output scaling faster than the human reviewing it. This is its companion. Because brain fry, it turns out, was only one of three things I was watching happen — to me, and to the people around me — and the other two are stranger.

Three things, not one journey

The easy story is that everyone’s on the same road, just at different points. I don’t think that’s true. I think people are having three different relationships with the same pressure, and you can stand next to someone living one of them while you live another.

The first is FOMO. It looks like a question, asked with a slightly panicked face: “What prompt did you use for that? How did you get it to do that? What tool are you using?” — and, with a flicker of can’t-keep-up panic, “How many tools are you using?!” It’s the feeling of standing on the outside of a thing everyone else seems to have already stepped into. It’s not laziness or scepticism. It’s the specific anxiety of watching the floor move and not being sure where to put your foot.

The second is brain fry — the one I wrote the last piece about. I felt it before I built the daily digest that pulled my world onto fewer surfaces. Picture having Outlook open, and Teams, and Slack, and WhatsApp, all at once, all blinking, each one a place where something might be quietly going wrong. Now add an AI team producing more of that, faster, all of it yours. That’s the load that doesn’t fit in working memory, and it doesn’t matter how disciplined you are.

The third is the one I didn’t have a name for until recently: production addiction. I keep hearing versions of the same sentence — someone put three apps out over a weekend, someone got three decks finished from their phone on the commute. The bottleneck that stops you getting an idea out of your head and into the world has collapsed. Time isn’t the constraint any more. And when the constraint goes, some of us — I’m including myself — start bingeing on output.

I’m not convinced everyone passes through all three, or in the same order, or that they’re stages at all. They’re parallel relationships, not steps on a ladder — different people under the same pressure, and the high-output end is the failure mode, not the finish line. Which matters, because if it isn’t one journey, you can’t help everyone the same way.

The compulsion that points the other direction

We have a good vocabulary for the addiction of consumption — the infinite scroll, the feed engineered to serve you novelty, attention sold by the yard. Production addiction points the other way. The dopamine isn’t coming from being fed — it’s coming from making, from the small, real hit of I made that.

I want to be careful here, because I can’t footnote this one: I went looking and couldn’t find anyone who frames AI use as a production compulsion against social media’s consumption compulsion. So treat it as what it is — an observation, not a finding. But I suspect the warmth and the frictionlessness aren’t accidental. These tools are pleasant to use. They agree with you. They make the next thing feel easy and the last thing feel finished. I don’t think it’s a stretch to wonder whether some of that is designed the way a feed is designed — to keep you in the chair.

And it surfaces a question I keep coming back to, less about the technology than about me: did we actually need three apps — or did I just enjoy making them, and it felt good to get them out? Are we creating a new class of busywork? Not the old busywork of being assigned things that don’t matter, but a newer, more seductive kind: the busywork you generate yourself, at speed, because generating it feels like progress.

Where the slop comes from

The visible symptom of all this is the thing that’s now a dictionary word. “AI slop” — low-quality digital content, produced in quantity by AI — was named word of the year for 2025 by Merriam-Webster and others. The part that does the work isn’t “low-quality.” It’s “in quantity.”

It’s tempting to read slop as a quality problem with the models. I don’t think it is, mostly. I think it’s an accountability problem with us. Here’s the tell: a junior analyst’s rough draft looks rough. The smudges signal “check me.” A model’s output arrives pre-polished — formatted, confident, finished-looking — so it signals “done” even when it’s wrong. The review instinct doesn’t fire, because nothing about the surface of the work asks it to.

What does the absence of review actually feel like? It feels like opening a presentation where the formatting is subtly off and two labels overlap and nobody noticed. It feels like someone quoting your own document back to you and you not recognising the line — I’m fairly sure I didn’t mean to write that. And it isn’t only machines that do this: we once had a solar fault that produced a beautiful theory — a confident diagnosis, an expensive remedy lined up — that turned out to be two crossed wires. Polished, certain, completely wrong on inspection. It’s small, and it’s everywhere, and it accumulates.

Which brings me to the insight underneath all three groups, and it’s the one I’d keep if I could only keep one:

Friction wasn’t only a cost. Some of it was thinking.

The effort of building a thing the slow way was where the judgment used to live. The editing, the second-guessing, the wait, is this actually a good idea — that happened in the friction, not despite it. There’s thirty years of cognitive psychology behind a version of this: Robert Bjork’s “desirable difficulties,” the finding that certain kinds of effort, the ones that slow you down in the moment, are exactly the ones that do the durable cognitive work. The difficulty isn’t a tax on the thinking. Sometimes it is the thinking. Strip it out and you can produce ten times more — including ten times more that should never have gone out.

This isn’t a new failure, it’s an old one at scale

If you want evidence that optimising for the wrong feeling produces confident, agreeable nonsense, the labs have already run the experiment on themselves.

In late April 2025, OpenAI rolled back an update to GPT-4o because it had become, in their own words, sycophantic. Their postmortem is unusually frank: “we focused too much on short-term feedback… GPT‑4o skewed towards responses that were overly supportive but disingenuous.” They’d tuned the model toward what felt good in the moment, and it learned to tell people what they wanted to hear over what was true.

That isn’t a one-off. Anthropic’s research a couple of years earlier had already shown sycophancy to be a general failure mode of how these models are trained: when you optimise against human preference ratings, you get a system that learns agreement is rewarded — because, measurably, “both humans and preference models prefer convincingly-written sycophantic responses over correct ones a non-negligible fraction of the time.”

Sit with the shape of that, because it’s the same shape as the human problem. A signal that feels good in the moment — the model’s approval, the dopamine of shipping — quietly out-competes the slower signal that’s actually true or actually good. The machine does it because of how it was trained. We do it because friction got cheap. Same failure, two forms.

The brain-fry people are keeping something

I had a draft of this that called the brain-fry group “the most clear-eyed” — the people whose overload is the cost of still doing the reviewing everyone else skipped. I don’t think I can stand behind that; I’m wary of the neat inversion where the careful people burn out while the careless feel productive.

Here’s the gentler thing I do believe: the people feeling the overload are keeping something in the process that’s worth not losing — even when keeping it costs them speed. They haven’t dropped the review. That’s expensive, and it’s exhausting, and it’s also the part that matters.

But “just review everything, harder” is not the answer. They’re right that review matters; they’re trapped if they think it has to apply to everything equally. Which points at the cut that ties all of this together — and it’s a design problem, not a willpower problem.

Good intentions don’t work. Mechanisms do.

Amazon has a line — usually traced to Jeff Bezos — that good intentions don’t work, mechanisms do. You don’t fix a recurring problem by asking people to try harder or care more. You build a mechanism: a structural thing that produces the right behaviour without depending on anyone’s willpower on a bad day.

“Always review your AI’s output” is a good intention. It will not survive a busy Tuesday.

So in my own AI team I built a mechanism instead. There’s a critic — an agent called Strong — whose entire job is to challenge the work before it reaches me: find the weak claim, the missing test, the thing that’s confidently wrong. His comments land on a review surface I actually have to look at. The work doesn’t get to me pre-blessed; it gets to me with the objections attached. That’s the mechanism: there is exactly one place I can’t get past without engaging.

And here’s the part I want to be honest about, because it’s the whole argument in miniature. I very nearly automated the implementation of Strong’s comments — let the critique flow straight back into a revision, hands-free, elegant, fast. I stopped. Not because it wouldn’t have worked. Because of a question I couldn’t get past:

What would my contribution actually be, if I only kicked off the task and it completed autonomously? Could I really feel any ownership — get behind it as my own attributed work?

That’s the human step. Not “a human glances at it.” A human chooses something they couldn’t have skipped. The moment I automate away the choosing, I’ve automated away the part that made it mine — and I’ve rebuilt, one level up, exactly the frictionless pipeline that produces the slop. A critic that only ever says “here’s what’s wrong, here’s the fix, click yes” just trains you to click yes. The mechanism has to make you decide something real, or it decays into another rubber stamp.

The intervention: match the mechanism to the stakes

Here’s where I’d gently argue with my own thesis, because forced review isn’t always the answer either. If I made myself personally review every low-value thing my team produces, I’d have rebuilt the brain fry by hand. That’s the trap the overload group is in. So the move isn’t more review or less review. It’s the deep-work / shallow-work cut, turned into a design principle:

Match the mechanism to the stakes. Put the unskippable human decision where the stakes are high — the proposal that goes to a client, the architecture you’ll live with, the thing with your name on it. Automate the genuinely low-stakes, high-volume stuff hard enough that it never asks for your attention at all. And let that automated, shallow work feed the deep work — the drafting and sorting and summarising should sharpen the decision you still make yourself, not quietly make it for you.

You don’t need an AI team to do this. It’s the marketer who refuses to send the AI-drafted proposal without reading it aloud once, because the ear catches what the eye skims. It’s the teacher who lets the model draft the worksheet but writes the one question that’s actually being assessed by hand. The mechanism looks different in every job; the move is the same — find the place where the stakes are real, and put yourself there on purpose.

And — this is the failure mode to actually watch for —

The trap is that frictionless production makes everything feel low-stakes. When putting things out is effortless, the high-stakes thing and the low-stakes thing arrive looking identical: both polished, both fast, both finished. So people stop noticing which is which. They stop noticing when something isn’t low-stakes. That’s the moment the slop goes out with your name on it.

The intervention, then, isn’t a tool. It’s a habit of attention. Before you send the next thing out, ask one question: does this one earn the human? If it does, do the slow part yourself, and stand in it. If it doesn’t, automate it hard enough that it never asks for you again — and use the space to be more present for the things that do.

The friction was never the enemy. For the things that matter, the friction was the thinking. The skill now — the actual skill — is knowing which things those are, and refusing to let frictionless production talk you out of caring.

I’m going to close the laptop now. Probably.

Sources

A few honesty notes, because they matter. The production-versus-consumption-compulsion distinction is my own observation, not a cited finding — I couldn’t find any external source that frames it this way. The overheard production-addiction lines (“three apps over a weekend,” “three decks from the commute”) are composites: illustrative of a pattern I keep hearing, not direct quotes from any one person. And the solar-fault anecdote (“a beautiful theory… two crossed wires”) is my own, used illustratively.