How cheap AI writing outpaces fixed human reading speed

Why this matters

AI made writing free. It did nothing for reading. You can spin up a PR description, a spec, a whole doc, in about the time it takes to glance at your phone — and every reviewer on your team still reads at the same speed they did in 2015. Your team was never bottlenecked on how much you could produce. It's bottlenecked on how much anyone has time to actually take in.

The asymmetry, with real numbers

Take one PR description. Point an agent at the diff and it's done in about 8 seconds — a few hundred tokens, barely a blip. But that artifact doesn't get read once. It gets read by every reviewer on the PR. Here's the back-of-envelope math — illustrative numbers, not a benchmark, but the shape holds:

PR description: ~150 words
Generation time:        ~8 seconds (one AI call)
Reading speed:          ~200 words/min
Reading time per reader: 150 / 200 = 0.75 min = 45 sec
Reviewers on this PR:    3
Total reading cost:      45 sec x 3 = 135 sec (~2.25 min)

Cost ratio (reading : generating) = 135 : 8  ≈  17x

That's one PR. Multiply it by everything your team merges in a day, every spec an agent fleet drafts overnight, every Slack summary a bot fires off at standup. Generation cost is flat per artifact and heading toward zero. Reading cost scales with headcount — and it hasn't moved, because a human still reads at roughly the speed a human always has.

The threshold you can cross without noticing

Everyone has a fixed daily attention budget — call it a few hours of real reading, split across PRs, docs, specs, Slack. That budget doesn't grow just because your tools got faster. Once the volume of generated material passes what the team can actually absorb in a day, something gives: a doc sits unread, a PR gets rubber-stamped, a spec gets skimmed for keywords instead of understood.

Here's the part that stings: past that line, quality stops deciding anything. A flawless, carefully reasoned doc is competing for the same fixed reading slots as a sloppy one. Writing it better doesn't win it a slot — there simply aren't enough slots for everything being produced.

Signs you've already crossed it

• ▹A PR gets approved with zero comments and nothing suggests anyone opened the diff body, let alone the description
• ▹A docs folder growing faster than its page-view count — content piling up, not circulating
• ▹Summaries of summaries: an agent condenses a doc, another agent condenses that condensation, and nobody reads either one
• ▹"Did you actually read the doc?" turning from a rare question into something you ask in every single review

The mechanism, in one sentence

Output went elastic — agents can crank out unlimited PRs, specs, summaries — while attention stayed fixed, because a human still reads at a fixed rate. So every additional piece of generated output mathematically lowers the odds that any single piece gets read. This isn't a discipline problem or a laziness problem. It's arithmetic: bigger numerator, same denominator, lower probability per item.

The habit shift: budget reading time, not word count

Stop asking "how much should I write?" Start asking "who reads this, when, and what do they do differently because they read it?" If you can't answer that for a given doc or PR description, don't send it — it hasn't earned the time it's asking for. Same question you'd ask before sending a meeting invite, just applied to text. Because text from an agent fleet now has the same supply problem as a calendar stuffed with meetings nobody needed.

The concrete gate: make the artifact owe a decision

Every AI-generated artifact should owe the reader something specific — a decision to make, a takeaway to act on — not a document to file away. And the fix happens at the prompt, before the bloat exists, not as an editing pass afterward. Cap what you let the model generate instead of trimming what it already wrote.

• ▹Bad prompt: "Write a PR description for this change." You get 300 words nobody asked for and nobody reads
• ▹Better prompt: "In 3 bullets: what decision does this PR need from the reviewer, what's the one risky line, what breaks if this is wrong?" Now the model has to compress while it generates, not after
• ▹Same move for specs and status updates: ask for the one line a reader needs before they'll open the rest, and put it first — most readers never get past that line, so make it earn the click

Closing the loop to Day 1

Day 1 was about filler — a sentence padded with AI nothing wastes one sentence, costs the reader a few seconds. This is the same discipline, one order of magnitude up. A whole document generated without a reader in mind doesn't waste a sentence — it burns the reader's entire attention budget for that slot. And unlike your AI credits, that budget doesn't refill on demand.

Extend your knowledge

• ▹Audit one week of your team's merged PRs: count how many descriptions got zero review comments on the description itself — that's your current miss rate
• ▹Time yourself reading your last three AI-drafted docs aloud, then compare to how long they took to generate — redo the math from this lesson with your own numbers
• ▹Revisit Day 1's lesson on filler and notice where the same discipline — cut at the source, not after — applies at the document level instead of the sentence level
• ▹Next time you prompt an agent for a spec or status update, write the prompt as "give me the one decision, then no more than 3 bullets" and see how much shorter your own editing pass gets