The Efficiency Paradox: Where the Work Went

AI does not remove the work of writing. It moves it. The hour you save by not drafting a first version does not vanish — it reappears downstream, in two jobs the blank page used to fold into one: verifying what the model asserted, and re-voicing what it wrote so the text sounds like a person reasoned it out rather than a machine averaged it. Those two jobs are harder to see than drafting, they are easy to leave out of a time estimate, and when you leave them out they take back the time you thought you had banked — frequently with interest. The writers and teams that come out ahead are not the ones who generate fastest. They are the ones who planned for where the work went.

This is the whole argument of this site compressed into one claim, so it is worth stating plainly before the evidence: the efficiency of AI writing is real at the point of generation and conditional everywhere after it. Skip the condition and you do not save time. You relocate it, lose track of it, and pay it late.

The paradox has a measurement

The cleanest proof that the work moves rather than disappears comes from a setting where you can actually time it. In early 2025, METR ran a randomized controlled trial with 16 experienced open-source developers across 246 real tasks on their own large repositories, randomly allowing or forbidding AI tools (mostly Cursor Pro with Claude 3.5/3.7 Sonnet). The developers expected AI to speed them up by 24%. It slowed them down by 19% — and even after finishing, they still believed it had made them about 20% faster (METR, 2025). The gap between the felt 20% faster and the measured 19% slower is the efficiency paradox in a single number. The generation step genuinely felt quick. The reviewing, correcting, and reconciling that followed it did not register as “the task,” so it did not register at all — even though it was where the time went. Tellingly, the developers accepted fewer than 44% of the AI’s suggestions, which means more than half of what the tool produced became something to read, judge, and discard: downstream work, created by the upstream shortcut.

That pattern is not confined to code. A 2026 survey of desk workers by Foxit found executives estimating about 4.6 hours saved per week against roughly 4 hours and 20 minutes spent verifying AI output, and ordinary users saving about 3.6 hours against nearly 3 hours and 50 minutes reviewing it — a net gain one write-up measured at sixteen minutes a week (Foxit, 2026). The Glean Work AI Institute’s 2026 study of 6,000 digital workers across the US, UK, and Australia put a name to the downstream job — “botsitting” — and a size to it: workers saved around 11 hours a week with AI and spent more than six of them checking, correcting, and re-prompting, with 37% of their weekly AI time going to botsitting against 36% spent producing anything (Work AI Institute, 2026). A separate 2026 Zapier survey of more than 1,100 enterprise AI users found the average employee spending 4.5 hours a week revising, correcting, and redoing AI output, even as 92% reported feeling more productive (Zapier, 2026). The feeling of productivity and the location of the labor have come apart. The work is downstream, and most people’s mental accounting still books it upstream, where it is cheap.

Why drafting is the cheap part

Verification is expensive for a structural reason, not a temporary one: a model optimizes for plausibility, and plausibility is exactly the property that makes a false claim expensive to catch. The costliest version is the fabricated citation, because it wears the uniform of scholarship. A 2026 review of the problem reports citation-hallucination rates running from 18% in GPT-4 up to over 70% in other frontier models, with as much as 88% in legal contexts — and underneath those headlines, controlled measurements put ChatGPT-4o at a 20.0% hallucination rate and Gemini Advanced at 76.7%, while the CiteME benchmark found language models citing scientific claims accurately only 4.2–18.5% of the time against 69.7% for human annotators (Misra & Udandarao, 2026). Notice what those numbers describe: a draft that looks finished and is not. The fluent surface is produced in seconds; establishing which of its references actually exist is the slow part, and it is slow precisely because the fabrications are built to survive a skim.

You cannot close that gap by asking the model to grade itself, which is the shortcut most people reach for first. It helps less than it looks. When Safran and Çalı had ChatGPT generate forty scientific references and then asked it to verify its own work, the share of fully accurate citations rose from 7.5% to 77.5% (Safran & Çalı, 2025, p. 698) — a real gain, and still a wall. After a careful self-check, roughly one reference in five was wrong (Safran & Çalı, 2025, p. 699), and the survivors are the confident fabrications polished enough to pass the model’s own inspection. The self-review is a cheap first pass, not a finish line; we unpack exactly where it holds and where it quietly fails in the self-critique pass. The only step that actually resolves a citation is external — checking each entry against a real bibliographic index. That is a distinct, engineered stage, not a favor the author model does for free: the fact-verification systems that work at all separate retrieval from judgment and check claims against an outside source, and even the strongest hybrid pipelines land around an F1 of 0.93 rather than 1.0 (Kolli et al., 2025). Verification is its own discipline with its own ceiling. Pretending the drafting tool performs it is how the downstream hour goes unbudgeted.

The honest shape of the task is old, and journalism has known it longer than AI has existed. Verification — “checking and verifying statements and sources” — is a basic routine of the newsroom, and Laura Postma’s study of data journalism is blunt that integrating AI “has its costs” and that the work it touches is “time-intensive” (Postma, 2024, p. 6). Her argument is that verification cannot be shunted to a specialist corner and treated as optional; it has to be built into the ordinary production flow, with the people doing the work sitting at the same table rather than bolted on at the end (Postma, 2024, p. 7). That is the same lesson the botsitting surveys are relearning the expensive way. The verification does not go away because a machine drafted the text. It gets larger, because now the text can be confidently wrong at scale. Before you trust a reference list a model handed you, run it through a checker that confirms each entry resolves to a real work — cytado.com’s source checker does this for free (cytado is owned by this site’s operator, so read that as a disclosed interest, not a neutral tip), and it catches the invented DOI your eye slides past exactly because the surface reads clean.

The second downstream job: getting the voice back

Verification is only half of where the work went. The other half is re-voicing — the editing you do to make an AI draft read as though a specific person thought it, rather than as the fluent, hedged, structurally identical prose every model converges on. This is the labor behind the 2026 term “workslop”: output that looks like work, passes a glance, and offloads the actual thinking onto whoever receives it. The Glean study found 41% of workers admitting they had forwarded AI-generated work they could not explain if pressed (Work AI Institute, 2026). Every one of those documents is a re-voicing job deferred, not avoided — pushed onto a colleague, an editor, or a reader who now has to reconstruct the meaning the draft only gestured at.

Re-voicing is harder to measure than verification because it has no DOI to resolve against, but it follows the same economics. The time you save by generating a competent-sounding paragraph is borrowed against the time it takes to turn a competent-sounding paragraph into one that actually says your thing in your register. And the two jobs interact: the most dangerous edit is the one that smooths the prose while leaving a fabricated claim untouched, so the text now reads more trustworthy and is no more true. The discipline that keeps re-voicing honest is the same boundary we draw in correct, do not rewrite — stay on the surface, keep the substance and the responsibility yours, and never let a polish pass launder an unverified fact into a confident one. Editing an AI draft well is the core skill of this whole editing pillar, and it starts with naming the job out loud instead of hoping the draft is already done.

Planning for where the work went

The failure is almost never the tool. It is the estimate. The writer who budgets an hour for a piece because the model “writes it in five minutes” has silently assumed the five-minute draft is the deliverable. It is not. The deliverable is a verified, re-voiced text, and the fifty-five minutes the model appeared to save are the fifty-five minutes you now owe to checking and rewriting — the work did not leave, it changed address. A few moves keep the account honest:

• Estimate the downstream, not the draft. Budget verification and re-voicing as their own line items. If a piece makes factual claims, assume checking them costs more than generating them, because the measured numbers say it does.
• Verify before you polish, never after. Resolve every citation and claim against an external source first. Polishing an unverified draft only makes its errors more convincing — you are investing editing time into protecting a fabrication.
• Re-voice deliberately or ship slop. Decide whose voice the text is in and edit toward it. A draft left in the model’s default register is a re-voicing job you have handed to your reader without telling them.
• Count the discards. More than half of what a model offers is often something you reject (METR, 2025). Reading and rejecting is work. If your plan assumes you keep what the model writes, your plan is wrong.

None of this is an argument against using AI to write. It is an argument for booking the cost where it actually lands. Generation is genuinely fast, and that speed is worth having — but it is the first and cheapest step of a longer process, not a replacement for it. The hour the model saves you is real. So is the hour it moves downstream. Whether you come out ahead depends entirely on whether you saw the second hour coming and planned for it, or whether you spent the first one celebrating and met the second as a surprise. The work did not disappear. It went somewhere specific. Go meet it there on purpose.