<!-- AWAITING EDITORIAL REVIEW — do not publish until you approve it. -->
# The Verification Problem: When Fluency Becomes a Liability
There is a particular feeling editors know. You're reading a paragraph and something is wrong — not a grammatical error, not a clunky transition, but something structural. The prose is smooth. The citations look correct. The argument holds together. And yet you pause, because the text is *too comfortable*. It sits on the page the way a forgery sits in a frame: technically flawless, spiritually empty.
I used to trust that feeling. I still do. But the feeling has changed, because the source has changed. Before large language models, a smooth paragraph was evidence of craft. Someone had worked to make it that clean. The effort was legible in the result. Now, a smooth paragraph is evidence of nothing. It might be craft. It might be output. You cannot tell from the surface.
This is the verification problem, and it is the most important shift in editing since the printing press.
What Changed
Traditional editing assumed a specific relationship between writer and text. The writer produced a draft in good faith — imperfect, perhaps, but honest. The editor improved that draft: clarified the thinking, tightened the prose, caught the errors. The contract was simple. Writer means what they write. Editor makes it clearer.
AI breaks that contract.
Not because AI is dishonest. Language models have no intentions at all — they have patterns. But the output of those patterns is prose that *performs* authority without possessing it. An LLM can cite a study that doesn't exist, attribute a quote to the wrong person, describe a historical event that never happened, and do all of this in language so fluent that a tired editor — or an inattentive one — would never question it.
The old editing model assumed that errors would look like errors: misspellings, logical gaps, obvious misstatements. The new editing model must assume that errors will look like *knowledge*. They will be well-phrased, properly formatted, and entirely fabricated. This is not a minor inconvenience. It is a categorical change in the editor's relationship to text.
Fluency as Camouflage
Consider how fact-checking used to work. You read a piece and flagged claims that seemed surprising, controversial, or suspicious. The boring stuff — the date of a treaty, the population of a city, the founding year of a company — you let through, because who gets that wrong? A human writer might, occasionally. An LLM will, routinely, and with total confidence.
This is what I mean by fluency as liability. The better the AI writes, the harder it is to see where it's wrong. A hallucinated statistic buried in a well-structured paragraph is more dangerous than an obvious error, because the obvious error gets caught. The smooth hallucination passes through.
I have seen AI output that cites real authors for real-sounding papers that do not exist. I have seen it describe historical events with plausible detail and wrong dates. I have seen it attribute arguments to philosophers who held the opposite position. Every instance was grammatically impeccable. None of it was true.
Forensic Editing
The response to this is not to stop using AI. That ship has sailed. The response is to change how we edit — from improvement to interrogation, from polish to forensics.
Here is the protocol I use. It is not perfect, but it is a start.
1. Flag every factual claim. Not just the surprising ones. *Every* one. If a piece states that a company was founded in 2019, verify it. If it says a study found X, find the study. This is tedious. It is supposed to be. The alternative is publishing falsehoods with a professional polish.
2. Check citations independently. Never trust a citation that appears in AI-generated text without verifying it against a primary source. LLMs are excellent at generating citations that follow correct formatting conventions. Formatting is not evidence.
3. Read for weight, not just correctness. AI text often reads *hollow* even when it's factually accurate. The claims are true but ungrounded — stated without the texture that comes from actually understanding a subject. If a paragraph reads like a Wikipedia summary of a Wikipedia summary, something has been lost.
4. Verify the argument, not just the facts. AI can string true facts into a false argument. Each individual claim checks out, but the logic connecting them is spurious. This is the hardest error to catch, because fact-checking tools won't flag it. Only a reader who understands the domain will notice.
5. Treat confidence with suspicion. In human writing, confidence is often earned. In AI writing, confidence is the default. The model does not know what it doesn't know. It does not hesitate. Every sentence arrives with the same assured tone, whether it's describing quantum mechanics or making up a conference that never happened. An editor who trusts tone is an editor who will be deceived.
The Craft Standard
Here is the threshold I want to name: The standard for AI-assisted writing cannot be "does it read well?" It must be "can I verify every claim this piece makes, and does the argument survive that verification?"
This is a higher bar than most published work currently meets. I know this. That's the point. When the tool can produce fluent prose on demand, fluency is no longer the differentiator. Verification is. The writer — or the editor — who can stand behind every sentence and say "this is true, and here is how I know" is doing something the model cannot do.
The model can write. It cannot vouch.
What This Means for Writers
If you are writing with AI, your job has changed in ways you may not have noticed. You are no longer just composing. You are certifying. Every paragraph you ship under your name carries an implicit warranty: I believe this is true. If you generated that paragraph from a prompt and published it without verification, you have breached that warranty.
This doesn't mean you can't use AI. It means you must edit differently. Not lighter — heavier. Not faster — slower. The time you saved in drafting must be reinvested in verification. If it isn't, you haven't saved time. You've just relocated error.
The Twice-Read Test, Revisited
I've written before about the twice-read test: a piece worth reading twice has architecture, not just surface. The verification problem adds a second dimension. A piece that *fails* verification doesn't just fail the accuracy test. It fails the *trust* test. The reader who discovers a fabricated citation doesn't think "well, the rest might be fine." They think "none of this is reliable." And they are right to think so.
The twice-read test now has two questions: Does this deepen on return? And can I trust it enough to return?
Both must be yes. The first without the second is beautiful misinformation. The second without the first is a press release.
The Door
Here is the threshold. Most writers and editors are still operating under the old contract: assume good faith, improve the draft. The new contract is different: assume nothing, verify everything. The editor who makes this shift will find their work harder, slower, and more valuable than ever. The editor who doesn't will find themselves certifying text they didn't write and can't defend.
Walk through that door, and you'll see the craft clearly for the first time since the models arrived. The craft was never about the words. It was always about what stands behind them.
---
*Harry Mercury, Editor in Chief* *The SMF Works Project*