Vetting AI content: the checks before anything ships

By Matija Konjić

Every company now has access to infinite drafts. The bottleneck moved overnight from producing content to judging it, and most teams have no judging process at all: generated text flows from a prompt to a publish button with nothing in between except optimism. The results are visible across the web, and increasingly visible in the policies search engines write about it.

This guide is the process for the judging part: how to vet AI-assisted content before it ships, what the checks are, who should run them, and when the honest verdict is that a draft costs more to fix than to rewrite. It is the same exam we run inside our editorial review service, on AI output, freelance work and guest submissions alike, because the origin of a draft changes nothing about the standard it has to meet.

Why vetting became the job

Generation collapsed the cost of producing plausible text, and plausible is exactly the problem. A weak human draft usually looks weak: thin paragraphs, obvious gaps, visible effort missing. A weak generated draft looks finished. It has confident topic sentences, tidy transitions and the cadence of expertise, which means the failures hide where skimming never finds them, inside the claims.

The economics moved with the bottleneck. When drafting cost hundreds of dollars a piece, review was a garnish; when drafting costs cents, review is the product. Buyers who understand that shift stop asking providers what they charge per word and start asking what stands between generation and publication, which is the question this whole guide answers.

Google’s position sharpened the stakes without banning the tool. Its guidance on generative content is consistent: how content is produced matters less than whether it is helpful, original and accurate, and its spam policies now name scaled content abuse, mass production without value, as a target regardless of how the text was made. The policy reads as permission and warning in one sentence: use the tools, own the output.

The liability is asymmetric

One fabricated statistic in a published piece costs more than a hundred clean drafts saved in fees. It burns reader trust, invites correction in public, and hands every skeptical prospect the reason they were looking for. Vetting exists because the downside of shipping wrong is bigger than the upside of shipping fast, and the gap widens with every reader who knows what generated filler looks like.

What AI drafts reliably get wrong

After enough reviews the failure patterns stop surprising you. Five account for most of the red ink:

• Confident fabrication. Statistics, studies and quotes that do not exist, cited with perfect formatting and full conviction.
• Stale ground truth. Prices, features, laws and interfaces described as they were in the training data rather than as they are today.
• Consensus blur. A smooth average of the existing top ten, with every original observation sanded off.
• Misattributed authority. Real sources attached to claims they never made, which is worse than no source at all.
• Voice drift. House style at the start of a long draft, generic marketing voice by the middle.

None of these are exotic malfunctions; they are the natural texture of generated text. The exam below exists because every one of them slips past a casual read.

A worked example of the first pattern, typical of what review catches weekly: a draft about invoicing law cites a government statistic, names the agency, formats the reference cleanly, and the statistic does not exist anywhere in the agency’s publications. Nothing about the sentence looks wrong. The only defense is the boring one: somebody opened the source and read it, which is precisely the step production pressure wants to skip.

The exam: five checks, three verdicts

The vetting sequence runs cheapest checks first, and every draft leaves with one of three verdicts.

Check one is claims: every number, date, name and assertion of fact gets highlighted and verified against a primary source, and anything unverifiable gets cut or softened. Check two is sources: links must exist, must say what the draft claims they say, and must be worth citing. Check three is originality: does the piece say anything the current top ten does not, and would an expert reader learn something. Check four is voice: does it read like the brand on its best day, and does it hold to the style guide from first paragraph to last. Check five is fit: intent, internal links, and whether the piece does the commercial job the brief assigned.

The verdicts, honestly applied

Keep means the draft survives with light touches, which is rarer than teams expect. Fix means the bones are right and the flesh needs work: real edits, real corrections, worth the hours. Kill means the fixing would cost more than a rewrite, and the honest response is a better brief rather than a longer edit. Teams that never issue kill verdicts are not vetting; they are formatting.

The verdict distribution is also a management signal. A rising keep rate means the briefs and prompts are improving; a rising kill rate means the generation step is being asked to do work it cannot, usually original analysis or anything requiring current ground truth. Read quarterly, the log tells you where to spend: better briefs, better sources fed into the drafts, or a human writer for the pieces generation keeps failing.

How deep to vet what

Not everything deserves the full exam, and pretending otherwise guarantees the exam gets skipped. Depth follows risk.

Anything with claims, prices, health, money or legal exposure gets every check, every time, the same logic that makes regulated niches the strictest corners of search. Public brand copy gets the voice and originality checks at full strength. Data-heavy comparisons get the source check hardest. Internal material gets a light pass, because the cost of an error there is a correction rather than a reputation.

The risk register habit

Teams in regulated spaces should keep the quadrant explicit: a one-line register naming which content types get which depth, agreed with whoever owns compliance. It ends the per-piece negotiation, survives staff changes, and turns an awkward conversation about AI policy into a working document the lawyers have already blessed.

Who should vet

The vetting seat needs two qualifications: enough subject knowledge to smell a wrong claim, and enough editorial skill to fix voice without flattening it. That is an editor with domain grounding, and specifically a person whose name is attached to the verdicts. Diffuse responsibility produces diffuse quality; a named editor with a written standard produces a bar.

What the seat does not need is the writer who produced the draft. Self-review catches typos and misses everything that matters, because the reviewer already believes the claims. Even a small team can separate the roles: whoever prompts does not approve, and the approval trail says who passed what.

Outsourcing the seat works when the outside editor holds both qualifications and a written standard you have read. It fails when review is bought as a checkbox from whoever also sold the drafts, because the seller grading its own homework returns the grade you paid for. Whoever vets, the test stays the same: show me a kill verdict from the last month, and show me what it caught.

Where the tools help

Detection and checking tools earn a place without earning trust. Plagiarism checkers catch copying, link checkers catch dead sources, and AI detectors are unreliable enough on modern output that a verdict should never rest on one. The checks that matter, is the claim true, is the source real, is the point original, remain stubbornly human, which is exactly why vetting has market value.

Fitting vetting into production

Vetting works as a stage, never as a rescue. The workflow that holds: brief, draft, vet, revise, approve, publish, with the vet gate owning the calendar slot the way editing always did in publishing. Squeezing review into the hour before deadline recreates the exact failure the process exists to prevent.

Budget the stage honestly. A full exam on a standard article is an hour or two of skilled time; on a data-heavy piece, more. Priced against a single public correction, or one burned prospect who spotted an invented statistic, the hours are the cheapest insurance in the content budget. On a program shipping eight pieces a month, the whole editorial layer costs roughly one thin article’s fee, and it is the only line in the budget that touches every piece. The arithmetic is the same one that makes vetting a guest post site the highest-paid minutes in link building.

The paper trail

Keep the verdicts in writing: what was checked, what was corrected, who approved. The log turns quality from a mood into a record, makes new editors trainable in a week, and gives whoever owns the brand an answer better than we are careful when a stakeholder asks how AI is being used.

When AI-assisted content is simply fine

The honest ending is that generated drafting, vetted properly, is a legitimate way to produce good content. The tools are excellent at structure, at first-pass explanation, at turning an expert’s bullet points into readable prose. Paired with real sources, a human expert’s corrections and a genuine editorial pass, the result can meet any bar on this site.

What has no defense is the unvetted version, and the market is pricing that in already: readers skip it, search systems target it at scale, and AI assistants decline to cite it. The dividing line in 2026 is not how the draft was made. It is whether anyone with judgment stood between the draft and the publish button, and that judgment is exactly the layer we sell.

Run the numbers on your own last quarter as a closing exercise: pieces shipped, pieces that would have survived the five checks, and what the gap cost in corrections, rewrites and rankings that never came. Most teams find the exam pays for itself before the second month, which is why the process outlives every tool cycle that was supposed to replace it.