Reviewing AI-generated content

Every organisation that adopts AI tools passes through the same sequence of moods. First, wonder: the drafts are fast, the summaries feel magical. Then, a quiet unease: something in paragraph three wasn’t quite right. Then, one bad afternoon: a document reaches a client containing a statistic that doesn’t exist, a citation to a source that never was, or a confident statement of terms that contradicts what you actually signed — and the cost of the discovery has a currency symbol attached. What separates the organisations that stay in the productive phase from the ones that retreat to banning things is not better models. It’s a review discipline: humans who know how to check machine-drafted work, quickly, in the right order, and who understand that responsibility for the output moved to them the moment they hit send.

Fluency is not truth

The core difficulty of reviewing AI content is that the failure mode is invisible to the reading habits most professionals have built. We skim for tone and structure; we notice clumsy sentences; we assume that fluent means informed. AI drafts break that assumption because they are optimised for fluency — an invented statistic arrives in a perfectly constructed sentence, the invented citation is formatted flawlessly, the wrong number sits inside a paragraph that would survive a copy-edit with full marks. Reading for language finds nothing, because there is nothing wrong with the language. The error is in the facts, and facts require a different kind of reading: verify-first, trust-second, applied selectively where risk concentrates.

This is why ‘just read it carefully’ fails as a review protocol. Carelessness isn’t the problem; the problem is that care, untrained, pays attention to the wrong layer. A reviewer who has been taught where errors concentrate — and who has felt, once, the particular chill of a fabricated source in a document they nearly sent — reads differently forever after. Training that judgement is what this guide is about, and it’s more learnable than it sounds: the failure modes are surprisingly few, and they repeat.

The facts-first pass

Review in layers, in a fixed order, because the layers have very different cost-and-risk profiles and the order prevents wasted effort. Layer one: facts. Every name, number, date, figure, and — above all — every citation or reference. This is where hallucination concentrates, because this is where the model is asked about things that exist in the world rather than things that exist in your prompt. Names come out subtly wrong; percentages drift from the source; the citation is a plausible blend of a real journal, a real author and a paper that was never written. The check is mechanical: anything the draft asserts about the world gets verified against your source material or against reality, and anything unverifiable gets cut. Fifteen seconds per checkable claim on a routine email; five minutes on a document that will be read by a customer.

The professional habit that makes this layer fast is knowing which claims to check. Not every sentence contains one: connective tissue (‘as we discussed’, ‘in today’s market’) is checkable only against the model’s mood. Risk concentrates where the draft is specific — a number, a name, a date, a source, a quotation. Specificity is exactly what makes AI drafts useful, and it is also precisely where they fail. So the review triage is simple: underline every specific claim; verify the ones that carry risk; delete the ones that carry risk and aren’t needed. Most drafts survive the deletion with no loss of value and all of the credibility.

The claims-and-argument pass

Layer two: reasoning. Facts verified, the question becomes whether the draft’s argument holds — and AI writing fails here in a distinctive way: the confident middle. A draft can open with a true premise, close with a reasonable recommendation, and connect them through an inference that no careful human would sign. The giveaway is the smoothing: ‘therefore’, ‘which means that’, ‘clearly’ — the connectives of an argument that has been rendered fluent without ever having been checked. Reviewing this layer means slowing down at every ‘therefore’ and asking the child’s question: is that actually true? Does the discount rate follow from the churn figure, or did the paragraph just wish it so?

A special subtype deserves its own alarm: the invented commitment. AI drafting an email about your services will, with cheerful confidence, promise things — ‘we’ll have the report to you by Friday’, ‘this is covered under your support agreement’, ‘our team can extend that deadline’ — that no one at your company ever authorised, because they are the statistically likely promises in such emails. They read as generosity and cost as liability. The review pass for commitments is short and non-negotiable: every promise, deadline, assurance and scope statement in a machine-assisted draft gets checked against what you have actually agreed to, and softened or struck where it outruns reality. The phrase to hunt is any future tense with a verb of obligation.

The tone-and-terms pass

Layer three: voice and boundaries. The subtle failures live here, and they cost trust rather than money — which makes them sneaky, because nothing ‘goes wrong’ in a way you can trace. The hedge that drifted: your careful ‘may reduce processing time’ has become the model’s ‘will significantly reduce processing time’, because the model’s native register is promotional. The boilerplate that contradicted your actual terms, arrived at by blending your industry’s generic language. The apology that accepted fault you haven’t established. The American spelling in your British document. Each is small; together they add up to the sense that a company’s writing is nobody’s — which clients notice before they can name it. The check is a house-style pass with teeth: does this sound like us, does it promise only what our terms support, is every hedge where we put it?

This layer is also where disclosure decisions live. Many companies now note AI assistance in their terms or on request; some client relationships warrant explicit mention. Whatever your line is, decide it once, write it down, and apply it — the unacceptable positions are the two extremes nobody chose on purpose: full disclosure theatre on every email, or silent adoption nobody authorised. A one-line sentence in the AI policy settles it, and the review protocol is what makes the chosen line true in practice.

Review depth by stakes

A protocol applied identically to everything collapses under its own weight within a fortnight, so the last layer of the discipline is proportionality. Full protocol (all three passes, second reviewer where available) for anything that leaves the company with your name on it and carries commitment: client deliverables, proposals, contracts support, public content, anything financial. Standard protocol (facts pass on specifics, commitments pass, house style) for internal documents that inform decisions — the summary a colleague will act on is a document worth five minutes. Light pass for the low-stakes bulk: internal notes, first-draft material that a named human will substantially rewrite anyway, exploratory work where the output is an input. The skill being taught to your team is not ‘check everything maximally’ — it’s knowing, per document, which protocol applies, and never letting an unreviewed draft cross the company boundary.

Ownership is the closing principle, and it’s the one that makes the rest enforceable. Every AI-assisted document has a human name on it, and that name means something: they read it, they checked it, they answer for it. Not ‘the AI wrote it’ — that sentence is an abdication wearing a technicality. The practical form this takes in well-run teams is simply that the sender is the reviewer: whatever the machine contributed, the person on the from-line of the email or the byline of the document performed the passes above. Once that norm is in place, review depth stops depending on individual diligence and becomes culture — the quiet, shared understanding that in this company, fluent is not the same as true.

Why this is trainable, and where to train it

Everything above compresses into a teachable checklist — facts first, then reasoning, then voice and commitments, at depth proportional to stakes, with a human name on every send. It takes about twenty minutes to learn and a week of practice to become reflex, which is a better exchange rate than any alternative on offer: the one bad afternoon, or a permanent low-grade distrust that slows every draft to the speed of suspicion. Teams that train the protocol describe the same outcome: they run faster with AI than before the training, because the trust boundary is explicit — speed on the safe side of the review, care exactly where the risk lives.

That’s Reviewing AI Output, the module this guide is the reading version of — deliberately seeded practice documents, the layered checklist as a desk card, and a company review policy you write in the final lesson. It sits at the end of every sensible rollout sequence, after the production modules, because it is the skill that makes all the other skills safe to deploy. Read the guide, teach the checklist, put names on sends — and the machine’s fluency becomes what it should have been all along: an amplifier with a competent adult holding the volume control.