Failure modes
Every review workflow should be designed backwards from the ways the thing being reviewed goes wrong. AI-assisted output has a characteristic set, and they differ from the mistakes a tired human writer makes: a person misremembers; a model composes. The output is fluent, structured and confident whether or not it is true, which is why proofreading, the review most teams already have, catches almost none of it. The table lists the failure modes a marketing reviewer meets most often and the check that catches each. Why a model composes confident falsehoods in the first place is explained in how language models generate text.
| Failure mode | What it looks like | Most likely in | Caught by |
|---|---|---|---|
| Fabricated facts and figures | A statistic, date, feature or price that sounds right and is not | Research summaries, product copy, reports | Source-grounded prompts; every figure traced to a source |
| Invented or misattributed sources | A study, author or URL that does not exist, or a real source that does not say that | Long-form content, thought leadership | Open every citation; quote-to-source check |
| Unsupported claims | "Fastest", "most trusted", "clinically proven", "customers save" without evidence | Ad copy, landing pages, email | Claim check against a substantiation file |
| Outdated information | Features, pricing or rules that were true when the model was trained | Anything about products, platforms or law | Source packs dated; reviewer checks currency |
| Brief drift | Correct, well written and not what was asked for | Long outputs, multi-step tasks | Checklist against the brief before anything else |
| Inherited errors | Mistakes in the input (an old spec, a wrong field) reproduced with confidence | Summaries, data-to-text, personalisation | Input quality check; source hierarchy |
| Over-confident tone | Hedged or uncertain matters stated flatly | Analysis, comparisons, advice content | Reviewer marks certainty against the evidence |
| Inconsistency | Different answers to the same question across outputs | Batch generation, agents, assistants | Spot checks across a sample, not one item |
| Policy and legal missteps | Health, finance, safety or environmental claims; comparative advertising; endorsements | Regulated categories, testimonials | Specialist review tier |
| Omission | The caveat, exclusion or step that matters is missing | Instructions, terms, offers | Reviewer works from a "must include" list |
Two patterns run through the table. First, the failures are mostly invisible to a reader who does not already know the truth, which is why the reviewer needs the sources, not just the output. Second, most of them can be reduced before review by changing what the model is given, which is where grounding comes in. When fabricated facts keep appearing despite the checks, the troubleshooting page for AI content with invented facts logs the recent inventions and traces the pattern to its cause.
Grounding in sources
Grounding means giving the model the material it should draw on and instructing it to draw only on that material. In practice this is a source pack: the product specification, the pricing page, the brand claims file, the previous approved version, the research the piece is meant to summarise. The instruction is explicit — use these sources, cite which source each statement comes from, and say "not in sources" rather than filling a gap. Retrieval tooling that pulls relevant passages from a document store automates the pack for teams producing at volume; for most marketing teams, a well-maintained folder and a prompt template do the same job.
A source hierarchy settles conflicts. Your own structured data (the product catalogue, the price list, the CRM) outranks your own prose (old pages, decks), which outranks public primary sources (a regulator's page, a platform's help centre, a published study), which outranks anything the model "knows". When a reviewer finds a statement with no source, the rule is simple: it is not in the output until it has one.
Grounding reduces fabrication; it does not remove it. Models paraphrase sources loosely, combine two sources into a claim neither makes, and cite a source for a sentence that came from elsewhere. It also does nothing about a source that is itself wrong, which is why the pack needs an owner and a date. For analysis tasks — summarising a dataset, explaining a trend — grounding means the model sees the query and the result, and the reviewer checks the figures against the data, as the AI for analysis and reporting guide describes. For written content, the AI for content guide covers what a source pack for a page or an email should contain.
Fact and claim checks
A fact check and a claim check are different jobs and are often confused. A fact is something that can be verified against a source: a product dimension, a release date, a quotation, a figure from a report. A claim is an assertion that needs evidence to be fair: that the product is faster than a rival, that most customers see a result, that something is the best in its class. Opinions and framing are a third category and are the editor's business, not the checker's.
The fact check works line by line. Every number, every proper noun, every date, every quotation and every statement of what a product does is marked, traced to the source pack and ticked or struck. The reviewer does not accept "the model cited it" as verification; the citation is a pointer to be followed, and the source is read at the cited place. For a single AI answer used in research or planning rather than published, the step-by-step fact-checking routine applies the same discipline claim by claim. Where a statement cannot be traced it is removed or rewritten as a question for the owner. The discipline is slow the first time and fast once the source pack is good, because most statements trace quickly.
The claim check works from a list. Comparatives ("faster", "cheaper", "more accurate"), superlatives ("best", "only", "leading"), outcome claims ("saves time", "customers see results"), category claims ("clinically", "certified", "sustainable") and implied endorsements ("experts recommend") are each flagged and either backed by something in a substantiation file or softened. Advertising regulators in most markets expect claims to be substantiated before they are made, and a model has no idea whether yours are. Search quality guidance pulls the same way: pages that demonstrate experience, expertise, authority and trustworthiness (E-E-A-TExperience, Expertise, Authoritativeness, Trust.) are pages whose claims are specific and supported, not fluent and generic.
When AI-written variants are tested against existing copy in an A/B TestControlled comparison of two variants., both arms go through the same checks before launch. A variant that wins because it makes a claim the control could not make has not won; it has been caught late. The A/B testing guide covers the test design; the point here is that review sits before the test, not after the result.
Review tiers by risk
Reviewing everything at full depth is how review workflows die: the queue grows, the checks get skimmed, and within a quarter the tier that should have caught the regulated claim is rubber-stamping. Match the depth of review to the consequence of error, using the risk tiers the AI governance hub sets out. When AI drafts now reach the editors faster than they can clear them and deadlines no longer move, the diagnosis for AI tools that are not saving time shows how uniform full-depth review becomes the step that sets the team's pace.
| Tier | Output | Review |
|---|---|---|
| Internal and reversible | Summaries, outlines, research notes, internal drafts | The person using it reads it; no record |
| Published with an editor | Page copy, email, ad text, social posts, decks for decisions | Automated checks; brief check; fact check against the pack; claim check; editor sign-off recorded |
| Customer-facing and automated | Chat replies, agent drafts, send-time copy, recommendations | Guardrails in the system; a sampled review on a schedule; an incident route; owner reviews logs |
| Regulated or sensitive | Health, finance, safety, environmental or comparative claims; testimonials; synthetic likenesses | Full published-tier review plus specialist sign-off (legal, compliance, subject expert); disclosure decision recorded |
The tier is assigned at the brief, not at the end, because it determines the source pack and the reviewer. A piece that changes tier mid-way — a blog post that acquires a product comparison, an internal summary someone wants to publish — is re-briefed. Automated checks belong on every tier above internal and cost almost nothing to run: banned terms, claim-pattern matching, unfilled placeholders, links that resolve, a required-elements list. They do not replace a reviewer; they free the reviewer to read for the failures only a person can see.
For automated tiers, the sampling plan is the review. Decide the sample size, the cadence and who reads it; log what the sample found by failure mode; and set a threshold at which the use pauses for a full review. The AI for marketing automation guide covers the guardrails that sit inside the system; this guide covers the human loop around it. The tutorial on writing a sampling plan for an automated AI use writes the plan for one use: the reader, the sample, the log and the pause rule.
Roles and sign-off
A review workflow needs four roles, which in a small team may be three people and in a large one may be a department. The operator prompts the model and assembles the source pack; they are responsible for the brief and the inputs. The reviewer performs the fact and claim checks and the brief check; for the published tier this is usually an editor. The approver signs off that the review happened and the output may go out; for the regulated tier this is a specialist. The owner of the use is accountable over time: they maintain the source pack, read the logs and decide when the tier or the tooling changes.
Two rules keep the roles honest. The operator does not review their own output above the internal tier, for the same reason a writer does not edit their own work: they read what they meant, not what is there. And sign-off is recorded — who, when, which version — because a review without a record is indistinguishable from no review when a problem surfaces months later. Where the output needs disclosure (a synthetic image, a conversational agent), the approver records that decision too; the disclosure and data privacy guide sets out when it applies. The content governance guide describes the editorial roles these map onto for content teams. For translated output the reviewer must read the target language, and the tutorial on reviewing localised copy in a language you do not speak sets out the checks that remain for an approver who cannot.
A software company reviews an AI-drafted comparison page
A project-management software company drafts a "vs" page against a competitor using an assistant grounded in its own feature list, pricing page and the competitor's public documentation. The page is assigned to the regulated tier because it makes comparative claims. The fact check traces 31 statements: 26 pass, 3 are rewritten (the model stated the competitor "has no API" when its documentation describes one in beta), and 2 are removed because no source supports them. The claim check flags 7 comparatives; 4 are kept with the substantiation file updated, 2 are softened to "in our testing" with the test method linked, and 1 ("the most secure option") is cut. Legal signs off after 45 minutes against a checklist. The review log records 5 defects per 31 statements, all in the "fabricated fact" and "unsupported claim" rows, and the operator adds the competitor's changelog to the source pack so the next draft starts from current information.
Logging and learning
The log is where a review workflow stops being a cost and starts being an asset. Record, per reviewed item: the tier, the model and prompt version, the source pack version, the reviewer, the number of defects found by failure mode, the decision and the time taken. From that log, a few measures tell you how the workflow is doing. Defects per reviewed item by tier is the main KPIA metric tied directly to an objective.: it should fall as prompts and source packs improve, and a rise after a model or prompt change is the signal to roll back. The share of items escalated to a higher tier shows whether tier assignment at the brief is working. Review time per item shows whether the workflow is sustainable at the volume the team produces.
The learning loop closes in three places. Prompts and templates absorb the recurring defects — if the model keeps inventing sources, the template says "cite only from the pack" more forcefully and the operator checks citations before handing over. Source packs absorb the gaps — a defect traced to a missing or stale source becomes an update to the pack. Tier assignments absorb surprises — a use that keeps producing regulated-tier defects on the published tier is moved up. The NIST AI Risk Management Framework describes these habits as measuring and managing: observe how the system actually behaves, then act on it. Before a new prompt, model or tool enters the workflow at all, the method in testing prompts and models shows whether its output is good enough to be worth reviewing. A marketing team does not need the framework's full apparatus to do that; it needs the log and a monthly hour to read it. The AI governance operations playbook books that read beside the incident list and the use register, so what review finds moves tiers, prompts and source packs on a regular rhythm.
For a single piece at the review step, the AI-assisted content review checklist turns these checks into a list a reviewer can tick, from primary sources to a recorded sign-off.
Common mistakes
- Treating review as proofreading. Fluent text with wrong facts passes a proofread every time.
- Accepting the model's citations as verification. A citation is a pointer; the reviewer follows it.
- Reviewing everything at the same depth. The queue grows and the regulated tier gets skimmed.
- Assigning the tier at the end. The tier decides the source pack and the reviewer; it belongs in the brief.
- Operators approving their own output. They read what they meant.
- No substantiation file. The claim check has nothing to check against, so comparatives slip through.
- Sign-off without a record. When a problem surfaces later, nobody can show the review happened.
- Never reading the log. The workflow cannot improve if no one looks at what it finds.
