First, log the last inventions with their prompt, tool, sources and reviewer
Start with a list, not a theory. Take every invented fact the team has found in recent weeks: the award nobody won, the quote nobody gave, the specification the product does not have, the statistic with no study behind it. For each one record the draft it appeared in, the prompt that produced the draft, the tool and, where the tool shows it, the model in use, what material was pasted or attached, whether the tool retrieved anything from the team's own documents, who reviewed the draft and what that review consisted of, and whether the piece was published before the invention was caught. The NIST AI Risk Management Framework, as the site's framework page applies it, calls this mapping and measuring: knowing where the tool is used and what went wrong before deciding what to do, and treating the validity of its output as something to be observed rather than assumed. A short log is enough to start.
Then read the log for the pattern, because the pattern is the diagnosis. Inventions spread across every task and every prompter, with no material attached to any of the prompts, point at the first two causes: the model was never given the facts. Inventions that reached publication point at the review step, whatever else went wrong, because something was supposed to catch them and did not. Inventions that carry a citation to the team's own help centre or document library point at retrieval. Inventions that began on a date, from a prompt that had been reliable, point at a change to the prompt, the model or the tool. And an invention that somebody "checked" by asking the tool whether it was right points at the checking routine itself. Most logs show more than one pattern at once, which is why the fixes below are taken one cause at a time rather than as a single rewrite of the whole process.
Two distinctions keep the log honest. An invented fact is different from a stale one: a price that was right when the model's training text was gathered and is wrong now is a currency problem with the same fix as invention, supplied sources, but a different lesson about what the model can know. And an invented fact caught at review is a near miss, worth logging because it shows the inputs are weak, while the same fact on a live page is an incident. Ranking consequences for pages built on unchecked drafts belong to the diagnosis for AI content that does not perform; the weight of a false claim about a product is the subject of the guide to AI claims and consumer protection. This page stays with why the inventions keep happening.
Diagnosis table
Diagnosis
| Symptom | Check | Likely cause | Fix |
|---|---|---|---|
| Figures, quotes, awards or features appear in drafts whose prompts attached no material | The prompt for each logged draft: was any source pasted or attached, and was the model told to use only that | Drafting without supplied sources, so the model fills gaps with likely text | Supply a source pack with every factual brief and instruct the model to write "not in source" for anything the pack does not cover |
| Inventions cluster on one recurring task or one person's prompts, and the prompt is a line or two long | The prompt against the six elements of a brief: audience, task, material, constraints, format, quality bar | A prompt that carries no material, audience or constraints, or asks for a fact the brief never gave | Rebuild the prompt as a brief; strike any request for a fact the brief does not supply, or supply the fact |
| Drafts were signed off and the invention was found later by a reader, a customer or a colleague | The review record: did the reviewer trace each figure, name and claim to a source, or read for tone and flow | Review that reads for fluency rather than tracing every fact to a source | Separate the fact check from the proofread, tier the review by risk, use a reviewer other than the operator and record sign-off |
| The draft cites a passage from your own library, and the passage is old, partial or about something else | Open the cited passage: is it the current version, and does it say what the sentence says | A grounded tool retrieving a stale or wrongly split passage and citing it | Archive superseded documents outside the library, rewrite split sections to stand alone, and read citations as pointers |
| A prompt that was reliable started inventing after a date, with no change the team can name | The tool's release notes or model setting, and the team's own edits to prompts and context blocks since the last test | A prompt or model changed since the last test, with no regression set to catch the shift | Store a test set beside the prompt and rerun it after every change to the prompt, the model or the tool |
| A claim was "checked" by asking the tool to confirm it, and it confirmed | The record of the check: was a source outside the tool opened and read at the cited place | A single answer checked by re-asking the tool instead of by the fact-checking routine | Run the claim-by-claim routine against primary sources and record the result with the piece |
Fix, verify and prevent each cause
Drafting without supplied sources
How language models generate text explains the mechanism: a model produces each piece of text because it is a likely continuation of everything before it, and nothing in that process checks whether a sentence is true. When the right facts are in front of the model, the likely continuation is usually the true one. When they are not, the likely continuation is something with the shape of a true answer: a statistic with a believable value, a quote in the right voice, a study with a convincing title, a feature a product of that kind might have. Nothing in the output marks the difference, and an early wrong turn is built on rather than corrected, because consistent continuations are likely continuations. The same page lists the requests that make invention most likely, and marketing briefs are full of them: a company's own prices, results and customer quotes, requests for sources, niche or recent subjects, and leading prompts that ask for "the research showing" a point.
Fix: put the material in the prompt. The AI for content guide states the rule as giving the tool sources every time the piece makes a factual claim: a draft written from supplied documents can be checked against them line by line, while a draft written from general knowledge has no reference point and can only be checked by researching the subject from scratch. Instruct the model to use only that material and to write "not in source" where it is silent, which turns an invisible invention into a visible question.
Verify: rerun the logged prompt with the material attached and the instruction added, and read the new draft against the pack; every figure, name and claim should trace, and the gaps should show as marked questions rather than as plausible prose.
Prevent: make a source pack part of every brief for factual work, and keep unsourced drafting for structure, framing and phrasing, where nothing has to be true.
A prompt that carries no material, audience or constraints
A promptThe instruction and material given to an AI language model, which works like a brief: the model knows language, not the audience or product, unless told. is a brief, and the prompt patterns guide sets out the six things a brief has to carry: the audience, a single task, the material, the constraints, the format and a quality bar. Its table of what happens when each is missing reads like the invention log. Missing material produces invented specifics that sound right. Missing constraints produce claims the business cannot support. A missing audience produces copy written for nobody in particular, which the model fills with whatever is typical of the genre, and typical product copy carries typical claims. A prompt that asks for a fact the brief never gave, "mention our recent recognition" with no recognition supplied, asks the model to make one up, and it obliges.
Fix: rebuild the prompt in the guide's brief-prompt shape, with the material under its own heading so pasted text is not mistaken for instructions, the claims that are not allowed listed under constraints, and the line that says what to do when the material does not cover something. Strike any request for a fact the brief does not supply, or supply the fact. For adaptation work, where facts drift as a piece is shortened or localised, use the guide's rewriting pattern with its "keep exactly" list of figures, legal wording and product names.
Verify: run the rebuilt prompt on the drafts that went wrong and count the claims that are not in the material; the guide's own example shows the count falling sharply once the six elements are present, and the few that remain should be caught against the pasted sheet.
Prevent: store the rebuilt prompt in the shared library with its test cases rather than in one person's chat history, and brief the tool as fully as a freelancer would need to be briefed, because if a capable person would have to ask a question, the model fills the answer in.
Review that reads for fluency
The accuracy and review workflows guide is direct about why proofreading, the review most teams already have, catches almost none of this: a person misremembers, a model composes, and the composed falsehood is delivered in the same fluent, structured, confident prose as the true sentence beside it. A reviewer who does not already know the truth cannot see the difference by reading. Review that works is a human reviewThe step where a named person checks AI-assisted work for accuracy, sources, originality, privacy and disclosure before it is published or sent. step with the sources in hand. The fact check works line by line, marking every number, proper noun, date, quotation and statement of what a product does, tracing each to the source pack and ticking or striking it; the claim check works from a list of comparatives, superlatives, outcome and category claims against a substantiation file. A citation in the draft is a pointer to be followed, not a verification.
Fix: separate the fact check from the proofread, give the reviewer the source pack, and apply the guide's rule that a statement with no source is not in the output until it has one. Assign the review tier at the brief, not at the end, so that product claims and anything regulated get the specialist pass they need, and have someone other than the operator review anything that reaches a customer, because operators read what they meant rather than what is there. Record the sign-off: who, when, which version.
Verify: take the published pieces from the log and run the fact check on them now; the inventions should be found by the check, and anything else it finds goes into the same correction.
Prevent: log defects per reviewed item by failure mode, read the log monthly, and let the recurring defects change the prompts and the source packs. The AI content review checklist turns the checks into a list a reviewer ticks before an AI-assisted piece is published.
A grounded tool citing a stale or wrongly split passage
GroundingMaking a language model answer from specified material, such as documents retrieved for the question, rather than from what it absorbed in training. narrows invention without removing it. How grounding and retrieval work describes the chain a retrieval tool follows, splitting the team's documents into passages, finding the passages that best match the question and placing them in front of the model, and the three ways it fails. The right material is never found, and a model with weak instructions answers anyway from general patterns. The wrong material is found: two versions of a policy are both in the library and the older one matches the wording better, or a passage was split away from the condition that qualifies it. Or the model strays from what it was given, combining two passages into a claim neither makes or adding a detail from training. A draft that cites a real document from the team's own library and is still wrong is usually in the second group, and those answers are especially convincing because they come with a citation. The citation records which passages were supplied, not which sentence came from which passage.
Fix: open the cited passage for each logged invention and read what it says. Where it is a superseded version, archive it outside the library rather than marking it old inside; where it is a section that lost its heading or its exception in the split, rewrite it to stand alone with the rule and its exception in the same paragraph; where the passage does not support the sentence at all, the instruction to the model needs to say that it answers only from the passages and says when they are silent.
Verify: rerun the question that produced the invention and trace the new answer to its passage; it should cite the current document, and the sentence should be supported by the cited place.
Prevent: give every document in the library one owner and a last-reviewed date, keep one current source per fact, and keep a set of real questions that is rerun after every change to the library. Where a customer-facing assistant is the tool giving the wrong answers, the diagnosis for an assistant giving customers wrong answers traces conversation logs through the same chain.
A prompt or model changed since the last test
The guide to testing prompts and models makes the point that a passed test describes one combination of prompt, context blocks, model and tool settings at one point in time. Change any of them and the result may no longer hold, and vendors update the models behind their products, sometimes with little notice. A prompt that drafted clean product pages for months and then starts adding specifications is the classic shape of this cause; the team's first instinct, that someone must have edited the prompt, is often right and otherwise points at a model update. Without a stored test setA collection of real inputs for an AI task, each noting what a good output must contain, run to compare prompts or models and rerun when they change., the shift is noticed only when the inventions reach review, or a reader.
Fix: find what changed. Compare the prompt and its shared context blocks with the stored version, check the tool's release notes or model setting, and check whether the tool's own settings for length or creativity were altered. Then build the test set the guide describes if none exists: real inputs from the task, including the hard cases, the edge cases where the tool should decline, and the sensitive cases where a mistake would matter most, each with notes on the facts that must appear and the facts that must not.
Verify: run the stored set against the current prompt and model, score it with the rubric in which factual accuracy against the source and handling of gaps are pass-or-fail criteria, and compare the failure log with the recorded baseline; the inventions should appear in the log as a new failure type, which confirms the cause and gives the pass bar for the fix.
Prevent: rerun the set whenever the prompt or a context block is edited, the vendor announces a model change or the team switches models, tool settings change, the source documents are restructured, or reviewers notice a change in the kind of errors they catch. Set the pass bar in advance, with invented facts, prices and claims listed as failures that are unacceptable at any rate.
A single answer checked by re-asking the tool
Some inventions in the log carry a note that says the claim was checked, and the check turns out to have been a second question to the same tool: is this correct? The fact-checking tutorial explains why that is not a check: the same process that produced the error produces the reassurance. A link the tool supplies is a pointer, not a verification, and answers that combine a real source with a claim the source does not make are among the hardest errors to spot, because the link looks like evidence.
Fix: run the tutorial's routine on the answer. Copy the question and the answer with the tool and the date, split the answer into claims one per row, mark each claim's type, find the primary source for each claim that will be used, check quotes word for word, check figures with their units, periods and dates, mark what stays unknown, and record the check with its date. The tutorial's red flags, a precise figure with no source, a citation that is hard to find, a real source cited for a specific claim, a quotation with a famous name, decide which claims to check first.
Verify: the working table shows every used claim as verified, corrected, unknown or dropped, with a source opened and read for each verified one, and unknown claims appear in the piece as questions or not at all.
Prevent: apply the routine in proportion to where an answer is going. An answer used to brainstorm needs a careful read; an answer whose facts will appear in a page, a brief, a slide or a decision needs the routine, and the record of the check travels with the piece so the next person can rely on it rather than repeat it. Where the invented figures sit in reports rather than copy, the diagnosis for AI report numbers nobody can reproduce traces them to the query and the data behind the narrative.
Prevention: no draft without a source pack, a pass-or-fail accuracy criterion and a rerun test set
Every cause above is a missing input or a missing check, and the routine that prevents the next pattern is built from three habits the site's guides already describe.
The first is that no factual draft starts without a source pack. The accuracy guide describes the pack: the product specification, the pricing page, the approved claims file, the previous approved version, the research the piece summarises, with an owner and a date, and a source hierarchy in which the team's own structured data outranks its own prose, which outranks public primary sources, which outrank anything the model knows. The prompt names the pack as its material, instructs the model to cite which source each statement comes from, and tells it to write "not in sources" rather than fill a gap. For teams producing at volume, retrieval automates the pack, which is grounding in the sense the glossary entry gives it, and then the library needs the care the grounding page describes: one current source per fact, superseded versions archived outside the library, self-contained sections, dates and owners. The source pack is also what makes the review cheap, because most statements trace quickly once the pack is good.
The second is that factual accuracy is a pass-or-fail criterion, in review and in testing alike. The testing guide's rubric separates criteria that are matters of degree, such as tone and audience fit, from criteria where a single failure makes the output unusable, and puts any fact not in the source and any gap filled by invention in the second group, so a good score for voice never averages away an invented specification. Review applies the same rule through the fact check and the claim check, sized to the tier the piece was given at the brief, with sign-off recorded and defects logged by failure mode. A reviewer who is not the operator reads anything that reaches a customer. Published pieces are gated by the AI content review checklist, and a single answer that is about to become a brief or a slide goes through the fact-checking routine.
The third is a stored test set that is rerun after any change. Keep the set beside each recurring prompt, draw it from real work with the hard, edge and sensitive cases included, fix the rubric and the pass bar before running anything, and rerun it when the prompt, the context blocks, the model, the tool settings or the source documents change, or when reviewers notice the errors changing shape. The share of outputs that pass review first time is the figure to watch in live use, and a rise in defects after a change is the signal to roll back.
Around the three habits sits the governance the AI governance section sets out: every use of AI in a register with an owner and a tier, the review routine written down, and an incident route for the invention that reaches a customer. The NIST AI Risk Management Framework, as the site's framework page applies it, is the shape of that routine rather than a rulebook: govern with a named owner, map where drafting happens and what could go wrong, measure with the test set and the review log, and manage by changing the prompts, the packs and the tiers when the evidence says so. Google's published guidance on AI-generated content, as the AI for content guide reads it, is the reason the habits matter on published pages as well as in drafts: Google's systems aim to reward high-quality content however it is produced, the guidance points to the qualities of experience, expertise, authority and trust, and a page carrying an invented award has none of them. The AI troubleshooting hub lists the neighbouring symptoms, from an assistant quoting old terms to automation that sent the wrong message, that share these causes.
A kitchen worktop fabricator traces a fictional certification (illustrative)
Wimpleburn Worktops
Wimpleburn Worktops, an invented kitchen worktop fabricator with 2 workshops and an office of 6, finds two inventions in one week. A product page for a composite surface claims a fire-resistance certification the material has never held, and a blog post quotes a trade association saying the composite "outlasts natural stone in 9 out of 10 kitchens", a line the association never published. The product page has been live for 11 weeks.
The marketing coordinator logs the 2 pieces and then the 14 other AI-assisted drafts from the same quarter. The log shows one shape: 12 of the 16 prompts attached no material at all, every draft was reviewed by the person who prompted it, and every review was a read for tone and length. Both inventions came from drafts produced after the tool's vendor changed the model behind it, 5 weeks earlier, and 3 other drafts from the same period carry specifications the product sheets do not support, caught at review by luck rather than by method.
The fixes follow the log. Every product and blog brief now carries a source pack: the material's specification sheet, the 6 approved claims and the previous approved page. The prompt is rebuilt around the six elements, ending with the instruction to write "not in source" for anything the pack does not cover. Factual accuracy becomes a pass-or-fail criterion, and the sales manager, not the prompter, runs the fact check on anything customer-facing. The coordinator builds a test set from the 12 hardest product pages, including the 3 materials that hold certifications and the 2 that hold none, and sets the pass bar at zero invented certifications and zero figures without a source. The first rerun after the rebuild produces 1 failure in 36 outputs, an unsupported "scratch-proof", against 9 failures in the baseline run on the old prompt. Certification wording is routed to the operations director, who owns the certificates, before any page is published. The live page is corrected the same day and the blog post's quotation is removed.
Common mistakes
- Switching tools before changing the inputs. Every model writes likely text; a new one invents in a new style.
- Reading the draft to see whether it sounds right. Fluency and accuracy are produced independently; only tracing finds the composed sentence.
- Accepting a citation as verification. It shows what the model was given, or what a reference usually looks like, not that the sentence is supported.
- Asking the tool whether it was right. The process that made the error makes the reassurance.
- Leaving superseded documents in the library marked as old. Retrieval matches wording, not dates.
- Testing once at adoption. A model update changes behaviour silently; the stored test set catches it before a reader does.
- Treating a caught invention as a non-event. A near miss at review is evidence that the inputs are weak; log it with the rest.
