How to verify AI output when an error costs your reputation
A model names non-existent provisions and figures as confidently as real ones — in the text they are indistinguishable. For a lawyer, a doctor, a finance professional or a writer, that makes the tool unusable right up until there is a verification process. Below it is, and it takes minutes.
A model names non-existent provisions and figures as confidently as real ones — in the text they are indistinguishable. For a lawyer, a doctor, a finance professional or a writer, that makes the tool unusable right up until there is a verification process. Below it is, and it takes minutes.
"I did a contract analysis in ten minutes" — and then it got frightening
There is a distinct group of people for whom AI worked out quickly and well. Lawyers, doctors, accountants, HR professionals, procurement specialists. They did not learn "AI in general" — they brought the tool into their profession and got a result immediately.
From my survey: a contract analysis that took a day was done in ten minutes.
And then comes the chilling thought: what if there had been an error in it? Not a typo but an invented clause, a non-existent provision, a figure transposed. The contract would have gone to the client. The report to the patient. The analysis to management.
In ordinary work an error costs time. In these professions it costs a reputation, and sometimes a licence. So here you cannot "just try it and see" — here you need a protocol.
Why a model states falsehoods confidently
It is not that the model is bad, and not that you asked wrongly.
The root cause: a model is optimised for plausibility rather than truth. It assembles an answer that looks correct — and when the fact does not exist it will still assemble an answer that looks correct. It has no blanks.
From that follows the main property to remember: an error looks exactly like a correct answer. You will get no signal saying "I am unsure here". A section number will look like a section number, a dosage like a dosage, a reference like a reference.
The most dangerous misconception here: "I am a professional, I will notice". You notice in your narrow area. And the error arrives where you are not the expert: an adjacent provision, a regional exception, last year's amendment. Plus the fluency effect works: smooth text reads as checked.
And a second: "the better the model, the lower the risk". The risk changes quantitatively rather than qualitatively. Rarer does not mean never, and what you need is a process that holds for the rare cases too.
The zero-trust protocol
There is one rule: nothing verifiable goes out unverified. Below is how that works step by step.
Step 1. Split the answer into two kinds of content
That is the key skill, and it takes seconds.
Verifiable statements: provision numbers, dates, sums, names, dosages, references, quotations, statistics. Anything with a source outside the model.
Reasoning and structure: the logic of an argument, the order of sections, the wording, the possible readings.
The second you can judge with your own expertise. The first is verified only against the source, always, without exception. Even when it looks obviously right.
Step 2. Ask for the source in the request itself
Phrase it so the model separates facts from assumptions: "for each statement, state what it rests on; where there is no source, say plainly that it is an assumption".
That is not a guarantee — a reference can be invented too. But it gives you a list of what needs checking instead of continuous text in which everything needs checking.
Step 3. Verify against the primary source rather than by asking again
The most widespread mistake: asking the same model "are you sure?". It will either agree or reverse itself — and neither means anything.
Verification happens outside: the official text of the provision, a register, a reference work, the manufacturer's site. Opened, found, compared. That is the work that cannot be delegated.
Step 4. Cross-check with two models where the cost of error is high
Different models err differently. Matching answers do not prove correctness, but a divergence is a reliable signal that this is where to dig. It is a cheap way of highlighting risky places in a long document.
Step 5. Never hand over generated text as it is
The final text always passes through you: you reread it and take responsibility for every statement. Not "the model wrote it" — you wrote it. If you are not prepared to answer for a sentence, it should not be in the document.
Where the risk is near zero and where it is unacceptable
That distinction saves the most time.
Almost always safe: drafts and structure, rephrasing something you already wrote, summarising a text in front of you, generating questions and checklists, translating the complex into plain language, finding contradictions inside a supplied document.
Note the pattern: if you supplied all the material yourself, the model has nothing to invent. That is the most reliable mode of working, and in professions with a high cost of error it is exactly those tasks that give the main benefit.
Requires full verification: any provisions, references, figures, dosages, deadlines, links, quotations and anything that was not in the material supplied.
Never handed to a model at all: a final decision with legal or medical consequences, and personal data that must not leave the building.
How long this actually takes
The main objection to a protocol is "checking takes longer than doing it myself". Usually untrue, and here is the arithmetic.
Analysing a contract took a day. With a model it takes ten minutes plus verification. What has to be checked is not the whole text but the list of verifiable statements — usually five to fifteen items: provision numbers, dates, sums. A minute each is a quarter of an hour.
Total: half an hour against a day. Even if verification takes an hour, the ratio does not change qualitatively.
The saving disappears in only one case — when a text contains hundreds of verifiable statements. That is a sign the task was chosen wrongly: things like that are not given to a model whole, they are broken up.
Separately on "they do not know at work"
Some people in this situation use AI quietly — because it is not accepted in the organisation or is explicitly not permitted.
I will say it plainly: that is a separate risk, and not a technical one. Work you delivered using a tool nobody knows about becomes, at the first error, your personal problem twice over: the error plus the concealment.
The protocol above reduces the probability of an error. It does not settle the question of whether your employer knows how the work is done. That is worth settling separately — and it is usually settled more easily than it seems, if you arrive not with "I use AI" but with a result and a description of how it is verified.
Three phrasings that reduce the risk at the input
Verification is always obligatory, but some errors are cut off in the request itself.
"Rely only on the text I attached. If the answer is not in it, say the data is insufficient." That is the strongest technique: it closes off the model's ability to draw on memory. It works for analysing documents, contracts, reports.
"Separate facts from interpretation: first a list of statements with the source of each, then your conclusions." You get a ready list to check instead of continuous text.
"List what is missing from this document and which questions remain open." The reverse request: instead of the model filling the gaps, it names them. For professions where an omission is more dangerous than an error, that is the most useful mode.
How to build this into your work
The platform has a dedicated express course on exactly this method — "Perplexity as an investigative tool". It is not about AI in general, it is about a skill: how to check a fact, a company, a figure or a reference in minutes instead of trusting a text. The same skill is needed for checking models.
Alongside are 200+ AI assistants across 14 categories, including the legal department and analytics: they work in the mode where you supply the material — that is, the safest one.
Registration is free. Three days of the Basic plan switch on with one button, no card needed: Basic includes 200+ AI assistants, prompt texts, skill files, five programme courses (including “Sales”) and six Basic mini-courses — among them “Freelancer positioning”, Freelance 2.0, Perplexity and Grok.
Do one thing today. Take the last document you made with a model and find every verifiable statement in it — numbers, dates, sums, references. Check them against sources. If everything matches, good, you now have a habit. If not, you have just avoided the thing this article was written for.
If you have a colleague who submits documents without checking the figures, forward this to them.
Where the verification skill comes from
Verification is a method rather than an intuition, and it is taught separately.
The course "Advanced prompting" covers sixteen methods in four groups: fundamentals and context · logic and reasoning · multi-step processes · autonomous methods.
The second group is directly about this: how to make a model reason step by step and show its working, so an error is visible in the process rather than only in the result.
The Perplexity express course is about a tool that searches and supplies references. The key skill from it: how to read the output critically, because a reference in an answer does not mean it supports what was said.
Adaptive RAG: answers from a knowledge base from the automation templates catalogue is the next level: the system classifies the question and picks a search strategy over your own knowledge base (factual, analytical, opinion, contextual) instead of answering from memory. The answer is tied to your documents — verification built into the process itself.
The "Advanced prompting" course — Basic; Perplexity — Basic; the Adaptive RAG template — Creator, Make, N8N and Full.
Verification as a profession in its own right
The express course "Perplexity as an investigative tool", Basic plan or $9. It closes exactly the pain this article is about — distrust of the result. The key points:
The "search first" approach — long detailed prompts work worse than short ones, because they mix search intent with output intent. They have to be separated.
Precise search techniques — operators, source filters, working with databases rather than the general web.
Structured synthesis — how to make a model work as an analyst rather than a paraphraser.
The course turns a user into a senior research analyst — that is the phrasing from the programme itself.
Assistants for verification, Basic plan, the catalogue: Analytical journalist — subject analysis and working with sources. Academic editor — where precision of wording matters. Brainstorm AI: Devil's Advocacy — it looks for the weak points in a conclusion you liked.
Skills where verification is built into the process
The protocol from this article can be performed by hand. Or you can take a tool where it is already assembled.
- Fact verification on the CoVe protocol — the strictest of them all: it extracts five types of statement, then runs a separate verifier in an isolated context that cannot see the original text. That removes the main problem of self-checking — the model does not confirm itself
- Bibliographic reference audit — a three-level check of every reference: does the source exist, is the metadata correct, and does it support the thing it was cited for. It catches invented co-authors and citations that do not apply
- Fact checking in documents — it checks against official sources and corrects only after your explicit approval. Oriented to what dates quickly: versions, prices, specifications
- Adversarial claim review — for calculations and figures: each is checked by a script that fails with an error on a mismatch. Four verdicts, including "empty claim"
- External claim verification — it catches three dangerous patterns: round marker numbers like 10 000, chains of "statistics from statistics", and data from an interested party
- Scientific claim verification — along the chain publication → method → data → replication, where the cost of error is high
All skills above are advanced: they open on Creator and above, or for $9 once, or for 1000 XP.
What to take from this without installing anything. The principle of the first item: the verifier must not see the original text. If you ask the same model in the same chat "are you sure?", what you get is not verification but politeness.
What to take on the platform for this task
- Legal assistant — a paralegal on Russian law; works with a document you supply. Basic plan
- RAG pipeline: from PDF upload to chat answers — answers strictly from your files; Creator, Make, N8N or Full plan
- Prompt engineering: a basic course on working with all AI models — an express course, $9, Basic plan and above
- AI tools encyclopedia: 198 verified services — an express course, $9, Basic plan and above
What to read next
How to verify a fact with AI search — the tool that does it quickly.
AI for lawyers — the protocol for a profession with a high cost of error.
AI for engineers — where calculations cannot be trusted.
AI in medicine — the boundary of what is permissible.
Best open LLMs 2026 — the full guide to open models.