Chapter 9: The TRUST Verification Method
⚡ FAST TRACK
The skill that builds confidence
Verification is the habit that turns an impressive answer into something you can use with confidence. Checking everything would waste the time AI can save, so the useful skill is learning what deserves verification and how much checking it needs.
What you need is a method: a way to decide quickly what requires verification, and a fast, repeatable procedure for doing it.
That is TRUST. Five steps, most of which take seconds, and a rule for how much of it to run.
T: Trace the claim to where it should live
R: Re-ask cold
U: Use an independent source
S: Scrutinize the specifics
T: Take responsibility
The name is a deliberate irony. You run TRUST precisely because you should not extend trust by default.
First: decide what to verify
Verification is a cost. Spend it where errors hurt.
Run the two risk questions (what does a wrong answer cost me, and would I catch it) and sort what is in front of you into three bins.
Bin one: no verification. Rewrites of your own text, brainstormed options you will evaluate yourself, explanations of well-established concepts you are learning informally, formatting, structure, tone. The majority of AI use lives here. Checking it is wasted effort.
Bin two: spot-check. Anything going to another person under your name, but with modest stakes. An internal summary, a first-draft plan, a set of talking points. Run the fast version of TRUST: steps S and T, about ninety seconds.
Bin three: full verification. Anything with money, legal exposure, health, safety, a deadline, a public audience, or your professional reputation attached. Run all five steps. Every specific gets checked at a source.
Making this decision explicitly, before you read the answer, is most of the discipline. Once you have read a fluent, persuasive response, your assessment of how much checking it needs will have already been corrupted by how good it sounded.
T: Trace the claim to where it should live
Before you check anything, ask a structural question: if this were true, where would it be recorded?
A court decision lives in a legal database. A statute lives in the code. A drug interaction lives in a pharmacological reference. A company’s revenue lives in a filing. A product’s price lives on the vendor’s own site. A study lives in a journal.
This step takes about five seconds and does two things.
It tells you whether verification is even possible. If a claim has no natural home — “most consultants recommend,” “studies suggest,” “it is generally accepted” — you are looking at an unverifiable assertion, and you should treat it as an opinion the model absorbed rather than a fact it retrieved.
And it gives you the destination for step U. You are not going to search for the claim in general. You are going to a specific authoritative place to look for a specific thing.
You can enlist the model here, but only for the address, never the answer:
For each factual claim you just made, tell me where it would be documented if true: the specific source type or authority, not a summary of what it says.
R: Re-ask cold
This step exploits a property of how these systems fail, and it comes with an important limit.
Open a new conversation. Not a follow-up: a fresh one, with no prior context. Ask the same question, phrased differently. If you can, ask a different tool entirely.
Fabrications are unstable. Because an invented citation is generated fresh each time from thin material, asking cold in a new context frequently produces a different invention. Well-supported facts, by contrast, tend to come back the same.
Two warnings, and they matter.
Asking in the same conversation is worthless. The earlier answer is in the context, and the model will simply agree with itself. This is exactly what happened in Mata v. Avianca: the attorney asked ChatGPT whether the cases were real, in the conversation that had invented them, and received a confident yes.
The signal is asymmetric. Inconsistency across cold re-asks is strong evidence of fabrication. Consistency is only weak evidence of truth: a widely repeated falsehood in the training data will come back identical every time. Passing this step does not end the process. Failing it ends it immediately.
U: Use an independent source
Only this step checks the claim against independent evidence. Everything else supports that check.
Go to the location you identified in step T and look at it directly. The bar is straightforward: the source must not be the model, and must not be another model summarizing the same territory.
Three things that are not independent verification:
Asking the AI to confirm. Covered above. It is the failure mode itself.
An AI-generated search summary. Many tools now search the web and produce a synthesis with links. This is a genuine improvement and it is not verification — the summary can misread, over-generalize, or cite a page that does not support the claim. Open the link. Read the relevant passage.
A second AI tool. Two systems trained on substantially overlapping text can share the same error. Agreement between them is not corroboration.
What counts: the primary document, the official site, the actual statute, the journal article, the vendor’s own pricing page, a qualified human.
One efficiency: you rarely need to verify an entire answer. You need to verify the specifics, which is the next step.
S: Scrutinize the specifics
Errors are not distributed evenly through an answer. They concentrate, with great reliability, in particular kinds of content.
Go through the response and mark every instance of:
• Numbers: figures, percentages, prices, measurements, dosages
• Dates and deadlines
• Proper names: people, organizations, products, places
• Citations and quotations: anything attributed to a source
• Legal and regulatory references: statutes, rules, requirements
• Superlatives and absolutes: “the only,” “always,” “the largest,” “required in all cases”
• Causal claims: “X causes Y,” “this leads to”
Those marks are your verification list. The connective prose between them is usually fine and rarely worth checking.
This is why the step is so fast in practice. A four-hundred-word answer might contain six specifics. Checking six things at a source takes a few minutes, and you have verified the parts capable of hurting you.
One prompt earns its place here:
List every factual claim in your answer that I should independently verify before relying on it. Separate the ones you’re confident about from the ones you’re not.
Treat the confidence sorting as a rough prioritization, not as evidence. The list itself is the value.
T: Take responsibility
The last step has nothing to do with the machine.
Decide. You have the verified material. The judgment about what to do with it is yours, and it involves things the model does not have: your situation, your risk tolerance, your obligations, your knowledge of the people involved.
Own it. Once the work goes out under your name, responsibility remains with you. The tool cannot serve as a co-author, a defense, or a substitute for the professional duties that apply to your work. Courts, employers, clients, and professional bodies increasingly expect people to verify AI-assisted output before relying on it.
Record what you checked. For anything consequential, a note costing fifteen seconds: what you verified, against what source, on what date. In Mata v. Avianca the sanctions turned not on the fabrications but on the failure to be forthcoming and the continued defense of the cases after doubt arose. A contemporaneous record of what you checked is what lets you respond honestly and immediately when a question comes.
Know when to stop. When verification cannot resolve a consequential question, bring in a qualified human. TRUST helps you filter claims; expertise remains essential when the stakes are real.
The ninety-second version
Most bin-two work needs only this:
1. Scrutinize: mark the specifics.
2. Use a source: check the two or three that matter most.
3. Take responsibility: send it knowing it is yours.
Full TRUST is for bin three. The abbreviated version, run consistently, will prevent nearly everything that actually goes wrong.
Worked example
The request: What are the record-keeping requirements for a small employer in my state, and how long must I keep payroll records?
The answer: four confident paragraphs, a specific number of years, a named federal act, a named state agency, and a citation to a state regulation.
T: Trace. Federal requirements live in the Department of Labor’s own guidance. The state requirement lives in the state’s administrative code and its labor department. The citation is checkable directly.
R: Re-ask cold. New conversation, rephrased: How long must a small employer retain payroll records, and under what authority? The federal retention period comes back identical. The state citation comes back with a different section number. That discrepancy is the alarm.
U: Use an independent source. The federal guidance confirms the retention period on the agency’s own page. The state regulation cited in the first answer does not exist at either section number; the actual provision is elsewhere and specifies a different period.
S: Scrutinize. Two retention periods, one act, one agency, one citation, one threshold for “small employer.” Six specifics, four minutes.
T: Take responsibility. The federal answer is sound. The state answer was wrong in a way that would have produced a real compliance failure. Note what was checked and the date. Given that the state rule was fabricated once already, confirm the correct provision with the state agency or an employment attorney before setting policy.
Elapsed: under ten minutes. Without it: a documented policy built on a regulation that does not exist.
Exercise 9.1: Run full TRUST once
Take a real question in your own field where you would use the answer, and run all five steps. Time it.
Most people are surprised by how fast it is, and by how much shows up in step S.
Exercise 9.2: Prove the cold re-ask
Ask a narrow factual question with a specific answer. Then open three new conversations and ask the same thing, worded differently each time.
Compare. Where the answers diverge, you have found the boundary of what the model actually knows, and you have learned to see it without needing an outside source at all.
Exercise 9.3: Write your own bin rules
Write down, in one page, what in your work goes in each bin. Be specific to your job.
The method becomes useful when this sorting happens automatically. Chapter 25 folds this into your personal AI policy.
Learn More About Artificial Intelligence
Learn more about the CLEAR, TRUST and LEARN methods for Artificial Intelligence
REAL-WORLD CASE STUDY
Quiet Title, Filed Without a Lawyer
One of the most practical ways I used AI was preparing a quiet-title action I filed without an attorney in Pickens County, Georgia. The property was 3.96 acres acquired through a tax sale. I had already completed the statutory process that ended the former owner’s right to buy it back, but no title insurance company would insure the property without a court order clearing the title.
AI helped me understand the difference between Georgia’s conventional and special-master quiet-title procedures and identify the statutes that applied to my situation. It helped me organize the chain of title, analyze the recorded deeds and tax-sale documents, and draft the petition, summons, legal description, supporting affidavits, proposed orders, and exhibits. Instead of starting from a blank page in unfamiliar legal language, I described the facts in plain English and received a structured draft I could verify and revise.
Serving the former owner turned out to be the hard part. The first records pointed to Alabama; later information pointed to the Philippines. Because international service can be governed by treaty, AI helped me research whether the Hague Service Convention applied, organize the address searches and attempted-service records, and prepare the necessary correspondence. The Philippines has been a party to that convention since 2020, a detail that mattered and had to be checked at the official source.
When the defendant could not be located and served through those efforts, AI helped me draft an amended motion, affidavit, and proposed order requesting service by publication. Georgia law permits publication in certain property cases involving nonresident or unknown claimants. The court authorized publication.
I chose this example because it was the relatively ordinary kind of quiet-title case. There was no opposing attorney, no active dispute over who owned what, and no witnesses contesting the facts. That is very different from attempting to handle a genuinely contested title dispute alone. An example of that much harder situation appears in the Bonus Content section.
Everything AI produced, I checked. A draft that sounds authoritative and a draft that is correct are not necessarily the same document. The only way to tell them apart is to open the statute, court rule, deed, or other primary source yourself.
AI did not file the case, sign the pleadings, or make a decision. I confirmed the law, verified the facts, dealt with the clerk, and submitted everything. What it gave me was an unusually capable research and drafting assistant that never billed by the hour and never complained. It also never once said “the Philippines?” out loud, which is more than I managed.
You now understand why AI can be both fluent and fallible, and you have a practical way to verify what matters. That combination gives you a stronger foundation than casual use ever will.
Learn More About Artificial Intelligence
Learn more about the CLEAR, TRUST and LEARN methods for Artificial Intelligence