Accuracy, Hallucination & Liability

"Hallucination" is the term of art for when an AI states something false as though it were true - an invented statistic, a misquoted standard, a court case that never happened, a citation to a paper that does not exist, a plausible-looking number that is simply wrong. It is not lying, because lying implies knowing the truth and choosing against it. It is closer to confabulation: the model generates the most *plausible-sounding* continuation, and a fabricated clause number can be more plausible-sounding than an honest "I do not know."

Understanding the mechanism tells you exactly where to be careful. Because Claude predicts plausible text, it is most dangerous precisely where plausible and true diverge: specific facts that look like other facts. A byelaw setback, an IS code designation, a product's fire rating, a supplier's lead time, a historical date, a page reference - these all have the *form* of things Claude has seen, so it will produce something of the right shape whether or not it has the right content. It is least dangerous on tasks where plausibility and usefulness are the same thing: restructuring your notes, suggesting concept directions, improving your prose, laying out an argument you then check.

Two aggravating factors are worth naming. First, Claude's knowledge has a cutoff: without web search it does not know the newest code amendment, product or event, and will happily reason from an outdated fact. Second, it is agreeable - press it, and it may "correct" a right answer to a wrong one just because you sounded doubtful, or invent support for a premise you fed it. Neither of these is a reason to distrust Claude across the board; that would waste an extraordinary tool. It is a reason to know the failure's shape so your guard goes up in exactly the right places - the specific, the numeric, the recent, the citable - and can relax where the work is divergent and low-stakes.

The antidote is not suspicion of everything; it is *proportioned* verification, applied as a habit rather than a mood. The governing rule is the one this course keeps returning to: scale your scrutiny to the stakes. A brainstorm of mood-words needs a glance. A specification clause that goes to site needs line-by-line confirmation against the actual source.

A practical hierarchy of checks, from lightest to heaviest:

SCRUTINY LADDER (match to stakes)
- Divergent / low-stakes  -> read for sense, curate
- Facts that shape a doc  -> spot-check the load-bearing ones
- Numbers that matter     -> recompute or check with a real tool
- Codes, clauses, specs   -> verify EVERY item at the primary source
- Anything you will seal   -> full professional review, as if a
                             junior drafted it - because one did

Some concrete techniques. Ask for sources, then check the sources - not the summary. A citation Claude gives you is itself a claim that can be hallucinated, so open it. Make Claude show its working on any calculation (extended-thinking or a plain "show each step") so you can audit the method, then verify the arithmetic yourself or in a spreadsheet, because the reasoning can be sound and a single figure still wrong. Cross-examine rather than accept: ask "what would make this wrong?" or "what are you least sure of here?" - Claude is often honest about its own uncertainty when you ask directly. And never let agreeableness settle a fact: if you challenge an answer and it flips, that tells you it was never anchored, so go to the source.

The deepest technique is a mindset: read Claude's output the way you read a talented junior's first draft - gratefully, and with a red pen always in hand. You are not looking for reasons to reject it; you are looking for the one confident sentence that is quietly false. Find that reliably and Claude becomes not just fast but safe.

Now the part no upgrade will change. Claude is not a lawyer, a structural engineer, a quantity surveyor or a code official, and it cannot become one. Its output is a draft or a lead to verify, never a ruling, a stamped calculation or a signed-off document. It cannot hold a professional registration, carry indemnity insurance, or stand behind its work in front of a client or a court. When a Claude-drafted spec, letter, estimate or clause goes out under your name, the professional and legal responsibility is one hundred per cent yours - exactly as it would be if a junior in your office had drafted it.

This is not a burden the tool has added; it is the definition of being the professional of record, and it is unchanged from the pre-AI world. "The AI told me" is not a defence, any more than "my intern told me" ever was. Your standard of care - the level of skill and diligence expected of a competent practitioner - applies to everything you issue, regardless of how it was drafted. If anything, using a fast drafting tool raises the bar on your review, because volume can outrun attention if you let it.

The figure captures the chain: Claude drafts, you verify, you decide, you seal - and liability lives at the seal, where it always has. The line never moves back up the chain to the tool. Practically, this means three commitments. Never issue Claude output you have not personally verified to the standard the document demands. Never represent AI-drafted work as more certain than your checking supports - if you have not confirmed a figure, do not let it read as confirmed. And for anything with legal, safety or contractual weight, route it through the real professional - the lawyer, the engineer, the QS, the code authority - because their sign-off is what Claude fundamentally cannot give. Handled this way, Claude reduces your drudgery without ever reducing your accountability, which is precisely the trade a professional should want.

Discipline that depends on remembering to be careful fails on a busy Friday. The fix is to bake verification into your process so it happens whether or not you feel vigilant. Three habits do most of the work.

First, mark the provenance of anything Claude touches until it is checked. A simple convention - keeping unverified figures in brackets, or a "CLAUDE-DRAFT, UNCHECKED" header on a working document - stops a plausible draft from silently graduating into an issued one. The most dangerous hallucination is the one that ages into assumed fact because nobody remembers it came from a machine.

Second, build source-checking into the task, not after it. When you ask Claude for anything factual, ask in the same breath for the source and the exact wording, so verifying is a click, not a research project. For codes and standards, keep the primary document open beside the chat; treat Claude as the thing that points you to the clause, and your own eyes as the thing that reads it - the cardinal rule from Module 2, now a whole-practice reflex.

Third, make the review a named step with an owner, as Module 9 taught for quality control. On anything that ships, someone - ideally not the person who prompted it - reads it as if a junior drafted it, because one did. A second pair of eyes catches the confident-wrong sentence the author's fluency-blindness slides past.

None of this slows a good practice down; it is simply where the time you saved on drafting goes, and it is a fraction of that saving. The reward is the one thing worth protecting above speed: you can put your name on Claude-assisted work and mean it, because you know exactly what you checked and why. That confidence - earned, not assumed - is what separates a professional using AI well from an amateur using it dangerously.