AI Learning Studio · Job aid 07

The Human Check.

A practical job aid for reviewing AI-supported work for accuracy, privacy, bias, usefulness, and appropriate disclosure. Five checks, sized to the stakes, finished with an honest sign-off.

Job aid · Sine Language

AccuracyPrivacyBiasUsefulnessDisclosurePrint-ready

Use it before AI-supported work leaves your hands — sent, filed, published, or acted on. A tool drafted it. Your name carries it.

01 Set the depth

Four questions size the risk. The stakes decide how hard you look.

02 Run the five checks

Work through what applies. A tick means it happened, not that it was meant to.

03 Sign off honestly

Disclose what the tools did, stand behind what you decided. Then send it.

Step 01 · Calibrate

How hard should you check?

Verification scales with consequence. Not every draft needs a forensic pass, and pretending otherwise is how checking quietly stops happening. Size the risk first, then spend your attention where an error would actually cost something.

Consequence If this is wrong, what happens?

LowAn awkward moment, easily corrected HighMoney, health, coverage, reputation, or trust

Reversibility Could a mistake be walked back?

LowEasy to recall and correct HighHard or impossible to undo

Reach Who will see or rely on it?

LowMe, or a colleague who knows the context HighClients, members, leadership, or the public

Detectability Would an error be obvious?

LowA mistake would jump out HighA mistake would look plausible and slip through

Unsized

Answer all four questions and this panel names the pass you owe the work. The arithmetic is printed below — the judgment stays yours.

Question	Low ☐	High ☐
Consequence — if this is wrong, what happens?	An awkward moment	Money, health, coverage, reputation, trust
Reversibility — could a mistake be walked back?	Easy to correct	Hard or impossible to undo
Reach — who will see or rely on it?	Me, or a colleague in context	Clients, members, leadership, the public
Detectability — would an error be obvious?	It would jump out	It would look plausible and slip through

The rule: 0 high → light pass (read as a stranger, spot-check what matters most) · 1–2 high → standard pass (all five checks) · 3–4 high → deep pass (five checks + independent sources + a second person) · unsure → round up.

Step 02 · Review

The five checks.

Run them in order — accuracy first, disclosure last. Each check is one question asked properly, with the moves that answer it. Tick as you go; on paper, a pen works the same.

Check 016 items

Accuracy

Is every claim that matters actually true?

Fluency is not evidence. Output never outranks an authoritative source.

Mark the load-bearing claims. Names, numbers, dates, dollar amounts, quotes, citations, policy statements — anything someone will rely on.
Open every citation. Confirm it exists and says what the work says it says. A complete-looking reference proves nothing.
Redo the math outside the tool. Calculator or spreadsheet. Generated arithmetic can be confidently wrong.
Compare against the record. The source document, the system of record, the policy wording, the original thread.
Ask what's missing. Work can be true line by line and still misleading by omission.
Date-check anything that changes. Prices, policies, people in roles. If it moves, it needs an "as of."

EscalateA claim that matters and won't verify: cut it, qualify it, or take it to someone who can settle it. A checked "likely" beats an unchecked "certainly."

Check 025 items

Privacy

Should this information be in here at all?

The check runs both ways: what went in, and what the work gives away.

Trace what went in. Did the prompt or attachments include personal, confidential, or client information — and is the tool approved for that class of data? Your organization's rules decide, and they decide before, not after.
Audit what comes out. Does the work share more than this reader needs? Necessity and proportionality apply to drafts too.
De-identify properly. Could someone still be recognized from role, dates, location, or circumstance? Removing the name is not the same as removing the person.
Apply the no-tool test. Would you share the underlying information with this exact audience anyway? If not, a tool in the middle changes nothing.
Watch for quiet accumulation. Small, harmless details across one document can add up to a profile no single line would justify.

EscalateUnsure whether something counts as personal or confidential? Ask your privacy contact before it goes anywhere. Before is the whole job.

Check 035 items

Bias

Who does this work for — and who does it quietly leave out?

Generated text reproduces its training patterns — including ones you'd never choose.

Find the default person. Does the work assume an age, ability, language, income, family shape, or culture without saying so?
Run the swap test. Change the name, group, or context. Does the tone, advice, or benefit of the doubt change with it?
Audit the examples. Who appears, who decides, who is the cautionary tale? Patterns teach, whether or not you meant them.
Distrust the average. "Works well overall" can hide failing badly for a smaller group. Ask who the exceptions are.
Read for access. Plain language, clear structure, described images, readable contrast. Work people can't use isn't neutral.

EscalateIf the work feeds decisions about people — who gets what, who gets flagged, who waits — one reviewer isn't enough. Add a second set of eyes, and say why.

Check 045 items

Usefulness

Is this good work, or does it just look finished?

Polish is the easiest thing to generate and the least valuable.

Re-read the ask. Does this answer the question that was asked — or a nearby, easier one?
Cut the padding. Generated drafts run long. If half the length serves the reader better, the cutting is the work.
Hunt the generic. Anywhere this could describe any team, any company, any quarter — your context is the missing ingredient. Add it or delete the passage.
Check for commitments. Does it say something specific enough to act on — and to be accountable for — or does it hedge everything?
Read as the recipient. The busiest person who'll get this: do they know what to do after one read?

EscalateIf fixing the draft is taking longer than starting over with a better brief — start over. Sunk cost applies to prompts too.

Check 055 items

Disclosure

Would anyone be surprised to learn how this was made?

Confident disclosure is a credibility move, not a confession.

Know the local norm. Does your team, organization, or audience expect AI use to be named for work like this? Where a rule exists, follow it.
Run the surprise test. If the reader learned about the AI step later, would their trust in the work change? If yes, tell them now.
Say it plainly, own it fully. "AI tools supported the drafting and research. I verified the facts, made the decisions, and I'm accountable for the final work."
Never claim a check that didn't happen. "Reviewed" means a person reviewed it. This page, honestly completed, is what makes the word true.
Keep your receipts. For consequential work, note what you verified and against what. Future-you defends the work with that list.

EscalateRules unclear and the stakes real? Default to disclosing, then ask whoever owns the channel to make the norm explicit.

Step 03 · Sign off

The part only you can do.

A checklist can carry you to the last step, but not through it. The sign-off is a human act: judgment, disclosure, and a name on the line.

The failure mode

Why checking fails.

The most common failure isn't a missing checklist — it's this one, rubber-stamped. After enough good outputs, review quietly turns into approval. The habit has a name, automation bias, and knowing about it is not the same as being exempt from it.

Time to look
Authority to say no
Evidence at hand
Competence to judge
A real chance of catching it

A human check is only real when the reviewer has all five. Missing one? Then it isn't a check yet — it's a signature. Fix the missing condition, or say plainly that the work went out unreviewed.

Three habits keep the stamp honest: check while you're fresh, not last thing. Be hardest on the parts you didn't write. And leave visible ticks — a mark you make is a decision you notice.

The busy-day fallback

The sixty-second version.

No time for the full pass? Do this much, then let the stakes decide whether you've earned the stop.

Verify the biggest claim at its source.
Justify or remove every piece of personal information.
Swap the subject — does the tone survive the change?
Cut the padding and answer the question that was asked.
Say how it was made if the making would surprise the reader.

If this were wrong and public, how bad? If the answer makes you wince, find the other four minutes.

Ready to put your name on it?

That's the real test. When you can say where every load-bearing line came from and why it's there, this stops being generated output with a wrapper and becomes your work — disclosed confidently, checked in proportion, defensible on a bad day.

0 / 26 checked

Ticks live on this page and nowhere else — nothing is stored or sent, and a reload gives the next piece of work a clean slate. Print a copy if you want a record of the review.

Reviewed by Date Checked against Depth of pass

Provenance

Where this stands.

01
NIST — AI Risk Management Framework: Generative AI Profile (AI 600-1) The failure-mode and risk vocabulary behind the five checks.
02
Office of the Privacy Commissioner of Canada — Principles for responsible, trustworthy and privacy-protective generative AI The privacy check's frame: necessity, proportionality, transparency, accountability.
03
Government of Canada — Guide on the use of generative artificial intelligence The disclosure norm: identify generated content, and review it for accuracy and bias before it ships.

This job aid is education and issue-spotting, not legal or privacy advice. Where your organization has its own AI, privacy, or disclosure policies, those govern — this page never overrides them.