Demystify.

A language model predicts the next word. That's the whole mechanism. It was trained on a large slice of human text, and it writes the most plausible continuation of whatever you give it.

Plausible is not the same as true. The model has no database it looks things up in. It has no flag that says "I'm guessing now." When it knows a topic well, plausible and true overlap. When it doesn't, it produces fluent, formatted, confident text that is wrong.

Two consequences matter for your work:

1. Closed-book vs open-book. Ask the model a question from memory and it's closed-book: fluent, shaky on specifics. Paste your source document into the prompt and it's open-book: accuracy jumps. An accountant who pastes the actual IRS notice gets a far better summary than one who asks "what does IRS Notice 2024-XX say?"
2. Training cutoff. The model learned from text up to a fixed date. It does not know this year's mileage rate, last month's bar opinion, or the funder deadline that moved.

Rule of thumb: supply the source whenever you can. When you can't, treat every specific claim as unverified.

Hallucination. The model invents specifics that look real. Attorneys have been sanctioned for filing briefs with fabricated case citations. Mata v. Avianca (2023) was the first famous one, not the last. A grant writer asking for "statistics on food insecurity in my county" can get a precise-sounding number with no source behind it. Invented citations, quotes, statistics, and product names are the most common failure in professional use.

Staleness. The model's knowledge stops at its training cutoff. Tax thresholds, GI Bill rates, state bar rules, curriculum standards: anything that changes on a schedule may be a year or more out of date. A veteran checking benefit eligibility or a bookkeeper quoting a standard deduction must verify against the official source for the current year.

Confidence without accuracy. The model states guesses in the same tone as facts. A teacher asking whether a lesson aligns to a specific state standard gets a confident yes either way. There is no hesitation to read. You cannot judge reliability from tone, only from checking.

None of these are rare edge cases. Expect all three in your first week. The point of this stage is that they cost you minutes, not your reputation.

The working rule for everything in ADAPT: AI drafts. You verify. You own.

Run three checks on any output that leaves your hands, whether it goes to a client, a funder, a student's parent, a hiring manager, or a court.

1. Fact check. Highlight every name, number, date, quote, and citation. Verify each against a source you trust. Not against the AI. Asking "are you sure?" proves nothing.
2. Fit check. Facts can be right and the document still wrong. Check length limits, required elements, audience, and whether it sounds like you. A military spouse freelancer whose client capped the proposal at two pages fails this check with a perfect three-page draft.
3. Name check. Read it once as the recipient. Would you sign it? "The AI wrote that part" is not a defense anywhere, not to a bar, an auditor, a principal, or a program officer.

Then apply the time test. If verifying takes more than about half the time of writing from scratch, the task is a poor AI fit. Move it back to your Stage 1 map as "draft only" or "not AI." That feedback loop, protocol results updating your leverage map, is what makes the map honest.

The universal rule: never paste anything you couldn't email to a stranger. That covers SSNs, client names tied to facts, student records, donor lists, medical details, account numbers. Replace identifiers with placeholders ([CLIENT A], [STUDENT 1]) before you prompt.

Second, know your tool's data setting. Some consumer tools may use your inputs for training; most business tiers do not. Find the setting before you paste anything sensitive. If you can't confirm, assume your input is not private.

Then the profession-specific layer:

• Attorneys: ABA Formal Opinion 512 (2024) ties AI use to your duties of competence and confidentiality. Many state bars have added their own guidance. Check yours.
• Accountants: IRC 7216 restricts use and disclosure of tax return information. Pasting client return data into a third-party tool can qualify as disclosure.
• Teachers and ex-teachers: FERPA instincts carry over. No student names, grades, or IEP details in prompts, and the same caution applies to a new employer's data after a career change.
• Nonprofits: funders are writing AI policies now. Some ask whether proposals were AI-assisted. Check before you submit, not after.

The Ethics One-Pagers below give you each profession's page in full.

Verification has a cost, so spend it where the risk lives.

Low risk, use with a skim:

• First drafts you will rewrite anyway
• Summaries of documents you pasted in (open-book)
• Reformatting: notes to outline, bullets to prose, prose to table
• Brainstorming options, naming, structure
• Critique of your own draft, practice questions, role-played objections

High risk, always run the full protocol:

• Anything with citations, statistics, or quotes
• Anything involving current rates, deadlines, laws, or standards
• Math beyond arithmetic you can eyeball. Models miscount; have the tool write the spreadsheet formula instead and check that.
• Anything going to a regulator, court, funder, or client under your name

One more calibration habit: keep a two-week log of what your tool gets wrong (the Verification Protocol includes the template). An accountant might find her tool nails reformatting but fabricates IRS form numbers. A grant writer might find his tool is reliable on structure and unreliable on local statistics. After two weeks you stop verifying blindly and start verifying where your tool fails. That pattern knowledge feeds directly into Stage 3, where you build these tasks into daily workflows.