← Back

Guide · Published November 14, 2026 · Reviewed November 14, 2026 · Ingenious AI team · Not legal advice

← All guides

When Does an ADMT Duty Not Apply? The Documented Negatives

The short answer: A duty does not apply when one of its conditions genuinely fails, and the kit's data names the conditions precisely: the tool is a minor input to a genuinely human decision (Colorado and California, with token review expressly failing California's test); the decision is not on the state's list (insurance in California; advertising expressly excluded there); a threshold is not met (California's CCPA criteria; Connecticut's 35,000 / sensitive-data / data-sale triggers); an express exclusion covers the use (Connecticut's scheduling and planning, health and safety, nonmaterial changes, productivity monitoring); or an exception position holds for a specific duty (California's opt-out exceptions). Each negative is a finding — facts, test applied, conclusion, date, review trigger — or it is nothing.

Check your duties and dates — free Build your first notice — free

What is a documented negative? It is the register entry that records why a duty does not attach to a specific tool, decision, and state — held to the same evidentiary standard as a positive duty set, because it will be read by the same readers (your future self, an internal auditor, an examiner) with the same scepticism. The scope tests themselves are in How to Tell If a Tool Is ADMT; this guide is about the findings practice around them.

Negative one: the genuinely minor role

The strongest negative, and the most abused. Its honest form requires the evidence the role guide prescribes: the removal test (the decision proceeds substantively unchanged without the tool for a week), the override behaviour (reviewers regularly reach outcomes the tool did not suggest), and California's three parts where relied on there (interpretation competence, full review, authority). All three recorded, with the measurements, not the adjectives.

Its dishonest form — 'a human is in the loop' asserted from the org chart — is token review's mirror image, and it fails for the same reason: the test is behavioural. A negative finding on role should therefore carry a re-test date and the metrics that would change it (a falling override rate is the canonical trigger). Roles drift toward the tool over time; negatives must be built to expire.

Negative two: the unlisted decision (and the express exclusion)

Listing negatives are textual: the decision is not on the state's list (the insurance-in-California pattern), or an exclusion expressly removes it (advertising in California; scheduling and planning in Connecticut's Act). The finding quotes the list or exclusion, maps the decision against it, and records the boundary analysis — especially the output-mapping check that the tool's output does not cross into a listed decision elsewhere in the workflow.

Boundary honesty is the difference between a finding and a hope. A marketing tool whose segments feed credit terms is not 'the advertising exclusion' case; the finding must say what the outputs feed, because that fact, not the product's label, decides the cell. Where the facts are mixed, the finding records the analysis and the advice taken — a 'probably out, analysis attached' entry is worth more than a clean assertion nobody can re-derive.

Worked example: Pioneer Valley Credit Services

Example — fictional business, real duty structure. Pioneer Valley, the fictional lender from the role-test guide, keeps a findings file where the negatives are documented to the same standard as its duty sets. Four of its current negatives:

Negative three: thresholds not met

Threshold negatives are arithmetic with sources: California's three criteria each measured (revenue against the adjusted figure; consumers or households counted as the Act counts them; data-sale share from the ledger), Connecticut's three triggers likewise (Connecticut consumers with payment-only excluded; sensitive-data classification from the inventory; data-sale offering from the business model). Each value with its period and method, per the threshold guides.

Growth is the re-test trigger, and it should be quantified in the finding itself ('re-test at 120,000 consumers or any data-product launch'), because unquantified triggers are never pulled. A threshold negative without a tripwire is a permanent conclusion drawn from temporary facts — the exact artefact an examiner discounts.

Exception positions: duty-specific negatives

California's opt-out exceptions are negatives of a narrower kind: they remove one duty (the opt-out) on stated conditions — the qualified-reviewer appeal offered, the security-only use, the evaluated-tool position for hiring, admission, and work assignment. The finding must therefore be duty-specific (the notice, access, and assessment duties stand untouched), condition-specific (the appeal exists and meets the reviewer standard; the security tool's output map shows no second use; the evaluation file is current), and dated like every other finding.

Exception findings decay fastest, because their conditions are operational: a reviewer rota lapses, a fraud score gets wired into a second decision, an evaluation ages past the tool's update. The review cadence for exceptions should be the shortest in the register, and any condition failure converts the finding back into the duty automatically — that conversion rule, written into the finding, is what makes the exception safe to rely on.

The findings file: one standard for yes and no

Assemble the negatives where the positives live: the register, one entry per tool-decision-state cell, positive or negative, each with facts, test, conclusion, date, owner, and review trigger. Uniformity is the point — a register where only the covered tools have files teaches a reader to distrust the gaps; a register where every cell has a finding teaches them the analysis was actually done.

Close the loop with the checker: every finding, positive or negative, should correspond to a checker run whose output is filed with it (the checker returns 'no duties found' with its reasoning for role negatives, and 'check the list' for listing questions). Where the finding and the checker disagree, the disagreement is resolved in writing before the programme proceeds — the file that contains a resolved disagreement is stronger than a file that contains none. Confirm your cells in the free checker and keep the printouts; they are the cheapest evidence in the whole programme.

  1. Write every negative as a finding: facts measured, test applied in the regime's own words, conclusion, date, owner
  2. For role negatives, attach the behavioural evidence — removal test, override rates, the reviewer's three parts — and set a re-test date
  3. For listing and threshold negatives, quote the list or measure the numbers, map the boundaries, and set a quantified tripwire
  4. For exception positions, record the conditions and the automatic conversion rule: condition fails, duty returns
  5. File the checker output with every finding, and resolve any disagreement in writing before proceeding

Sources, as carried in the kit's own data: the duty summaries in app/rules.py and the checker data in app/checker_data.py, verified against primary or authoritative sources on October 28, 2026. Run the free duties checker for your jurisdiction, decision type, and the tool's role, and compare the kit's plans for the record, builder, and review queue. More from Ingenious AI is on the hub.

Build your first notice — free

Related guides

See all ADMT guides for the full set.

Frequently asked questions

When does an ADMT duty genuinely not apply?

When a condition fails on the facts: the tool's role is genuinely minor to a human decision; the decision is not on the state's list or is expressly excluded; a coverage threshold is not met; or a specific exception's conditions hold for that duty.

Is 'a human approves it' enough for a negative finding?

No. California's test requires the reviewer to know how to interpret the output, review it with other relevant information, and hold authority to change the decision — and token review expressly fails. The finding needs behavioural evidence, not an org chart.

How long does a negative finding last?

Until its review trigger fires: a role re-test date, a growth tripwire, a pipeline change, a vendor update, a lapsed condition. Findings are dated positions about changing facts, not permanent classifications.

What happens when an exception's conditions fail?

The duty returns. Write the conversion rule into the finding itself — if the reviewer rota lapses, the security tool gains a second use, or the evaluation goes stale, the opt-out duty activates and the finding is re-issued.

Should negatives be kept with the positive duty sets?

Yes — in the same register, to the same standard, each with its checker output filed. Uniform findings are what make the register credible as a whole.

Build your first notice — free