Practical guide

What evidence shows intellectual property controls are operating in AI-enabled work?

Evidence that an intellectual property control is operating must show what information or content was classified, which rights or restrictions were checked, how the approved AI route was enforced, what happened in representative tests and real exceptions, and who decided the disposition. A policy, training completion, or tool configuration alone is design evidence—not proof of operation. This.

By Two Prune

Evidence that an intellectual property control is operating must show what information or content was classified, which rights or restrictions were checked, how the approved AI route was enforced, what happened in representative tests and real exceptions, and who decided the disposition. A policy, training completion, or tool configuration alone is design evidence—not proof of operation. This guide provides an evidence-testing method, not legal advice. Intellectual property categories, ownership, licensing, confidentiality, inventorship, authorship, and remedies vary by jurisdiction, contract, facts, and use. Route legal conclusions to qualified reviewers.

Define the control objective

Write one testable objective for the workflow. For example:

> Protected source material and third-party content enter an AI-enabled workflow only through approved tools and uses, with rights, confidentiality, provenance, human contribution, output review, exceptions, and downstream disposition recorded for the stated decision.

WIPO’s 2024 publication Generative AI: Navigating intellectual property highlights confidential-information and trade-secret exposure, terms of use, input and output issues, and combinations of technical, legal, and practical safeguards. NIST’s AI RMF Core calls for third-party and supply-chain policies that address infringement of intellectual property or other rights, plus documented roles, risks, testing, incidents, and monitoring. These sources support the evidence categories below; neither determines ownership or compliance in a particular case. WIPO: Generative AI—Navigating intellectual property NIST AI RMF Core

Sources: [1] [2]

Test operation with PROTECT

PROTECT is a seven-part evidence packet. Each item needs an owner, period, sample, result, exception rule, and limitation.

### P — Population and provenance

Identify the population of in-scope AI-assisted artifacts, prompts, source materials, tools, users, and time period. Retain:

Evidence artifact: source-and-artifact manifest. A sample without a defined population cannot support a control-rate claim.

### R — Rights, restrictions, and terms check

For each selected case, record the question actually reviewed:

Evidence artifact: dated review record linking the applicable text, reviewer, conclusion, scope, and expiry. Do not reduce a complex rights decision to a universal “safe” flag.

### O — Operating access and tool enforcement

Inspect actual operation:

Evidence artifact: configuration snapshot, access review, change record, and representative log entries. A screenshot without version, scope, or owner is weak evidence.

### T — Tests of representative and adversarial cases

Test cases should cover:

Record expected result, actual result, tester, environment, version, evidence, and disposition. Evidence artifact: test protocol and results, including failures.

### E — Exceptions and incidents

Sample approved exceptions, denied requests, policy bypasses, suspected disclosures, disputed outputs, and supplier or terms changes. For each, verify:

Evidence artifact: exception and incident register. Zero recorded exceptions can mean strong operation or weak detection; test the reporting route before interpreting the count.

### C — Content and output disposition

For sampled outputs, record whether the artifact is approved, revised, attributed, restricted, withdrawn, escalated, or held. Verify that:

Evidence artifact: output review and downstream-disposition record.

### T — Trend, ownership, and review

Aggregate only comparable cases. Report population, sample, failures, exceptions, missing evidence, configuration periods, repeat causes, overdue actions, and limitations. Do not present “no known incidents” as proof that no issue occurred.

Assign every corrective action an owner, due date, verification method, evidence artifact, and closure approver. Re-review when tools, terms, licences, sources, workflows, jurisdictions, or distribution routes materially change.

Evidence artifact: control review with supported conclusion—operating as tested, partially operating, not operating, or inconclusive.

  • source identifiers, owners or licensors, versions, acquisition routes, and dates;
  • confidential, trade-secret, licensed, public, customer, employee, and generated classifications;
  • approved and prohibited uses;
  • human contributors and review roles where relevant; and
  • downstream artifacts and distribution routes.
  • was the source obtained and used under the applicable permission, licence, contract, or exception?
  • do tool terms address submitted content, output use, model training, retention, confidentiality, and deletion as required by policy?
  • do customer, supplier, partner, or employment agreements add restrictions?
  • does the workflow require a jurisdiction-specific copyright, patent, trademark, trade-secret, or other legal review?
  • approved tool and tenant settings;
  • identity, role, and least-privilege access;
  • input restrictions, masking, retrieval boundaries, export controls, and retention;
  • monitoring and alert coverage;
  • blocked attempts and user-facing guidance; and
  • configuration changes during the period.
  • an approved public source;
  • confidential or trade-secret material;
  • third-party licensed content with a usage restriction;
  • a mixed-source artifact where provenance can be lost;
  • a generated output that closely resembles a supplied source;
  • a request to export, publish, or reuse the output; and
  • a missing, ambiguous, or disputed rights record.
  • intake time and owner;
  • containment or hold decision;
  • sources, tools, people, and downstream artifacts traced;
  • legal or specialist review where required;
  • notification or communication decision;
  • corrective action, expiry, and retest; and
  • closure authority and residual uncertainty.
  • source facts and generated material remain distinguishable where required;
  • required attribution or licence conditions follow the artifact downstream;
  • confidential content does not become public through summaries, exports, or copied prompts;
  • a qualified person reviews material legal uncertainty; and
  • downstream recipients receive the correct restriction or correction.

Copyable evidence record

| Case ID | Source and classification | Rights / terms basis | Approved tool and control | Test or observed event | Output disposition | Exception / corrective action | Owner / approver | Residual uncertainty | | --- | --- | --- | --- | --- | --- | --- | --- | --- | | IP-01 | stable source, owner, version, confidentiality | dated text and scoped reviewer decision | tenant, role, setting, version | expected versus actual evidence | approve / revise / restrict / withdraw / hold | action, due date, expiry, retest | named people | unresolved fact or legal question |

Worked test: product launch brief

This synthetic workflow lets a product team use an approved enterprise AI workspace to draft a launch brief from internal research, licensed market reports, and public product information. The control says confidential inputs must remain in the approved tenant, third-party restrictions must follow the output, and publication requires source and rights review.

The population contains 42 briefs produced during the quarter. The reviewer selects all four exceptions, all three tool-configuration periods, and a risk-based sample of eight ordinary briefs. The source manifest shows that one licensed market report permits internal analysis but not republication of its charts.

The operating evidence passes the identity and approved-tool checks, but one brief includes an AI-redrawn chart based closely on the restricted report. The output review marked it “original visualization” without recording the source restriction. The downstream record shows it was shared with an external agency but not yet published.

The control result is partially operating. The team restricts the brief, notifies the rights reviewer, traces agency copies, and replaces the chart only after a scoped decision. It does not assume that redrawing removed the restriction. The test also finds that the model workspace changed its retention setting mid-quarter; cases before and after the change are reported separately.

Corrective actions add licence conditions to the source manifest, require the output reviewer to see those conditions, test mixed-source chart generation, and alert on external sharing from restricted projects. Closure requires test evidence and approval from the control owner and qualified rights reviewer. The example demonstrates an evidence conclusion, not a legal determination about infringement or ownership.

Failure and hold conditions

Conclude not operating when a critical restriction is repeatedly bypassed, protected content is used through an unapproved route, output disposition is ignored downstream, or corrective actions close without evidence. Conclude inconclusive when the population is unknown, provenance is missing, configuration history cannot be reconstructed, logs exclude important paths, or the sample cannot test the stated objective.

Hold publication or reuse when the source or rights basis is disputed, confidentiality may be compromised, a required qualified review is absent, terms changed materially, the downstream scope is unknown, or a representative control test fails.

Two Prune may use an intellectual-property-control task in a realistic AI-enabled assessment to evaluate how a participant frames the issue, navigates evidence, uses AI support, verifies sources and restrictions, exercises judgment, and communicates a disposition. It provides assessment evidence, not legal advice, an ownership determination, proof of compliance, or an automated employment decision.

Sources

World Intellectual Property Organization: Generative AI—Navigating intellectual property

  • NIST: AI Risk Management Framework Core

Sources: [1] [2]

Sources

  1. 1.WIPO: Generative AI—Navigating intellectual property · www.wipo.int
  2. 2.NIST AI RMF Core · National Institute of Standards and Technology