AI Tool Compliance Checklist for Patent Practitioners

Identify the geographic jurisdiction of the servers handling your inputs. This affects both export control analysis and data protection law applicability.

Some providers retain inputs for abuse monitoring (30 days is common), others indefinitely. Retention terms may differ between consumer, API, and enterprise tiers of the same product.

A terms-of-service opt-out is different from a contractual no-training clause. Check whether the opt-out is a toggle in settings, a contractual commitment, or a statement in a privacy policy that can change with notice.

Enterprise tiers often provide DPAs. Consumer and API tiers usually do not. A DPA creates contractual obligations around data handling that a privacy policy does not.

Confirm deletion timelines and procedures. Some providers retain data in backups for extended periods after account closure.

Practitioners must make reasonable efforts to prevent inadvertent or unauthorized disclosure of information relating to the representation of a client. The USPTO AI guidance flags two specific hazards: material retained for training, and tools operated outside the United States.

Opinion 512 calls for evaluation of: the client, the matter, the task, the sensitivity of the information, the safeguards, and the particular tool. Depending on how a system uses or exposes matter information, informed client consent may be required.

These are two different regimes and conflating them produces the wrong answer. A foreign filing license under 35 USC 184 governs filing patent applications abroad. Routing data to a server in another country is not itself a foreign filing. Ask the filing question on its own terms: was the invention made in the US, is a foreign or international application being prepared, and has a license issued on the filing receipt?

Export rules reach technical data, and a release of controlled technology to a foreign person counts as an export even when it happens inside the United States. So the question is who can reach the material: provider personnel, subprocessors, support and abuse-review teams, and any distributed infrastructure. Note that a USPTO foreign filing license does not authorize exporting subject matter abroad to prepare a US application. Data residency is a useful input here, but it is not a substitute for the analysis.

Voluntary disclosure to a third party can waive privilege unless adequate confidentiality protections are in place. Enterprise agreements with confidentiality terms reduce this risk. Consumer-tier tools with broad license grants may increase it.

Check for encryption in transit (TLS 1.2+) and at rest (AES-256 or equivalent). Verify authentication supports MFA or SSO integration.

The provider's employees with access to your data are, for privilege analysis, third parties. Determine the scope of human review, whether it is opt-in or opt-out, and which subprocessors handle data.

For matters subject to litigation holds or regulatory scrutiny, you may need to demonstrate who accessed what and when. Check whether the provider offers access logs and how long they are retained.

Review the provider's incident response commitments: notification timeline, scope of disclosure, and remediation procedures. Compare against your jurisdiction's breach notification requirements.

Local inference removes the cloud provider from the data path entirely. Open-weight models now handle many patent-adjacent tasks (claim drafting, prior art summarization, disclosure review) at usable quality levels on workstation hardware.

Local models are smaller and slower. On a 64 GB machine, the best local model scores roughly 12 points behind the best cloud model on patent drafting benchmarks. On 32 GB hardware, the gap is wider. The tradeoff is real and quantifiable.

Published patents and applications carry lower confidentiality risk when processed through cloud tools. Unpublished disclosures, draft claims, and prosecution strategy carry the highest risk. Note that publication status of the attached document is not the whole test: a prompt built around a public patent can still carry claim strategy, planned amendments, or your own impressions. The same practitioner may reasonably use different tools for different material on the same day.

Opinion 512 addresses competence, not just confidentiality. A tool that produces fluent, well-formatted output raises the cost of review rather than lowering it, because errors no longer look like errors. Name the reviewer before the workflow starts.

This one is not hypothetical. In our own drafting benchmark, every local model tested produced quotations that were real and verbatim but taken from the claims being evaluated rather than from the specification, in one run 81 times out of 81. The claims were being offered as evidence for themselves. "The quote is real" is not verification. Confirm the document, the location, and the evidentiary role. See the benchmark.

USPTO submissions carry a personal certification. AI assistance does not change who is responsible for the accuracy of the assertions, the existence of the citations, or the consistency of positions taken. There is no general duty to disclose that an AI tool was used, but the duty of candor creates a narrow exception where the use itself is material.