Product

Solutions

Security

Company

Resources

Request a Demo

Product

Solutions

Security

Company

Resources

Request a Demo

Article

2026-10-09

AI protective orders: what courts now limit in discovery

Protective orders are starting to say which AI tools may touch confidential discovery material.

Halfway through a call about a trade secrets protective order, counsel asks everyone to open the new AI paragraph. It would keep CONFIDENTIAL and attorneys’-eyes-only material out of any tool whose provider can train on it or share it. Both teams use AI, but nobody on the call has the agreement covering the account where the pricing workbook would go.

An AI protective order can limit which tools receive confidential discovery. In Morgan v. V2X, the court required contractual restrictions on training and third-party disclosure, subject to a narrow service-delivery exception, and required the plaintiff to identify his tool. At Turbo Law, we put our terms in writing so counsel can compare them with the order before a production enters the workspace.

What is an AI clause in a protective order?

It sets conditions on putting protected discovery into an AI system, including information copied into a prompt. Uploading the pricing workbook is the obvious example. Typing its confidential pricing formula into a question about deposition strategy can raise the same issue.

Check the order already entered.

Federal Rule of Civil Procedure 26(c)(1)(G) allows a court, for good cause, to restrict disclosure of trade secrets or other confidential research, development or commercial information. Restrictions in an existing order may cover transmission to an AI provider even though the parties never discussed generative AI.

An expert may be an authorized recipient. That permission does not, by itself, establish that the expert can pass the workbook to a separate service, so find the provision authorizing that recipient’s access before the file leaves counsel’s hands.

What did the court require in Morgan v. V2X?

The provider needed contractual restrictions on confidential inputs, and the plaintiff had to identify his AI platform. In Morgan v. V2X, Inc., No. 1:25-cv-01991-SKC-MDB, ECF No. 65 (D. Colo. Mar. 30, 2026), Magistrate Judge Maritza Dominguez Braswell granted in part a defendant’s motion to amend a stipulated protective order in an employment action brought by a pro se plaintiff.

Under the amended language, a provider had to be contractually prohibited from storing or using inputs to train or improve its model. Disclosure to third parties was also prohibited unless essential to delivering the service, and those recipients had to be bound by equally protective terms. The order contains further conditions. Read those before borrowing its language.

Disclosure had a deadline.

Within ten days, the plaintiff had to identify any AI platform he had used with designated CONFIDENTIAL information. Although his AI-assisted litigation preparation could receive work-product protection under Rule 26(b)(3), he had not established that identifying the tool would expose protected strategy.

We would assess a demand for the provider’s name separately from a demand for the prompts used to prepare a witness examination. The latter calls for examining the litigation preparation being sought.

What is an AI clause in a protective order?

Check the order already entered.

What did the court require in Morgan v. V2X?

Disclosure had a deadline.

What AI terms are parties negotiating?

Parties are negotiating restrictions on the provider’s handling of protected information and the records needed to check that handling. Start with the proposed use. For the pricing workbook, these are the provisions we would mark up, without assuming any court requires all of them:

No training: Make the restriction enforceable and cover training or model improvement using protected inputs. Follow it through to providers behind the application.
Limited third-party disclosure: Specify permitted service-delivery disclosures and require equivalent confidentiality obligations downstream. Establish who receives the workbook’s contents.
Tool disclosure: Set out any advance-notice requirement, the identifying information required and how service changes will be reported. Leave room to resolve objections without automatically demanding litigation prompts.
AEO limits: Address attorneys’-eyes-only material expressly. Permission to use AI should not quietly enlarge the permitted-recipient provisions.
Deletion at case end: Cover uploaded files and retained derivatives, with terms that account for backups and legal holds. Read the retention agreement before promising a deletion deadline.
Access logs: Identify the records to retain and the process for investigating suspected disclosure. Preserve privilege and work-product objections concerning substantive queries.

Try administering the language before signing it. A receiving party may promise compliance, only to discover that an expert’s assistant uses an unidentified service whose terms nobody reviewed during negotiations.

Have the person responsible for the firm’s contracts examine any disputed requirement before counsel agrees to it.

How should both sides raise AI terms at the Rule 26(f) conference?

Circulate the proposed AI restrictions with the ESI protocol and privilege provisions before the conference. Federal Rule of Civil Procedure 26(f)(3) requires a discovery plan addressing electronically stored information and privilege issues, including whether the parties should seek an order under Federal Rule of Evidence 502.

Bring the sales ledger.

Ask counsel to explain how the receiving team could analyze that confidential document under the proposed clause, then repeat the exercise with an AEO technical drawing whose permitted recipients may be narrower. Agreement about training can conceal disagreement about “access”. Work through that disagreement while the draft is open.

Production-hosting language needs attention here, too. An ESI protocol should not appear to permit processing that the protective order forbids.

Under Federal Rule of Evidence 502(d), an order can provide that disclosure connected with the pending litigation does not waive privilege or work-product protection, including in other federal or state proceedings. Its effect depends on the order entered. It supplies no independent permission for an otherwise prohibited upload.

Where a service’s eligibility remains disputed, record the missing contractual protection and follow the court’s discovery-dispute procedure before sending protected material to that service.

How do you check whether an AI tool meets a protective order?

Match the order’s restrictions to binding terms for the account that will receive the production. A general security page may point you toward the documents, but it cannot answer a contract question unless the applicable agreement supplies the required obligation.

Use the actual subscription agreement.

Alongside its incorporated data terms, read the training restriction closely enough to determine whether model improvement is covered, and follow any service-delivery exception through the downstream providers that may receive inputs. Ask for written answers where the documents leave a gap. Optional features and connections to another application need checking, too.

Keep a record of the reviewed account and contract version, with unresolved restrictions identified. Then tell the production paralegal which workspace is approved and what must be held back. “We reviewed the vendor” leaves too much unanswered when the upload is waiting.

Counsel’s professional duties require a separate assessment. ABA Model Rule 1.6(c) requires reasonable efforts to prevent unauthorized access or disclosure. Under ABA Formal Opinion 512, Generative Artificial Intelligence Tools (July 29, 2024), lawyers must evaluate disclosure risks before entering representation information into a tool; some circumstances may require informed client consent. Apply the professional rules adopted in the relevant jurisdiction.

California’s approaching effective date belongs in that review. Approved September 30, 2026, Senate Bill 574, Chapter 858, Statutes of 2026, takes effect January 1, 2027. The new Business and Professions Code section 6068.1 restricts entering confidential, personally identifying or other nonpublic information into systems unless access is limited to the attorney and authorized persons obligated to protect confidentiality.

Other provisions address delegation of legal practice and the verification and correction of AI output. The legislation also addresses disclosure of AI use in court submissions; its amendment to Code of Civil Procedure section 128.7 requires personal citation verification by an attorney responsible for submitting the filing. These provisions are not yet effective as of this post. Calendar the effective date for the firm’s policy review.

How does one secure workspace keep confidential discovery in bounds?

Keeping authorized work together reduces the occasions for copying protected material into a personal account to complete an assignment. Counsel must still determine whether that workspace’s terms satisfy the particular order, including restrictions that apply only to AEO material.

Our security commitments prohibit training on client data through contractual and technical enforcement. Client environments are logically isolated. Staff access requires an explicit support ticket and client-side approval, while every access, query and export is logged and exportable for the firm to examine when a handling question arises.

The source stays available.

Within the authorized matter file, Review reads records and builds a connected picture of the case, linking facts to source pages and lines and updating that picture as new records arrive. Pleadings and productions can be considered alongside deposition testimony and expert reports, together with work the team has already produced.

An associate preparing an examination can use the production and later testimony without assembling another packet for a separate AI account. Review, drafting, research and reporting start from that shared case understanding. The firm’s DMS remains the system of record, and Turbo Law connects to it.

We do not give legal advice or certify compliance with an order. Before uploading affected material, counsel must compare our terms with the entered restrictions and resolve any mismatch.

FAQ

Can I upload confidential documents produced in discovery to an AI tool?

Permission depends on the governing order and the applicable contractual and professional requirements. Before uploading, check the terms for the account you intend to use; paying for a subscription does not establish that the upload is permitted.

Do I have to tell the other side which AI tools I use?

Begin with the order and any discovery request. Morgan required disclosure on its facts. It created no nationwide automatic disclosure obligation. A demand for tool identity needs its own assessment, separate from a request seeking protected litigation preparation.

Does removing the confidentiality stamp make an upload permissible?

The designation survives removal of the stamp. Copying the substance into a prompt can still transmit protected information; challenge a disputed designation through the procedure in the order before treating the material as unrestricted.

Does an AI clause cover experts and vendors?

Read its wording together with the authorized-recipient provisions. An expert’s confidentiality acknowledgment does not establish what the expert’s software may do with a production. Before sending files, ask about proposed AI processing and resolve whether the order permits it.

Sources

How does one secure workspace keep confidential discovery in bounds?

The source stays available.

We do not give legal advice or certify compliance with an order. Before uploading affected material, counsel must compare our terms with the entered restrictions and resolve any mismatch.