Do I have to disclose that I used technology-assisted review?

What the rules actually require#

There is no provision in the Federal Rules that obliges a producing party to state which review methodology it used. The obligations run to outcomes and process, not to tooling:

• Rule 26(b)(1) measures discovery by relevance and proportionality. A method that returns the responsive material proportionately satisfies it; the label on the method is not the test.
• Rule 34(b) governs the form of production and the response to a specified form. It says nothing about how documents were identified.
• Rule 26(g) is the one with teeth. Counsel certifies that a response is, after a reasonable inquiry, complete and correct. That certification is about the reasonableness of the search — which is why a methodology that cannot be explained is a problem regardless of whether it was disclosed.

The Sedona Conference's sixth principle points the same way: the responding party is ordinarily best situated to evaluate the procedures and technologies appropriate for producing its own electronically stored information.

So why does nearly every sophisticated case disclose it anyway#

Because the alternative is arguing about it later, from a worse position.

An undisclosed methodology becomes discoverable as a dispute the moment the requesting party suspects a gap. At that point the producing party is defending its process retrospectively, on the requesting party's timeline, often without having preserved the validation record that would settle it. Disclosure negotiated up front converts that from a motion into a term.

The practical answer is that disclosure is cheap and validation is what matters. Courts have shown far more interest in whether a process was reasonable and documented than in which vendor's classifier was used.

What to negotiate instead of arguing about it#

The useful conversation at the Rule 26(f) conference is not "are you using TAR" but what the parties will do about quality:

• Validation, not seed sets. Sampling protocol, the measures reported, and the acceptance threshold. This is the term that actually protects both sides.
• What happens to the null set. Whether a sample of documents the system classified as non-responsive gets reviewed, and by whom.
• Treatment of families and non-text items. Spreadsheets, images, audio and short chat messages behave badly under text classifiers and usually need a separate path.
• What is preserved for later. Which statistics get retained so the process can be defended months afterwards without reconstructing it from memory.

A party asked to hand over seed documents is usually being asked the wrong question. The reasonable version is a validation protocol both sides agreed to before review began.

Where it goes wrong#

The recurring failure is not the technology. It is a review that was never validated, run by a team that cannot describe the sampling, defended by a declaration written after the challenge arrived. That is the fact pattern that produces an adverse ruling — and it produces the same one whether or not the methodology was ever disclosed.