Legal

eDiscovery.
Defensible at volume.

eDiscovery has used machine learning for well over a decade, and the courts that accept it accept a documented, statistically validated process rather than a clever model. Anything new has to fit that frame or it cannot be used.

Last reviewed 4 September 2026 · Engineering guidance, not legal advice.

WHAT IS ACTUALLY DIFFERENT NOW

Technology-assisted review is established practice. What language models add is not a replacement for it but two capabilities the older approaches lacked: the ability to work from a description of what is responsive rather than only from labelled examples, and the ability to explain why a document was classified as it was in terms a reviewer can evaluate.

The first shortens the cold start. A traditional workflow needs a seed set built by senior reviewers before the classifier is useful. A model that can be given the issues in the matter, in prose, produces a usable first pass immediately — and the seed set then refines rather than bootstraps it.

The second matters more for defensibility. A reviewer disagreeing with a classification can see the reasoning and the passages it rested on, which surfaces systematic misunderstandings early — a whole category of document being misread because of an ambiguous term of art, for instance — instead of after a full pass.

WHAT DOES NOT CHANGE

Statistical validation

Recall and precision estimated on a properly drawn random sample, with confidence intervals, documented before production. This is the evidence that the process was reasonable, and no model quality claim substitutes for it.

A documented protocol

Agreed with the other side where required, and followed. Changing the approach mid-review without documenting it is the thing that gets a process challenged, regardless of the technology involved.

Privilege review stays human

An inadvertent privilege waiver is not a recoverable error. Models can prioritise and flag candidates; the determination stays with a lawyer, every time.

Reproducibility

Model version, prompt, parameters and the classification of every document recorded. If the process is challenged in eighteen months, you must be able to reconstruct exactly what happened.

THE CONFIDENTIALITY POINT

A discovery corpus is the most sensitive collection of documents an organisation will ever assemble in one place: internal communications, draft strategy, privileged material and personal data belonging to people who are not parties. Sending that corpus to a third-party API is a decision that deserves far more scrutiny than it usually receives, and in several jurisdictions it raises data protection questions of its own.

Self-hosted open-weight models generally handle classification and prioritisation of this kind comfortably. Running the review inside a perimeter you control removes the transfer question entirely and makes the arrangement far easier to explain to a court, a client and a data protection authority.

HOW THE PROCESS STAYS DEFENSIBLE

REVIEW PROTOCOL

Collection

inside your perimeter
Seed and describe

issues stated in prose
Classify

with reasoning recorded
Validate

random sample, stated intervals
Document

protocol, version, parameters
Privilege

determined by a lawyer, always

Solid: the validated review path Dashed: the determination that never leaves a lawyer

What a court accepts is a documented, statistically validated process. The model is one component inside it, not a substitute for it.

QUESTIONS

QUESTIONS — 3

In the US and England, courts have accepted technology-assisted review for over a decade where the process was documented and statistically validated. The technology is not what is accepted or rejected — the process is. Whether a given protocol will be accepted in your forum is a question for counsel, and many civil-law jurisdictions have no equivalent disclosure regime at all. Build the validation properly and this is a continuation of established practice rather than a novelty.

It can prioritise likely-privileged documents so lawyers see them first, which is genuinely useful. It cannot make the determination. The consequences of a false negative are severe and not undoable, and no accuracy figure justifies removing the human here.

Handled better than by older approaches, which frequently needed separate models or translation per language. Validate per language rather than in aggregate — an overall recall figure can conceal a weak language entirely.

Keep reading

Evaluating AI Systems

→

Infrastructure

Air-Gapped Deployment

→

30-minute technical call · no deck