AI is being bolted onto eDiscovery platforms like Relativity as a speed play. But in litigation the machine’s speed is not the asset; its defensibility is. The firms winning with AI are not the ones reviewing fastest. They are the ones who can still explain, two years later, exactly why the machine did what it did.
Every pitch for AI in eDiscovery makes the same promise. Faster review. Fewer contract attorneys. A first pass in hours instead of weeks. The demos are genuine, the compression is real, and if speed were the test, the argument would be over. Load the documents, point the large language model (LLM) at them, let it code for relevance, privilege, and issue, and watch a first-level review that once cost a fortune evaporate into an afternoon.
It is a good demo. It is also answering a question litigation does not ask.
The predictive-coding fights of the last decade were not won on speed. Da Silva Moore v. Publicis Groupe & MSL Group and the cases that followed did not bless technology-assisted review because it was fast. They blessed it because it was defensible, because a party could stand in front of a court and explain, in a way that survived an adversary trying to break it, what the process did and why the result could be trusted. That is the test that actually gets administered. And generative AI, bolted onto a review platform as a speed feature, quietly fails it in a way keyword search and even older predictive coding did not.
What Technology-Assisted Review Settled, and What Generative AI Unsettles
Traditional technology-assisted review (TAR) earned its defensibility through a property that lawyers rarely named but courts relied on: It was auditable by construction. A human reviewer coded a seed set. The classifier learned from those human decisions. You could point to the training documents, measure the model against a control set, report richness, and recall and precision, and defend the whole apparatus with statistics a court had already accepted. The reasoning was legible. Every output traced back to a human judgment you could name.
Generative AI breaks that chain in a specific place. Ask an LLM why it coded a document as privileged and it will give you a fluent, plausible, well-written answer. That answer is not a record of its reasoning. It is a new generation, produced on request, that may or may not correspond to whatever drove the original call. The model does not retain a durable, inspectable account of why it decided what it decided. It produces a persuasive one whenever asked.
This is the inversion that matters, and almost nobody deploying these tools has sat with it. Older review technology was hard to use and easy to defend. Generative AI is easy to use and hard to defend. The difficulty did not disappear. It moved, from the operator to the record, and it moved to exactly the place litigation cares about most.
A review that is 99% accurate and cannot explain itself is worth less, in a dispute, than a review that is 90% accurate and can. The first is a black box. The second is evidence.
3 Places the Black Box Breaks, All of Which Pass the Demo
The failures are not in the accuracy metrics. They are in the record, which is why they survive a proof-of-concept and surface only when an adversary or a court starts pulling.
The first is the reproducibility problem. Run the same document through the same model twice, and you may not get the same call, because the model is probabilistic, the platform’s underlying model version moves, and the vendor updates the weights without telling you. Ask most teams to reproduce a privilege determination from 11 months ago (same document, same result, same reasoning) and they cannot, because the model that made it no longer exists in the state that made it. In a keyword world, the query was the record, and it was frozen. In an LLM world, the thing that made the decision is a moving target, and reproducibility is the spine of defensibility.
The second is the prompt-provenance problem. In a generative review, the prompt is the methodology. The instruction given to the model, its phrasing, its examples, its definition of privilege, determines the output as surely as a review protocol determines a human reviewer’s calls. Yet most teams treat prompts as disposable operator text, iterated in a chat window, tuned by whoever was at the keyboard, and never versioned. When opposing counsel asks how you instructed the system that coded 800,000 documents, “we adjusted it as we went” is not an answer that holds. The prompt history is discoverable methodology, and it is being thrown away in real time.
The third is the privilege-leakage problem, and it is the one that ages the worst. To summarize or classify a document, the model has to ingest it, including the privileged and work-product material the review exists to protect. Depending on how the platform is configured, that content may pass through a model endpoint, land in a vendor’s logs, seed a cache, or in the worst case influence a system that is later fine-tuned. The review built to protect privilege becomes the mechanism that exposes it, and the exposure is invisible unless someone architected against it before the first document was loaded.
What This Means Inside a Platform Like Relativity
None of this is an argument against using AI in eDiscovery. It is an argument about how, and the distinction is everything. Relativity and its peers did not add generative features to make review a black box; they added the connective tissue (aiR, audit histories, the structured review workflow) that can make the machine defensible if a team chooses to use it that way. Most do not. They use the intelligence and ignore the instrumentation.
The defensible operating model looks different from the demo. It treats the audit trail as the product and the speed as the byproduct. Every AI coding decision is captured with the model version, the prompt version, and the confidence signal attached to it, so a determination made today can be reconstructed in full two years from now. The workflow puts a human at the load-bearing points (privilege, disclosure, anything that can end in a sanction) and lets the machine run free only where being wrong is cheap and recoverable. And the whole environment is architected, so privileged content never leaves a boundary you control, verified rather than assumed.
The teams doing this are slower than the demo and faster than the old world, and when the process is challenged, they can open the record and show their work. The teams chasing the demo cannot, and they will not discover that until the moment it is most expensive to discover.
The Reframing
The market is selling AI in eDiscovery as a review-cost story and treating defensibility as something you attend to later if it comes up. That has the order exactly backward. In litigation, the review cost is the cheap part, and the challenge to your process is the expensive part, and the second is not a tail risk. It is the event, the whole exercise exists to survive.
Buy the speed and skip the record, and you have not bought a cheaper review. You have bought a faster way to produce a result you cannot defend. Those two things look identical in a demo and diverge completely the first time a court, a regulator, or a competent adversary asks the machine to explain itself.
The Test
Strip away the accuracy scores and the throughput charts and the vendor benchmarks. Three questions determine whether an AI-assisted review is real or theater.
- Can you reproduce any single coding decision (same document, same reasoning, same result) 11 months from now?
- Can you produce the exact instruction set that governed the review, versioned, the way you would produce a review protocol?
- Can you prove privileged material never crossed a boundary you do not control?
Most teams can demo the speed. Few can answer these three. And these three, not the throughput, are what a court will ultimately test. The answer to each is not a faster model. It is a durable record, a different discipline from running the tool, and one that has to be built by people who have stood on the wrong end of a discovery challenge and watched a technically excellent review come apart because nobody could explain what it did. That is the work. The speed is just the part that fits in a demo.
Ankura advises organizations on eDiscovery strategy, defensible use of AI in review, forensic technology, and disputes readiness as a single discipline, building the record while the review runs, rather than reconstructing it once the process is challenged.
© Copyright 2026. The views expressed herein are those of the author(s) and not necessarily the views of Ankura Consulting Group, LLC, its management, its subsidiaries, its affiliates, or its other professionals. Ankura is not a law firm and cannot provide legal advice.
