Which is better for document review, technology-assisted review or generative AI? Neither wins every time. Technology-assisted review (TAR) learns from documents that lawyers have already coded. Generative AI follows written instructions and explains what it finds. Each can cut the number of documents people must read, and each needs validation to be defensible.
Teams adopting generative AI in eDiscovery are now asked to choose between the two, and sometimes to explain the choice to a court or a client. This article defines both approaches, shows how they differ, summarizes what courts have said about TAR and sets out how to document whichever method you use.
Key takeaways
- TAR trains a model on lawyer-coded examples and ranks or predicts the rest of the set.
- Generative AI uses a large language model that follows plain-language instructions and can summarize and explain.
- Courts have accepted TAR for years. They have not held it to a higher standard than keywords or manual review.
- Whatever the method, counsel remains responsible for a reasonable inquiry under Rule 26(g).
- Documented testing, not the brand of tool, is what makes a result defensible.
What is technology-assisted review?
Technology-assisted review is a process in which software uses lawyers' coding decisions on a sample of documents to predict which of the remaining documents are relevant. You may also hear it called predictive coding or computer-assisted review. Reviewers code examples, the system learns from them, and the system scores or ranks the rest. Lawyers then test the results and decide where to stop.
TAR is a mature approach. In Da Silva Moore v. Publicis Groupe (S.D.N.Y. 2012), Magistrate Judge Andrew Peck wrote that computer-assisted review is an acceptable way to search for relevant ESI in appropriate cases. In Rio Tinto v. Vale (S.D.N.Y. March 2, 2015), he wrote that it is "now black letter law" that courts will permit a producing party to use TAR, and that it is inappropriate to hold TAR to a higher standard than keywords or manual review.
How does generative AI differ from TAR?
Generative AI uses a large language model to read each document against instructions written in plain language, such as "Is this document about the 2024 supply contract, and why?" It can label a document, summarize it and point to the passage behind its answer. Because the instructions do the steering, a team may not need a large training round before starting.
That difference shapes the work. The table below summarizes it.
| Technology-assisted review | Generative AI review | |
|---|---|---|
| How it learns | From documents coded by lawyers | From written instructions (prompts) applied by a language model |
| Typical output | A relevance score or ranking | A label, a short explanation and sometimes a summary |
| Setup | Training and sampling rounds | Prompt design and testing on a sample |
| Best fit | Large sets where relevance is the main question | Issue tagging, privilege screening, summaries and quality control |
| Main risk | Poor training examples | Prompts that miss the issue, or inconsistent output |
| Defensibility step | Sample-based testing of results | Sample-based testing of results, plus a record of prompts and versions |
Neither column replaces the other. Many teams use TAR to prioritize a large set and generative AI to tag issues or check privilege in the documents that remain.
What do courts expect from technology-assisted review?
Courts expect a reasonable, documented process. The Rio Tinto opinion approved a TAR protocol that the parties had negotiated, and Judge Peck stressed that his approval did not mean the same protocol would suit every case. He also identified an open question: how transparent and cooperative parties must be about the seed or training sets. He did not decide it because the parties had agreed to disclose the non-privileged documents in the control sets.
Rule 26(g) adds the lawyer's side. A signature on a discovery response certifies that, to the best of the signer's knowledge and after a reasonable inquiry, the response is complete and correct. The Advisory Committee Notes describe that standard as objective. Using software does not remove it.
Generative AI review is newer, so expect more questions about how you tested it. A record that shows your instructions, your samples, your results and the people who made the final calls will answer most of them.
How should you validate and document a review?
Validation is the same idea for both methods: check the machine's work against human judgment. A practical record includes:
- The goal of the review and who set it.
- The method, tool and version, and any instructions or prompts used.
- Sample sets drawn at random and the reviewers who coded them.
- Results of testing, such as how many relevant documents the process found and how many it missed.
- Changes made after testing, with dates.
- The decision on where to stop and who approved it.
Our earlier article on AI document review walks through sampling and oversight in more detail.
Which approach fits your matter?
- Choose TAR when the set is large, the question is mainly relevance and the team is comfortable with training rounds.
- Choose generative AI when you need issue tags, privilege flags, summaries or a second-pass quality check.
- Combine them when you want TAR to prioritize and generative AI to explain.
- Stay with human review for small sets or sensitive documents where automation adds little.
Before you adopt a generative tool, check the ethics side too. ABA Formal Opinion 512 (July 29, 2024) says lawyers must understand a tool's capabilities and limitations, review its output with care and generally obtain informed consent before putting client information into a self-learning tool. The level of review should rise with the stakes, and document review calls for more checking than brainstorming.
How Gemean helps with eDiscovery and AI
Gemean's eDiscovery services in the USA include data hosting, review platforms, managed review and technology-assisted review. The firm also offers PythAI, a generative AI tool built for litigation and investigation document review. According to Gemean, PythAI is managed by its consultants in a secure, private environment and is designed to work alongside human reviewers rather than replace them. To discuss which method suits a matter, contact our team.
FAQs
What is technology-assisted review in eDiscovery?
Technology-assisted review is a process that uses software, trained on lawyers' coding decisions, to predict which documents in a large set are relevant. It is also called predictive coding or computer-assisted review.
Is technology-assisted review accepted by courts?
Yes. In Rio Tinto v. Vale (2015), Judge Peck wrote that it is now black letter law that courts permit a producing party to use TAR, and that TAR should not be held to a higher standard than keywords or manual review. Approval of one protocol does not guarantee approval of another.
What is the difference between TAR and generative AI?
TAR learns from documents that lawyers have coded and then ranks the rest. Generative AI applies written instructions to each document and can explain its answer. Both need testing against human review.
Does using AI change a lawyer's responsibilities?
No. Rule 26(g) still requires a reasonable inquiry, and ABA Formal Opinion 512 says lawyers must understand the tools they use and review the output with care.
Can generative AI replace first-level reviewers?
It can reduce how many documents people read, but people still set the instructions, test the results and make the final decisions. Treat it as a way to focus reviewers, not to remove them.
Sources
- Rio Tinto Plc v. Vale S.A., 14 Civ. 3042 (S.D.N.Y. Mar. 2, 2015), Magistrate Judge Peck, full opinion (Steptoe & Johnson LLP copy): https://www.steptoe.com/a/web/3oS1yD7SDvnQ7xm4REpsZf/3ZFuHj/20150303-rio-tinto-opinion-re-computer-assisted-review.pdf
- Judge Peck Wades Back Into the TAR Pits with Da Silva Moore Revisited, quotations from Rio Tinto and Da Silva Moore v. Publicis Groupe, 287 F.R.D. 182 (S.D.N.Y. 2012) (eDiscovery Daily, CloudNine): https://cloudnine.com/ediscoverydaily/judge-peck-wades-back-into-the-tar-pits-with-da-silva-moore-revisited-ediscovery-case-law/
- Federal Rules of Civil Procedure, Rule 26(g) (Legal Information Institute, Cornell Law School): https://www.law.cornell.edu/rules/frcp/rule_26
- ABA Formal Opinion 512, Generative Artificial Intelligence Tools (American Bar Association, July 29, 2024), as summarized in The Bar Examiner (National Conference of Bar Examiners, Fall 2024): https://thebarexaminer.ncbex.org/article/fall-2024/generative-artificial-intelligence-tools/
- PythAI for Document Review (Gemean): https://gemean.com/pythai-for-document-review/
- Electronic Discovery (Gemean): https://gemean.com/electronic-discovery/
