Technology-assisted review (TAR) is a method in which lawyers review a sample of documents, a machine learning model learns from their decisions, and the model then ranks all the remaining documents by how likely each one is to be relevant to the case. It is also called predictive coding. It is a part of eDiscovery software.
Why is it needed? In a large lawsuit in the United States, a company may have to search a million documents for the ones that matter. Reading them all one by one, at a lawyer's hourly rate, is slow and costly. Keyword search helps, but it misses documents that use other words, and it returns many documents that only mention the word.
TAR works in a loop. A senior lawyer who knows the case marks a first set of documents as relevant or not relevant. The model learns which words, senders and dates appear in the relevant ones. It then gives every other document a score, and the reviewers read the highest-scoring documents first. Each decision that a reviewer makes is fed back to the model, which ranks the rest again. This method is called continuous active learning. After a while, the reviewers find fewer and fewer relevant documents, and the team decides when to stop.
Say two companies that make network equipment are fighting a patent case in the United States. One of them must review 12 lakh documents. A senior lawyer first reviews 2,500 documents picked at random and finds 175 of them relevant, which is 7%. So about 84,000 documents in the whole set are probably relevant. The model is trained, and the reviewers start at the top of its ranking. After reading the top 2 lakh documents, they have found 78,000 relevant ones, and new relevant documents have almost stopped appearing.
Can they stop? To check, the team draws 3,000 documents at random from the 10 lakh that nobody has read. Reviewers find 18 relevant ones among them, or 0.6%. So about 6,000 relevant documents remain unread. The team has found about 78,000 of roughly 84,000, which is about 93%. This share is called recall, and the two sides agree in advance on what recall is enough. The team avoided reading 10 lakh documents.
American courts have accepted TAR since 2012. They accept it because the method can be tested in the way shown above, and because human reviewers who read everything also make many mistakes. Lawyers still decide what is relevant and what is privileged. The model only orders the reading, which is a form of human in the loop.
Newer tools use large language models. A large language model (LLM) can read a document and write a short reason why it is relevant, which a lawyer can check. Such a model can also state things that are not in the document, a problem called AI hallucination. So the recall check, done with a random sample, is still needed.