Frequently Asked Questions
There were some scary stories about people who used public AI tools that used their data for training. At eDiscovery AI, we only use private LLMs and would never use client data for training models. Additionally, we never send data to any third parties so we can ensure your data is secure.
Absolutely. Lawyers have been using computers to make document classifications for over a decade. It has been used on thousands of cases without a single successful challenge.
Read our blog post to see an in-depth discussion about defensibility.
For more fun, read our other blog post that walks you through the steps of a defensible review.
The reason why AI review is going to take over discovery is because it is significantly better than humans in just about every aspect of document classification.
It is far more accurate than human review. We are seeing AI identify over 90% of relevant documents in nearly all cases.
It is far less expensive than human review.
It is significantly faster than human review.
We are compatible with Relativity One and Relativity Server.
It is incredible.
AI doesn’t care what language the documents are in. It will read the documents, provide you with accurate classifications based on instructions in any language, and will provide summaries and explanations in any language.
With eDiscovery AI, we can summarize and classify audio files just as we do with text.
With eDiscovery AI, we can summarize and classify image files just as we do with text.
The difficulty with short messages are usually caused by spelling errors and the use of slang or abbreviations.
These are things that AI excels at. As long as a reasonable person would understand what is being said, AI will be able to properly classify the short messages.
We can usually get over 95% recall on most issues.
While 90%+ precision is not difficult, we have been seeing in most real cases that people prefer to lean towards over-inclusiveness so in real-life matters it’s more common to see precision in the 70s or 80s.
It is true that AI can hallucinate from time to time, but with proper validation this is not an issue.
Hallucination is most commonly an issue when you are generating content with AI. It is not much of a problem with classification. The result would be merely an incorrect classification.
So if AI makes a categorization error 5% of the time, but a human makes a categorization error 25% of the time, would you tell me you are concerned about humans hallucinating?
This is why we TRUST BUT VERIFY. Any time you are using a computer to assist your review, you must properly validate the results. As we have worked on over a thousand cases using computer assisted review, we know exactly what process we need to do to ensure a defensible result.
If you want to see what steps are required to make a review defensible, please read our blog about the steps to a defensible review.
Unlike traditional predictive coding tools, AI does an incredible job understanding context of language. When someone replies with “I agree” the AI understand what they are agreeing with. It also understands sentiment and tone of the speaker.
Our favorite workflow is a combination of AI review with Continuous Active Learning.
Stay tuned for a case study explaining the full process!