The most viral story in the legal industry in 2021 was the attorney who assured a judge that he was not, in fact, a cat.
The year 2023 may have provided this definitely human attorney’s most potent challenger in the form of the Mata v. Avianca. Plaintiff’s counsel utilized ChatGPT to draft a court filed document, ChatGPT “hallucinated” and provided several fictitious case citations. After attempting to cover up the use of Chat GPT, the attorney was sanctioned, and the case directly led to multiple bans on the use of generative AI from judges across the country.
But what are hallucinations, anyway? And can anything be done about them?What are AI Hallucinations?
Hallucinations are instances in which a generative AI tool (such as Chat GPT, Gemini, Claude, or others) fabricates information.
The technology that underpins generative AI is called large language models, or LLMs. LLMs analyze billions of inputs and then construct sentences based on a probability analysis – that is, based on the previous word, what is the most likely word to follow?
Hallucinations can happen because of a few factors, such as:
- Biased Training Data: If the software is trained on inaccurate or incomplete data, it can internalize and amplify those biases, generating false connections and conclusions.
- Overfitting: When an AI gets too fixated on specific patterns in its training data, it can mistake those patterns for universal truths, leading to inaccurate generalizations in new situations.
- Lack of Context: AI thrives on context. Without a proper volume of inputs, and proper training of the model, a generative AI-enabled product can misinterpret information and provide erroneous responses to user prompts.
Are Hallucinations a Major Concern in eDiscovery?
Multiple workflows and use cases for generative AI in eDiscovery will develop over the coming months and years. In the short run, it appears the two areas most likely to gain traction will be fact finding (did Custodian X have any knowledge of fraud in Company Y?”) and document review (“e.g. find all responsive documents).
Hallucinations are more likely in fact finding due to the open-ended nature of the inquiries and potential of vague language. For example, most custodians won’t write in an email “on February 1st I became aware that Larry was committing fraud.” When ambiguity is present and interpretation is required, that may cause the software to draw incomplete or incorrect conclusions.
With respect to document review, Cristin Traylor, the Director of AI Transformation & Law Firm Strategy at Relativity, gave an interview to Legaltech news in which she made some excellent points:
We can show [during a demo or active matter] the citations, the rationale, the considerations, which again, makes it less of that sort of black box that you think about with [Technology Assisted Review]. You can actually point people to what’s in the document, and why the product is coming up with the prediction that it is, which is a little bit different, I think, than TAR.
We agree, Cristin!
Trust But Verify
Hallucinations are another example in the long list of the promises and perils new technology offers. Technologists are always excited about new tools, but they need to remember that mitigating AI hallucinations is a technical challenge for them, eliminating them is an ethical imperative for attorneys.
By acknowledging the risks, implementing safeguards, and promoting responsible AI development, we can harness the power of AI while preserving the integrity of the legal system.
Want to read more? Check out our eBook "AI in Document Review. Skeptical? We are Too."that addresses the top concerns about CAL and Generative AI in document review.
