LexisNexis on Wednesday commercially launched Nexis+AI, a platform that uses generative AI capabilities to enhance companies' research, intelligence gathering and business decision-making.
The release comes after LexisNexis introduced a beta version of the service in April.
The commercial version uses over 20,000 licensed media titles from major publishers such as The Associated Press, Gannett and Benzinga.
LexisNexis says Nexis+AI can research and analyze business, financial and legal information and news, summarize documents, create first drafts of intelligence reports and summaries, and derive actionable business insights.
In addition to news data, information and analytics provider Nexis+AI also sourced data from SEC filings and other global registries.
LexisNexis began working on Nexis+AI after publishers criticized and filed lawsuits against several generative AI vendors for not obtaining permission to use their content.
The New York Times, for example, is suing OpenAI and Microsoft for infringing its content.
Nexis+AI comes as other media companies, such as Getty Images, are also developing tools and services using generative AI. Getty previously sued Stable Diffusion's developer, Stability AI, for copyright infringement, alleging that the vendor had violated its copyright by using Getty Images to train its image generation models. A few months later, Getty unveiled its own image generation tool.
Getty's approach to using its vast collection of images and data to train its AI tools is similar to the approach taken by LexisNexis' Nexis+AI.
LexisNexis leverages a variety of data to create GenAI tools for a wide range of users.
So the idea behind Nexis+AI – using an array of licensed data to create generative AI tools for a wide range of users – isn't exactly new, said Michael Bennett, an AI policy adviser at Northeastern University.
The power of AI at scale
But what's novel is the scale at which LexisNexis operates — the range of publishers, content creators and content licenses it offers, Bennett said.
“The fact that they're probably able to put all that knowledge to use suggests that their product is going to be pretty amazing,” Bennett said.
This will enable LexisNexis to offer subscribers access to a wider range of content and eliminate the threat of intellectual property infringement, he said.
“We think we're going to get a lot of benefit from powerful AI being able to understand large, broad, diverse datasets and provide insights, while also having insurance against the future where a content creator or licensee comes up to us and says, 'You're infringing my copyright,'” Bennett continued.
Resolving hallucinations
For legal professionals and researchers, Nexis+AI is a time saver and more mature than earlier versions of the tool released in the nascent stages of generative AI, said Katie Gardner, an intellectual property attorney at Gunderson Dettmer LLP.
Early models were not well-validated and were trained on whatever data they found on the internet.
This has caused problems for some legal professionals, for example, there was a notorious case where a litigator used one of the early tools to create case references for legal briefs and petitions to the courts, cases that were invented by a hallucinated law master.
“My guess is that Lexis has probably tried very hard to resolve that issue,” Gardner said.
Meanwhile, LexisNexis has sought to mitigate and eliminate the illusion with its own guardrails.
Nexis+AI uses two LLMs: Anthropic Claude via AWS Bedrock and the GPT model on Microsoft Azure.
The tool also uses search expansion generation for licensed content.
“This ensures that you're not just asking the LLM and getting the answer from the LLM,” says Snehit Cherian, CTO of Global Nexis Solutions at LexisNexis. “We're basically looking from our content repository at what sections of text within an article make sense for this question, and providing that as context.”
The tool uses a fidelity score to find out which parts of the answers provided are contextually accurate and which parts are inaccurate or fabricated.
We also examine how much context is used to calculate the answer to the question.
This tool searches for the relevance of your answer across two LLMs.
This means that an answer provided by one LLM is passed on to another LLM to see what question the previous LLM answered. If the question is similar to the original question, it is indicated as relevant.
There's also a human in the loop element: LexisNexis has a team that analyzes answers based on statistical sampling and calculates their usefulness, accuracy, reliability, and plausibility.
Finally, citations are provided for each answer, allowing users to verify the answer for themselves.
Due Diligence and Renewal
Despite the potential time-saving benefits and the work LexisNexis has done to reduce hallucinations, Gardner said users should still exercise caution.
“I think the burden is still on the end user to verify accuracy,” she said. “I think if they're relying on software to summarise a primary source rather than directly referencing the source, they need to be aware of the risks involved.”
Another concern for people using the tool is the update rate of the material – how quickly the platform can output the latest information, Bennett said.
“I imagine that because the material is coming from different sources, the speed of updates will vary,” he said.
Bennett said the faster rate of updates will be beneficial for organizations and teams that need to make decisions based on insights gained from the service.
Esther Ajao is a news writer and podcast host at TechTarget Editorial covering artificial intelligence software and systems.
