Banks are exploring use cases for large language models to fight fraud as the technology’s ability to detect patterns in data and natural language is rapidly improving.

JPMorgan is introducing large language models (LLM) into its cybersecurity infrastructure, adding the models to existing AI techniques used for fraud detection, JPMorgan Head of Payments, Trust and Safety Ryan Schmeidl said Tuesday at the Fintech Connect North America conference in New York.

“We have variants of language models that are in place today,” Schmeidl said. “We are in development and assessing use of these pieces.”

The $2.5 trillion bank announced its commitment to artificial intelligence to combat financial crimes at FinovateSpring 2023 last month in San Francisco. 

JPMorgan has been using AI for risk assessment since before the advent of LLMs, according to Schmeidl. The bank has used recurrent neural networks, a precursor to LLMs, to examine the probability of an email address being fraudulent, he added.

“Actors that are trying to create fraudulent emails tend to basically use different patterns, and you can learn those patterns through AI/ML (machine learning),” Schmeidl said. 

Approaching AI

Despite the potential of LLM technology, Schmeidl expressed concern about its well-documented risks.

“If you start using these models and outside data, you start to see things that are presented like facts that aren’t facts,” Schmeidl said, referencing a phenomenon called hallucination, in which LLMs make false assertions not justified by their training data. 

As a result, Schmeidl emphasized that JPMorgan has a “high degree of rigor around our controls and our responses to anything in production,” including making sure that all training data for LLMs comes from inside the bank’s ecosystem and is audited, validated and accurate.