A recent trailer for Francis Ford Coppola’s upcoming film “Megalopolis” has made headlines, but not for the reasons it was supposed to.
Intended to reinforce the director’s image as an iconoclast, the trailer cited negative reviews of some of Coppola’s masterpieces, such as “Apocalypse Now” and “The Godfather,” from the time the films were released. The only problem was, the quotes weren’t real. The marketing consultant tasked with finding them had, evidently, generated them using artificial intelligence (AI).
This isn’t the first high-profile case of this particular form of AI — a large-scale language model (LLM) — inventing and misattributing information. We’ve seen lawyers file briefs citing cases that do not exist. These fabrications, or hallucinationscan be written in a fairly authoritative way, in a way that can sound plausible enough for people to accept it without thinking twice. So when it comes to the future of AI-powered search tools, accuracy is everything. That’s why when the Washington Post decided to create such a tool to help users better access its own archives, it turned to Naren Ramakrishnandirector of the Sanghani Center for Artificial Intelligence and Data Analyticsbased at the Virginia Tech Innovation Campus in Alexandria.
Needless to say, the bar for such a tool coming from a source like the Washington Post must be higher – much higher – than it has been for some of these other LLMs.
“If you’re going to use a language model, you have to be sure that the answers you produce are based on actual reporting by The Washington Post,” Ramakrishnan said.
Ramakrishnan had worked with Sam Han, the Washington Post’s head of data and AI, on an earlier project to predict the popularity of stories. They had enlisted a group of Virginia Tech students to try to predict future social unrest by analyzing media coverage. So in the fall of 2023, Han and Ramakrishnan met with the vice president and executive director of Innovation Campus, Lance Collins, and the Post’s new CTO, Vineet Khosla, in Alexandria.
“Last year, looking at the industry and seeing big changes coming with language models and generative AI, we thought this might be another opportunity to work with Virginia Tech,” Han said.
The goal is to combine AI’s ability to synthesize massive amounts of information with The Washington Post’s archives to provide a kind of digital library resource tool that can access every piece of information in the newspaper on any topic in seconds. The model they developed differs from both traditional search engines and the LLMs that have generated so much media coverage in several key ways by using a process called augmented retrieval generation.
A traditional search engine works by indexing all the results it can find for each keyword you give it, then serving up a flood of information, pulled from the web, related to the combination of terms provided. Open LLMs provide answers to questions, but often stray outside the scope of the data they were trained on.
Augmented content generation is a two-step process. First, it searches the available dataset (in this case, the Washington Post archive) for articles related to the query. Then, it extracts information exclusively from those articles, running only that information through a language model to produce a summary for the user.
Although the augmented data generation only draws on the Washington Post archives, it is nonetheless trained on a larger corpus or dataset. This of course presents its own challenges in terms of reliability. But this is where the power of having one of the world’s largest newspapers leading the editorial decision-making process on this project comes into play.
“We had technicians who were in the same testing rooms as our journalists, the subject matter experts,” said Phoebe Connelly, the Washington Post’s editor of AI, innovation and strategy.
Through this collaboration, the tool that The Post and Virginia Tech have developed is designed to avoid the biggest pitfalls that LLMs like ChatGPT have encountered since their release. Needless to say, for an institution like The Post, accuracy couldn’t be more important. For Virginia Tech, the project represents an opportunity to advance the Innovation Campus’ mission of using AI as a force for positive change.
“Building trust in the online news space is a vast area of research,” Ramakrishnan said. “The entire journalistic process needs to be reimagined for the generative AI revolution.”
Two Virginia Tech doctoral students in the Department of Computer ScienceSha Li and Shailik Sarkar worked on the project, then interned at the Washington Post that summer. It was an opportune time for Sarkar, because he had been working on defending against fast-forward injection techniques, or the kinds of user directives that hackers implement to break such tools, instructing them to bypass their original programming. He immediately got to work on the Post’s AI model, which was already under development at the time. Climate Responseswhich was launched this summer.
“Seeing this in real time, in a real-world application, was really interesting,” Sarkar said.
