AI Goes to K Street: ChatGPT Turns Lobbyist

Concerns around how professional lobbyists distort the political process are nothing new. But new evidence suggests their efforts could soon be turbocharged by increasingly powerful language AI. A proof of concept from a Stanford University researcher shows that the technology behind Internet sensation ChatGPT could help automate efforts to influence politicians.

Political lobbyists spend a lot of time scouring draft bills to assess if they’re pertinent to their clients’ objectives, and then drafting talking points for speeches, media campaigns, and letters to Congress designed to influence the direction of the legislation. Given recent breakthroughs in the ability of AI-powered services like ChatGPT to analyze and generate text, John Nay, a fellow at the Stanford Center for Legal Informatics, wanted to investigate whether these models could take over some of that work.

In a matter of days, he was able to piece together a rudimentary AI lobbyist using OpenAI’s GPT-3 large language model (LLM), which is the brains behind ChatGPT. In a paper published on the arXiv preprint server, he showed that the model was able to predict 75 percent of the time whether a summary of a U.S. congressional bill was relevant to a specific company. What’s more, the AI was able to then draft a letter to the bill’s sponsor arguing for changes to the legislation.

“The lawmaking process is not ready for this,” says Nay. “This was just a simple proof of concept built over a few days. With more resources and more time spent on this, especially with more focus on building out the workflow and a user experience tied in with the day-to-day of human lobbyists, this could likely be built into something relatively sophisticated.”

Nay’s approach involved feeding the model with text prompts via OpenAI’s API. He provided the model with the title of the bill, a summary of the bill, the subjects of the bill as determined by the Congressional Research Service, the name of the concerned company and the business description the firm filed with the U.S. Securities and Exchange Commission.

Alongside this, he told the model to imagine that it was a lobbyist and use the provided information to work out if the bill was relevant to the company. The model was also asked to explain its reasoning and rate how confident it was in its decision out of 100. For bills deemed relevant, the model was then prompted to draft a letter to their sponsor(s) persuading them to make additions favorable to the company in question.

Nay collated data on 335 bills and then challenged the model to predict whether they were relevant to 121 unique companies. Because most legislation doesn’t affect most companies, he found that you could guess right 70.9 percent of the time by just always saying no. When he tested the approach on an older version of GPT-3 released in March 2022, it did considerably worse than that, managing a prediction accuracy of only 52.2 percent. But when tried out on the GPT-3.5 model, which powers ChatGPT and was only made public in November 2022, it achieved an accuracy of 75.1 percent. On bills where the confidence score was over 90, the accuracy rose to 79 percent.

The paper doesn’t attempt to assess how effective the drafted letters would be at influencing policy, and Nay makes clear the approach is still nowhere near being able to do the bulk of a lobbyist’s job. But he says the significant boost in prediction performance seen between models released just months apart is noteworthy. “There is a clear trend of quickly increasing capabilities,” he says.

That could potentially spell serious trouble for the legislative process, according to Nay. It could make mass influence campaigns significantly easier, particularly at the local level, and could lead to a flood of letter writing that could overwhelm already thin-stretched congressional offices or distort their perception of public opinion. Laws could also be useful resource for helping future AI systems understand the values and goals of human society, says Nay, but not if AI lobbyists are distorting how those laws are made.

It’s impossible to predict how quickly AI will become sophisticated enough to effectively influence the lobbying process, says renowned security expert Bruce Schneier. He notes that letter writing is probably not the bottleneck here, and in fact learning how to understand political networks and develop strategies to influence them will probably be more important skills. But he says this research is a sign of things to come. “It’s just a baby step in that direction, but I think it is the direction that society is going,” he says.

And political lobbying is only one way in which AI is likely to warp society in the future, he adds. In a new book out next week called A Hacker’s Mind, Schneier outlines how the powerful hack everything from the legislative process to the tax code and market dynamics, and he says AI is likely to turbocharge these efforts. “There are a lot of possibilities and we really are just beginning to scratch them,” he says. “The law is completely not ready for this.”

31 Jan. 12:15 p.m. ET Update: A previous version of this story misstated the number of unique bills and companies in the study and mischaracterized the context behind the statistical terms relevancy, confidence score, and accuracy. This version corrects the errors.