Lobbying depends on reading proposals, judging who they may affect and making a case to lawmakers. Researchers at CodeX, a legal informatics center at Stanford University, tested whether GPT-3.5 could help with those tasks, from screening U.S. Congressional bills to drafting advocacy letters.
Screening bills for company relevance
The researchers asked GPT-3.5 (text-davinci-003) to assess whether bills could affect selected public companies if enacted. The model received each bill’s title, summary and subjects, along with a company’s name and business description from its 10K SEC filing. When a bill was too long to process, the model summarized it before assessing its relevance.
For each bill, the system returned a yes-or-no answer, an explanation and a confidence estimate. That makes the task more than a simple search for matching terms: the model was asked to connect a proposal’s content with a company’s business.
Across 485 bills, GPT-3.5 reached 75.1 percent accuracy. When it gave itself a confidence score above 90 percent, accuracy among those classifications rose to 79 percent. The researchers also noted that a system that labeled every bill irrelevant would have been accurate 70.9 percent of the time.
That baseline matters when interpreting the headline result. The model performed better than the always-irrelevant approach, but the comparison also shows why accuracy alone does not tell the whole story: many bills in the set were irrelevant to the companies being assessed.
From analysis to advocacy letters
The experiment went beyond deciding which bills deserved attention. For bills it considered important, GPT-3.5 also generated letters addressed to the relevant member of Congress, seeking changes that would favor the company.
In one example, the model wrote on behalf of Alkermes Plc about the Medicare Negotiation and Competitive Licensing Act of 2019. Its letter supported provisions on price negotiations by the Centers for Medicare & Medicaid Services and competitive licensing, then suggested adding incentives for pharmaceutical companies to negotiate.
The generated letter shows how one classification could lead into the next stage of lobbying: once a bill is flagged as relevant, a model can produce an initial argument for a lawmaker. The example also illustrates that the output can take a company’s stated business interests and connect them to the bill’s provisions.
Why the performance gap matters
The researchers compared GPT-3.5 with GPT-3 (text-davinci-002), which achieved 52.2 percent accuracy on the same legal texts. They viewed the difference as a sign that lobbying-related capabilities could improve rapidly as language models improve.
That finding points to a possible practical use: automating parts of the work that involve reviewing bills and preparing routine communications. The researchers suggested this could free lobbyists to spend more time on strategic considerations and lower the cost of lobbying. If costs fell, they said, lobbying could become feasible for nonprofit organizations or individuals as well.
These are possibilities raised by the experiment, rather than evidence that the model can independently manage a lobbying campaign. The test examined relevance judgments and generated letters; the reported results do not establish how lawmakers would respond or whether the proposed advocacy would succeed.
The democratic stakes of AI lobbying
The researchers also identified a risk: an advanced lobbying AI could pursue goals that do not match citizens’ actual preferences. They warned that AI lobbying activities could shift policy discussion in directions that human-driven lobbying might not have pursued.
The concern reaches beyond the quality of an individual letter. The researchers argued that legal systems contain information AI needs to align with society’s needs. If AI systems also significantly influence the law, that process could be compromised, because law is described as the only available democratically legitimate societal-AI alignment process.
The experiment therefore presents AI lobbying as both a productivity question and a question about influence. Models may help classify legislation and prepare advocacy, but the researchers’ warning is that widespread use could affect whose priorities shape policy. The ability to produce a plausible letter is only one part of that larger issue.