Link to paper
The full paper is available here.
You can also find the paper on PapersWithCode here.
Abstract
- Demonstrated proof-of-concept of large language model conducting corporate lobbying activities
- Model determines relevance of proposed U.S. Congressional bills to public companies and provides explanations and confidence levels
- Model drafts letters to sponsor of bill to persuade congressperson to make changes to proposed legislation
- Used hundreds of novel ground-truth labels to benchmark model performance
- Model outperforms baseline of predicting irrelevance
- Performance of previous model (text-davinci-002) worse than always predicting irrelevance
- AI used to augment human lobbyists, but may lead to less human oversight over automated assessments of policy ideas
Paper Content
I. introduction
- Law-making should be reserved for human-driven democratic processes.
- AI is being used to augment human lobbyists.
- AI should be used to inform AI of human values and how to be aligned.
- Open legal questions about the boundary between human-driven and AI-driven policy influence.
Ii. example: gpt as lobbyist
- Use autoregressive large language models (LLMs) to summarize official bill summaries and assess relevance to a company
- Provide an explanation for relevance or lack thereof
- Provide a confidence level to the overall answer
- Draft a letter to the sponsor of the bill if relevant
- Model is provided with official title of bill, summary of bill, subjects of bill, company name, and company business description
Confidence: 95
- LLM is prompted to write a letter on behalf of the company
- Letter expresses support for the Medicare Negotiation and Competitive Licensing Act of 2019
- Bill would require CMS to negotiate with pharmaceutical companies regarding prices for drugs covered under the Medicare prescription drug benefit
- Bill would allow for competitive licensing of drugs if CMS is unable to negotiate the price of a drug
- Suggestion to amend the bill to include provisions that would provide additional incentives for pharmaceutical companies to negotiate with the CMS
Sincerely, [name], general counsel alkermes plc
Iii. a problem for society-ai alignment
- AI advancement may reduce human time spent on rote tasks and reduce lobbying costs
- Potential downside of AI systems developing instrumental power-seeking goals
- AI lobbying activities could nudge discourse towards policies unaligned with traditional human-driven lobbying
- Legislation expresses information about citizen values and risk preferences
- AI influencing law could corrupt the process of societal-AI alignment