You (Probably) Don’t Know How Pressure Politics Work
Capitol Hill Climbing, Part 3
For the third and final piece in this series on how tech firms actually influence AI policy, let’s talk brass tacks — maneuvers — strategy.
The midterms are fast approaching, and interest is surging in the rivalry between wealthy “pro-innovation” and “pro-safety” interest groups in AI policy. Typically, this is framed as a spending battle, where total money flowing into congressional campaigns is a proxy for influence.
Whilst this framework is intuitive and supports a compelling adversarial framing, the literature on special interest group tactics does not support it. Hard as it is to believe, there is no clear evidence that the more special interests spend on an issue, the more likely they are to succeed — vote-buying is not the paradigm of pressure politics. So… what is? And what might debunking some misconceptions tell us about how special interests will shape AI policy come November?
Pressure politics are (mostly) not a game of persuasion or exchange.
There are two historical models that have sought to explain the relationship between special interest groups and policymakers: the exchange model and the persuasion model.
Under the model of exchange, interest groups and legislators horse-trade campaign contributions for votes, with groups making down-payments on legislators during electoral cycles and providing additional contributions once the legislator proves loyal. Yet the empirical evidence that this approach occurs in practice is negligible. Baumgartner et al. ‘s 2009 book Lobbying and Policy Change: Who Wins, Who Loses, and Why provided clear evidence that there was no meaningful relationship between the amount of money spent and the likelihood of lobbying success — vote-buying was not the answer.
Under the model of persuasion, interest groups have a comparative advantage over policymakers in generating relevant information about constituents’ views and policy outcomes. They bide their time and transmit this information to legislators when it’s mutually beneficial. However, if this were the case, pressure groups would focus their efforts on persuading legislators on the fence, as providing their information here would have the greatest marginal impact. In practice, however, lobbyists “concentrate on their allies, avoid their enemies, and lobby undecideds infrequently” — imperfect information at strategic times was also not the answer.
Finally, in 2006, Hall and Deardorff came to spread the good word — lobbying functions primarily as a legislative subsidy, an overall grant to policymakers who are already very closely aligned in their ideology which allows them to devote their internal resources to other policies they find important. Because both actors are ethically aligned, legislators don’t expend resources on verifying this information, as this would render the subsidy pointless. Instead, they adopt these resources wholesale, especially if they are highly technical pieces of information that congressional staff simply don’t have the expertise to draft themselves. This model explains, for example, why lobbyists focus their efforts on allies and end up writing entire sections of bills, much to the chagrin of the public.
What does this mean for AI lobbying in the upcoming cycle? Expect some of the text you see in bills to closely align with the public education efforts of organisations like Build American AI on the innovation side and Americans for Responsible Innovation on the safety side. Policymakers in aggregate create an equilibrium, but individual legislators choose the grant of one side or the other.
Whilst this explains in-house lobbying by single-issue groups, we also need to account for how professional lobbyists hired by pro-innovation or safety SIGs are likely to strategise.
When do lobbyists succeed or fail?
Imagine you’re a down-on-your-luck K Street advocate looking for a gimme. You want to find the easiest policy on which a client is looking to lobby, and you’d like to achieve close to 100% of your agenda, and bring new clients flocking to your door. How would you find an easy issue on which to succeed? Thanks to Mahoney, we can select based on a few points —
The larger an issue’s scope, the less likely you are to succeed.
This one is simple: play small ball. In the US, lobbyists report fully attaining their project goals 60% of the time on niche issues, 33% of the time on large sectors and a mere 7% of the time on system-wide issues.
The more voices speaking about an issue, the smaller your own.
Mahoney finds that in the US (albeit across only 65 cases), the percentage of lobbyists who reported fully attaining their goals on a particular policy drops from 40% for policies with 0 stories in the press to 0% for those with 51 or more.
Moreover, you’d rather pick an issue that’s as uncontroversial as possible. 30% of advocates fail completely to achieve their objectives on issues in which they’re the only lobbying voice, but 70% fall flat on their face if there are two conflicting camps.
You’d rather fight for the status quo than for change.
In the US, policies are like New Year’s resolutions: everyone makes them, but very few follow through. Getting a policy off Capitol Hill and onto the President’s desk requires the will and political capital to push proposals through the gauntlet of subcommittees, committees, and floor votes, all of which may strike them down. Congress is an oil tanker, and incumbent policies are highly sticky. 81% of US advocates lobbying for the status quo reported some level of success, compared with only 41% for those lobbying for change.
So what, come the midterms?
So, in sum, if you want the highest chance to influence Congress, you want to find an issue that is small in scope, uncontroversial, has little public interest and doesn’t disrupt the status quo. Artificial intelligence is… none of those things, meaning that both the safety and innovation camps are going to have a difficult time achieving the majority of their goals in Congress, even with deep pockets. Nonetheless, the fact that innovation groups are arguing for the status quo of a light-touch regulatory environment, one which the White House will continue to endorse until 2028, is likely to be a significant advantage.
The takeaway then is that the overall strategies of these organisations and their bank balance matter less than the number of congresspeople who are willing to accept their technical information about AI without further verification. For lobbyists, it is not about the overall impact of the dollars you spend. It is about which individual is in each and every seat. Unfortunately, I’ve done the annoying thing where I’ve told you our current predictive framework, captured in tools like Transformer’s AI Campaign Finance Tracker, is not adequate, but I’ve failed to provide a new one. The uncomfortable truth is that influence on AI policy is irreducibly local — one ought to track it seat by seat, not dollar by dollar. The closest you might be able to get is watching polls closely to see if AI and automation become more salient with the general public or following CSPAN religiously to identify lobbyist talking points repeated by congresspeople on niche issues related to AI.
What I do believe the literature implies is that individual citizens have more influence over policy than they may initially believe; your time really is best spent on compelling your local officials to verify information received as a grant.
This is likely my last piece for the Alignment Desk. Whilst none of them are perfect, I’ve enjoyed the publish-or-perish philosophy, and would recommend anyone poodling around Cambridge to give up a few Saturdays to complete a few pieces. It’s worth it!