Speaking at The Hill’s Nation Summit last week, Florida Republican Kat Cammack claimed that the Chinese government was deploying “digital twins” of every member of Congress. These twins, digital replicas of individual lawmakers, are reportedly built from publicly available data and used to manipulate American politics. Cammack offered no evidence for this allegation, while the broader media largely ignored it.
American academics had already published a proof of concept
Yet, Cammack’s statement is more interesting than critics might recognize. China may or may not have built 535 virtual legislators. But it easily could. More importantly, somebody almost certainly will. Public figures have spent their entire professional lives publishing the training data for their digital replicas. A member of Congress’s votes, speeches, tweets, articles, press releases, donor relationships and committee interventions form a remarkably complete behavioral record. Taken together, this material allows any technically competent actor to not only imitate how a politician speaks, but to model how they are likely to behave.
This isn’t theoretical. In April of last year, researchers from Johns Hopkins University, Nomic AI, and Helivan, two companies that build monitoring and evaluation systems for AI and organize unstructured data for use with LLMs respectively, published a paper entitled “Toward a digital twin of US Congress.” Using a daily updated archive containing every tweet sent by serving members of Congress during their terms, they equipped large language models with politician specific data. The resulting models could produce tweets that were “largely indistinguishable” from the people upon whom they were modeled. More consequentially, the researchers used them to predict roll-call votes and estimate which lawmakers might cross party lines. Where Cammack described a Chinese intelligence program, American academics had already published a proof of concept.
The phrase “digital twin” is misleading. Nobody is attempting to capture and recreate a senator’s soul, assuming one can be located. A useful model need only predict him more accurately than a lobbyist, intelligence officer, or diplomat might be able to. Human political profiling is imperfect; campaigns divide voters into crude demographic categories; intelligence agencies produce psychological assessments of foreign leaders; lobbyists hire former staffers to explain what makes their former employers tick. AI makes the resulting profile interactive and continuously updating. Diplomats once studied their counterparts. Soon, they may rehearse against them.
Before negotiating with an opponent, officials could simulate countless permutations of the same meeting. Does he respond to flattery, or pressure? Is he exhausted by detail, or does he relish an hours-long philosophical digression? Which concessions might he accept, and what causes him to leave the negotiating table altogether?
The same capability could industrialize blackmail. Digital twins would not magically discover a politician’s secrets, but they could help determine which secrets matter. Intelligence services could test which threats produce favorable outcomes. It could simulate how a target’s family, donors, and voters might react to disclosure. Eventually, blackmail could become less like a guessing game at what a target fears, and more like running an online advertising campaign: continuously tested, refined, and deeply personalized. Disinformation campaigns would change as well. Before releasing a fabricated statement, an adversary could simulate how a target would react, which journalists would propagate the narrative, and what counterattacks might be effective in neutralizing a response.
There are limits. Politicians can be inconsistent, events alter incentives, and Trumpian behavior can make prediction seem futile. A machine trained on publicly available data will also reproduce the version of a politician that he and an army of advisors have carefully manufactured. But the same objections apply to every form of analysis. Polls, focus groups, and intelligence profiling are all useful despite their flaws.
Whether Beijing already possesses a virtual Kat Cammack remains a question rather than an established fact. But ignorance is not a defense policy. Writing her claim off as science fiction is dangerous because a political twin does not need to be correct all the time. It only needs to give its operator a slight advantage in predicting the next move of the real person. The first great application of AI in government is unlikely to be in replacing our representatives. It might be in allowing officials to practice on their adversaries.
Comments