By Benjamin Guggenheim

Not a subscriber? Sign up here to get this newsletter in your inbox.

• Exclusive: Senate Commerce Committee leadership under Chair Ted Cruz (R-Texas) has requested that all Republicans on the committee submit their priorities in preparation for a potential vote on an AI package.
• A new study led by researchers from the University of Oxford and the AI Security Institute demonstrates that AI models have become more effective at political persuasion than even the most skilled debaters and canvassers.
• I speak with the lead author of the study, Kobi Hackenburg, about how post-training can make AI models more persuasive — and what it all means for the political influence ecosystem.

This is Benjamin Guggenheim. Welcome to WP Intelligence’s AI & Tech Brief, where we examine the transformative technology of artificial intelligence at the intersection of innovation, policy and power. Get in touch at Benjamin.Guggenheim@washpost.com or on Signal at ben_guggenheim.88.

The Lead Brief

Leadership of the Senate Commerce Committee under Chair Ted Cruz (R-Texas) has asked down-dais Republicans on the committee to submit their priorities on AI policy in preparation for a potential mark-up, according to three people familiar with the plans, who spoke on the condition of anonymity to speak candidly.

A spokesperson for Cruz confirmed the development — though one of the people tells me that Cruz made the request before news reports surfaced about negotiations between Senate Commerce Committee member Marsha Blackburn (R-Tennessee) and the White House over a federal AI framework.

The move signals that Cruz may be looking to plant a flag on Senate Republicans’ views on AI, especially now that multiple bipartisan efforts in the House to create an AI framework have sputtered.

Cruz was a major force pushing for preemption of state AI laws last spring during negotiations over the GOPs “One Big, Beautiful Bill.” But Blackburn blocked the effort over concerns surrounding children’s online safety, among other AI guardrails.

Movement on Blackburn provisions: The Senate Judiciary Committee unanimously passed one of Blackburn’s priorities for an AI package Thursday morning. The bill, called the No Fakes Act, would protect victims of unauthorized “deepfakes” of themselves and their creative work. It now heads to the Senate floor.

Blackburn has also been in conversation with the White House over two key pieces of children’s online safety legislation that the senator is looking to move with a limited preemption framework.

Some sources had anticipated that the White House would endorse the two bills, but things have been quiet since I reported that key administration offices had met last week with children’s safety groups on the matter.

The Throughput Advantage

AI systems are now more persuasive than the most persuasive humans, according to a new study by researchers from the University of Oxford, the AI Security Institute, Stanford University and the London School of Economics and Political Science.

The study pitted AI systems against laypeople, professional canvassers and elite debaters (including world championship debaters) in 18,878 conversations involving a total of 6,923 people in four different experiments.

The research found that frontier models, including Gemini, Grok and ChatGPT, outperformed all of the humans at persuading people to agree with one of 10 prespecified policies in the U.K. This didn’t change even when humans were trained on the AI that beat them or offered cash incentives to win.

Why exactly? The study suggests that it’s because AI is able to produce written content at a much faster pace than humans can — and that AI can pack more fact-checkable claims into each conversation (the AI’s “throughput”). In one part of the experiment that involved real money, AI was much more effective than employees of a U.K. canvassing firm at fundraising for Save the Children.

What it means: The study has significant implications for political communications. But it’s difficult to tell exactly how the findings shift the balance of power in the political influence ecosystem — since only a fraction of political persuasion happens in contexts available to AI (written contexts, texts, etc.), rather than during in-person interactions.

A similar study published last December shed light on what exactly makes AI politically persuasive.

The study, which involved 42,357 people and 19 large language models, found that conducting post-training of models (using proven persuasive responses) was the biggest predictor of model persuasiveness — rather than having the largest or most advanced model.

And again, models were found to be the most persuasive when they used dense answers with a great deal of new facts and information. Personalizing the model to the human, or teaching it rhetorical devices such as moral reframing or storytelling, was less effective, the study found.

The outlook: The findings suggest that firms able to collect and leverage data from voter conversations to train models will be best equipped to take advantage of this.

But it’s also concerning from a public policy perspective. There are few laws and regulations currently on the books that would help inform users as to whether AI models are designed to be persuasive toward a specific end. Additionally, and alarmingly, the researchers also found that the factors that made the models more persuasive also made them less factually accurate.

I had a brief chat Thursday with Hackenburg, the lead author of the study, regarding his thoughts on the implications of the research. Hackenburg is a member of the technical staff of the AI Security Institute but spoke in his capacity as a Phd candidate in social data science at Oxford.

This interview has been edited for brevity and clarity.

Q: You found that information density was a key determinant of persuasiveness. But it sounds like, as models were trained to be more persuasive, they also became less accurate. Can you explain that?

A: The things that made the model more persuasive also made the model less factually accurate.

There’s two possible explanations for why this is the case. One is that it’s a causal explanation — that these untrue facts the model is deploying are more persuasive than the true ones. And this is why, when the model uses untrue facts, it’s more persuasive.

Another explanation is more of a by-product explanation. And it’s just that, as the models learn to pack their conversations with more and more facts, after a certain point, they run out of the good facts and end up scraping the bottom of the barrel of the facts that they have. And so, you see this decline in factual accuracy.

We don’t know which of those explanations really is the case, but we’re looking at that.

Q: Given that the most effective tool to enhance persuasiveness is post-training models, can you explain what that is and how you did it?

A: Broadly, in a very simple way, you can think of pre-training and post-training. Pre-training is the stage where the model is trained on huge chunks of the internet, huge amounts of text data. That’s very, very expensive — and requires lots of data and compute.

Then there’s the post-training phase where the model learns how to engage in a conversation instead of just completing whatever text you give it. It learns to act like a chatbot.

This stage can be done with much less data, basically, and is much more flexible. PhD students can post-ttrain models themselves. OpenAI offers a fine-tuning API where people can upload their data and fine-tune versions, which is what we did.

Q: Obviously, only a fraction of political persuasion is done on the internet. How do you see this kind of model capability manifesting in practice?

A: It’s a really difficult question. I actually don’t really know.

It could be that people find ways to train models that do message testing for them, and so AI systems are used to help make human messages more persuasive. It could be that we see this explosion of these really personalized texting campaigns and texting operations, which do exist.

There’s lots of ways in which chatbots can be sort of integrated into the public sphere, but it’s not clear at all exactly which ways are going to be the ones that people are most keen to talk to or engage with.

This newsletter is published by WP Intelligence, The Washington Post’s subscription service for professionals that provides business, policy and thought leaders with actionable insights. WP Intelligence operates independently from The Washington Post newsroom. Learn more about WP Intelligence.