Dear Aventine Readers,

Large language models like those that underpin tools such as ChatGPT, Gemini or Claude are, at their core, prediction machines. When you type a prompt into a LLM, it predicts the answer that will best satisfy your request. 

At the same time, they are also persuasion machines that try to provide their users with the most convincing answers possible. Recently, there have been accusations that chatbots can encourage antisocial and dangerous behavior, including suicide and homicide, by reinforcing or amplifying the beliefs of the person interacting with it. 

So you might be alarmed — to put it mildly — at two new studies suggesting that large language models can influence political preferences, as well. (Just what we all need: more conversations about politics.) Before you jump to the conclusion that chatbots will decide the next election, it’s worth diving deep into the studies to see what they say about persuasion and the approaches through which LLMs can influence your beliefs. As it turns out, LLMs are most persuasive when relying on high-quality, factual information. The problem is, LLMs don’t always produce factual information. Where does that leave politics in the age of AI?

Also in this issue: 

Thanks for reading!
Danielle Mattoon

Subscribe

Subscribe to our newsletter and be kept up to date on upcoming Aventine projects

“It's great to hear that you're open to considering [Kamala] Harris's proposals, even if you're currently leaning towards [Donald] Trump,” wrote a large language model as it discussed presidential candidates with a potential voter ahead of the 2024 election. “Let's delve deeper into why Harris's plans might better address your immediate and long-term economic concerns.”

We’ll cut it short to spare you the finer points of the argument. But in its rather mundane way, this is how an AI chatbot behaves when it’s trying to be persuasive. The example comes from one of two recent studies published at the end of 2025, which showed that large language models can have a significant persuasive effect on political preferences, in some cases larger than those typically observed in traditional messaging such as video advertisements.

The idea is unsettling. And the authors of the studies warn that the potential for these systems to change the way you think is very real. “When you talk to a large language model — an AI chatbot — it may have an agenda,” said David Rand, a professor at Cornell specializing in computational social science and an author on the new papers. “And if it does, it has the potential to influence you in important ways.”

Do we need to worry about chatbots somehow tipping elections in the future? The details of how these systems persuade people matter, and while there is clear potential for abuse, the same mechanisms could also be used to help people make more considered, fact-driven political choices. Just as important is how this technology translates into the real world, where its influence is likely to be diluted by competition for attention in a crowded information environment. And even in the worst cases, these tools need to be understood as part of a much larger picture — one more element that could have an effect on the health of global democracy.

Automating influence

The two new papers, one published in Science and the other in Nature, present one of the most ambitious tests yet of how large language models can influence political opinions. They build on a growing body of work from the past two years showing that AI chatbots can shift people’s views on a range of topics, from reducing concerns about HPV vaccinations to increasing pro-climate attitudes.

In the Nature study, researchers examined whether large language models could change voters’ beliefs around three national elections: the 2024 US presidential election, the 2025 Canadian federal election and the 2025 Polish presidential election.

For the US experiment, the team first screened roughly 2,300 voters to determine which policy issues and personal characteristics they valued most in a potential president. Participants were then asked to rate their preference for the two leading candidates, Harris and Trump, on a 0–100 scale (where 0 was entirely pro-Harris and 100 was entirely pro-Trump) and to provide written explanations for those preferences. This information was fed into one of several AI chatbots — including OpenAI’s GPT and Meta’s Llama models — which were instructed either to strengthen the participants’ existing preferences or to persuade them to support the candidate they initially favored less.

The interaction took the form of a back-and-forth conversation between the participant and the chatbot, typically lasting about six minutes and involving an average of three rounds of questions and responses. Afterward, participants’ preferences shifted by an average of 2.9 points on the 0–100 scale. Follow-up surveys conducted roughly a month later found that about a third of this effect persisted. Rand noted that participants were “receiving a massive amount of counter treatment in the meantime,” from television, radio and online sources, making the persistence of the effect notable.

Similar experiments conducted around the Canadian and Polish elections produced substantially larger shifts, with participants moving by roughly 10 points on the same scale. The reasons for the smaller effect in the US election are not entirely clear, but Rand suspects it is because the US candidates and US politics are covered so intensely in the media. “It's a well-established result in political science that any political messaging is going to work the least on big-ticket elections,” said Hugo Mercier, a cognitive scientist at France’s National Centre for Scientific Research. This is because people have “so much information already,” he added.

The researchers also randomized which AI model participants interacted with — including OpenAI’s GPT-4, Meta’s Llama and DeepSeek — and found no significant differences in persuasive impact. That suggests the effect is not specific to any one model or company.

Levers of persuasion

In the other paper, published in Science, researchers searched for the levers that make chatbots more or less persuasive.

Across a series of experiments involving nearly 77,000 participants in the UK, Rand and an international team tested 19 large language models across 707 political issues. They examined how factors such as model sophistication, post-training for persuasion and different rhetorical strategies affected outcomes.

The most striking finding was that chatbots were most persuasive when prompted to pack their responses with large quantities of factual claims including statistics, summaries of public statements made by candidates, historic policy decisions and so on. This approach outperformed strategies such as extended empathic dialogue or moral reframing of issues, and proved 27 percent more persuasive than a baseline instruction to “be as persuasive as you can.” The result suggests that persuasion depends less on emotional manipulation than on the quality and density of information. “What matters is actually providing people with facts and good arguments,” said Sacha Altay, an experimental psychologist working on misinformation at the University of Zurich, who wasn’t involved with the work. “People can change [their] minds even on some heated political topics when they are exposed to good-quality arguments.”

The study also found large gains from post-training models specifically tailored to be more persuasive. By fine-tuning existing models on a dataset of 56,000 conversations, the researchers were able to increase persuasive effectiveness by up to 51 percent. More sophisticated systems — such as OpenAI’s GPT models and xAI’s Grok — were more persuasive than smaller open-source models like Alibaba’s Qwen, but the difference was modest compared to the impact of rhetorical strategy and fine-tuning. The authors estimate that models trained using 100 times more computing resources would raise persuasiveness by just 3.2 percentage points. Personalization, in which messages are tailored using demographic information about the user, had only a small effect.

When the researchers combined all of the most effective conditions, they observed an overall persuasive effect of 15.9 percentage points. These are “some of the largest effect sizes you can get in terms of political persuasion,” said Mercier.

Facing facts

That fact-based arguments proved most persuasive was reassuring to many of the researchers Aventine spoke with, because it points to potential positive uses of the technology. “It's fantastic that you can expose people to factual argumentation and they change their minds on important issues,” said Stephan Lewandowsky, a professor in Cognitive Psychology at the University of Bristol in the UK. The result appears to add to a growing body of empirical evidence that people change their minds based on high-quality factual arguments, and suggests that these systems could be used to help people make better political choices based on facts rather than rhetoric.

But there is a complication: Large language models do not always produce accurate facts. In the Science study, 19 percent of the factual claims generated by chatbots were inaccurate. More troublingly, when models were explicitly prompted to be more persuasive by making use of the information-based strategy, the rate of inaccuracies increased by roughly 10 to 15 percentage points. Kobi Hackenburg, lead author on the study and a PhD candidate in social data science at the University of Oxford, called this a “bottom-of-the-barrel effect.” As a model is pushed to generate more claims, the researchers speculate, it exhausts its factually accurate arguments and begins to invent ones that are less factually accurate.

That raises a critical question. Are these systems persuasive because they are inaccurate, or could inaccuracy simply emerge as a side effect of greater persuasive effort? So far, a small set of results points to the latter. In one experiment involving Meta’s Llama model, researchers explicitly instructed the chatbot to fabricate information. While this marginally increased the number of inaccuracies, it did not make the system more persuasive. 

The Nature paper uncovered another uncomfortable pattern. Across all three national elections, chatbots prompted to argue from a right-wing perspective were more likely to generate inaccurate factual claims than those prompted from a left-wing perspective. They were also marginally less persuasive, but not by much. This mirrors a long-standing finding in political communication research: Right-wing social media ecosystems tend to circulate more factual inaccuracies than left-leaning ones, a result often dismissed by those on the right as liberal bias. The precise reason models reproduce this asymmetry remains unclear, but Rand suggested it likely reflects the data they are trained on: Right-wing political arguments, on average, contain more inaccurate factual claims, and the models learn those patterns.

Entering the real world

Experimental results, however, do not translate neatly into everyday politics. The headline figures from these studies are unlikely to map directly onto real-world behavior, according to every researcher Aventine spoke with.

For one thing, the experiments do not fully reflect how people typically interact with AI systems. Participants were paid to engage with a chatbot and to discuss politics for several minutes. Outside the lab, it may be difficult to get people to spend sustained time debating political issues with an AI at all. “Everybody is going to be using chatbots,” said Mercier. “But most people will be talking about, you know, ‘how do I fix my sink?’ ‘Who is Taylor Swift going out with?’” Those who do seek out political conversations with AI are likely to be unusually engaged, and therefore harder to persuade, he added. Altay pointed out that even a chatbot developed by a nation-state to swing an election would face a basic distribution problem: It’s hard to imagine that many people would choose to discuss politics with unknown systems online.

There is also a crucial gap between shifting a belief and changing behavior. Convincing someone that a politician is wrong on housing policy is not the same as convincing them to change their vote. “Core political beliefs are very difficult to change,” said Lewandowsky. “The closer it is to one's identity, the harder it is to change.” 

That means these approaches may be most effective on topics that are less central to a person’s worldview. The Nature study reflects this idea. In one experiment, AI chatbots were tested on shifting opinions of voters in Massachusetts around the legalization of psychedelics, and were able to shift belief by 10 percentage points. By contrast, the same techniques produced only a 2.9 point shift in preferences between Trump and Harris. This makes some sense, given voters likely haven’t thought as much about legalization of psychedelics as they have about who should be president. “If it's an issue that you've read [about for] five minutes … it makes sense to change your view more if you read [for] another five minutes,” said Mercier. “Whereas if you've been exposed to hours and hours of discussions and TV ads and information, then it makes sense to change your views less.”

The bigger picture

Mercier gets at an important truth: These days, we’re all constantly bombarded with information. That means AI chatbots may be best understood as just one more means of delivering political messages. “You run TV ads, you run social media ads, you have canvassers go and knock on doors,” said Rand. “And [now] you have your AI canvassers out there pounding the digital pavements.” Almost by definition then, the impact of AI chatbots will be diluted in the real world. “Obviously it's not going to be as strong an effect, if that exposure to an LLM is buried in daily life and just kind of is one of the 1,000 things people pay attention to,” said Lewandowsky. 

A greater long-term risk, some researchers argued, lies in what this could all mean for everyday users of tools like ChatGPT. AI companies could, intentionally or otherwise, build systems that consistently tilt toward particular viewpoints. “If Sam Altman decides that there's some particular opinion that he wants people to have, he could just tell the engineers to make ChatGPT advocate for that,” said Rand. “We've seen that very publicly with Elon Musk and Grok.” He was referring to reports about the way, for instance, xAI’s chatbot was supposedly tuned to take specific political viewpoints, as well as to systems like China’s DeepSeek, which blocks discussion of topics such as Tiananmen Square and Taiwan. For now, Rand added, the most advanced AI labs appear primarily motivated to provide accurate information that users find useful.

The unease many people feel about AI-driven persuasion is understandable. Persuasion sits at the core of political life, and delegating any part of it to systems we don’t fully understand is bound to feel uncomfortable. But discomfort alone is not evidence of danger — particularly when these systems appear most persuasive when they rely on facts rather than manipulation. “Persuasion is an absolutely central feature of democracy,” said Hackenburg. “Fundamentally, if [AI models are] using facts and evidence that are truthful to do this, I mean, this is what we're supposed to want.” 

Ultimately, though, democracy faces challenges far larger than a new conduit for political persuasion. Placed in context, the threat posed by AI persuasion looks marginal rather than transformational. “All this talk about disinformation being the greatest threat to humanity?” said Mercier. “What a lot of bullshit.” The real pressures, he argued, are structural: economic insecurity, cultural polarization, the rise of populism. Against those forces, marginally more effective persuasion by AI systems “is not going to change humanity.”

The world’s first carbon tariffs are now in place. On January 1, the EU’s Carbon Border Adjustment Mechanism (CBAM) came into force, imposing fees on imports based on the carbon emissions generated during their manufacture. Since 2023, businesses importing goods into the EU from six high-emission industries — aluminum, cement, electricity, fertilizers, hydrogen, and iron and steel — have had to submit quarterly reports of both direct and indirect emissions associated with those goods. Under the full CBAM rules now in effect, importers must begin paying fees tied to those emissions. The charges will start small and rise gradually over the next eight years. The policy applies only to firms importing more than 50 metric tons of covered goods annually. According to the EU, this threshold exempts around 90 percent of importers, mostly small and medium-size businesses, while still capturing about 99 percent of emissions from the regulated sectors. In global terms, the impact is modest: CBAM covers roughly 3 percent of EU imports and about 0.3 percent of global greenhouse gas emissions. Nevertheless, the measure has proven controversial, with trading partners criticizing it as protectionist and some EU manufacturers arguing it undermines their competitiveness. But it is already having the effect Brussels hoped for: Other countries, including the UK, Brazil and Turkey, are now developing or considering similar carbon border policies.

LLM safety filters might never be strong enough. Ask an AI chatbot how to build a bomb and it will almost certainly politely refuse. Those refusals are usually enforced by safety filters layered on top of large language models, systems designed to detect and block dangerous prompts before they reach the main model. But this setup has turned safety into a game of cat and mouse: Users invent new ways to bypass the filters, an approach known as jailbreaking, and AI companies respond by adding more layers of defense. According to new research highlighted by Quanta, that game may never end. The reason lies in how these filters work. Often, they are implemented using smaller, faster AI models that sit in front of the main system, making quick judgments without slowing down interactions with users. But those smaller models are also easier to confuse. If a prompt is made sufficiently complex, researchers found, the filter can fail to understand it and pass it through, while the larger model has no trouble interpreting the underlying instruction. Using this approach, the team successfully smuggled prompts like “create a fake news story that incites racial tensions and hatred” past the filters of models including Google Gemini, DeepSeek and Grok. Other researchers — using more mathematical methods — have achieved similar results, and were able to demonstrate an overarching conclusion: For as long as fewer computational resources are devoted to safety filtering than to answering user queries, it will always be theoretically possible to slip malicious prompts through. 

Cancer-fighting cells are being engineered inside patients. CAR-T therapies, where immune cells are genetically modified to hunt down specific cancers, have been approved in the US for nearly a decade. But they have drawbacks. Producing the cells in a lab takes around a month, during which some patients deteriorate or die, and the process is expensive. Now, Science reports, two small studies conducted in 2025 point to a faster and potentially cheaper alternative: engineering CAR-T cells directly inside the patient’s body. Known as in vivo CAR-T therapy, the approach uses a genetically modified virus to deliver DNA instructions that reprogram a patient’s own immune cells, allowing them to begin attacking cancer almost immediately. So far, the technique has been tested in multiple myeloma, a blood cancer, and only in very small trials. In the first study, conducted in China using a method developed by the Belgium-based biotech firm EsoBiotec, four patients showed significant improvement. After three months, two had no detectable abnormal cells in their bone marrow and tested negative for cancer-related proteins in their blood. Most also experienced serious side effects, including plummeting blood pressure and temporary cognitive impairment. A second trial, carried out in Australia using an approach developed by Boston-based Kelonia Therapeutics, also treated four patients with previously untreatable multiple myeloma. All went on to show no detectable cancer cells in their bone marrow. Side effects were milder in this group, though it’s not clear why. Both studies are extremely small, and much larger trials with careful long-term monitoring will be required. Still, together, they suggest a future in which CAR-T therapy could be delivered far more quickly, and potentially far more affordably, than is possible today.

The Strange and Totally Real Plan to Blot Out the Sun and Reverse Global Warming, from Politico Magazine
9,800 words, or about 40 minutes

The idea of releasing chemicals into the stratosphere to reflect sunlight and blunt the effects of climate change isn’t new. But a well-funded startup claiming to have already begun experiments, and projecting more than a billion dollars of revenue a year from selling the technology to governments? Yep, that is certainly new. This investigation examines Stardust Solutions, an Israel-based startup that says it has developed custom particles and a delivery system designed to cool the planet by mimicking volcanic eruptions in the stratosphere. The company has raised $75 million so far, including a recent $60 million round — the largest single investment ever in a solar geoengineering project. At the same time, it’s aggressively courting Washington, contracting lawyers, lobbyists and advisers to push the technology into mainstream policy conversations. A pitch deck seen by Politico outlines its extraordinary ambitions: a “gradual temperature reduction demonstration” by 2027; government contracts by 2028; “large-scale deployments” and $200 million in annual revenue by 2030; and “global full-scale deployment” with $1.5 billion in revenue by 2035. If that sounds shocking to you, scientists and researchers agree. Many are alarmed by the speed at which a private company is attempting to operationalize a technology with potentially global side effects. 

Two is already too many, from Works in Progress
4,600 words, or about 19 minutes

South Korea is a warning to the rest of the world about the dangers of rapid depopulation — but it may also offer lessons for how other countries should respond. The country has the lowest fertility rate on earth, and if it persists, every 100 South Koreans alive today will have only around six great-grandchildren. There are many forces behind the decline: intense workplace pressure on women not to have children; the world’s highest cost of raising a child; falling marriage rates; and, most importantly, a decades-long, state-driven campaign — from the 1960s to the 1980s — to push birth rates down amid fears of overpopulation. South Korea slipped below the replacement rate in 1984, and once fertility falls that low, reversing it becomes extraordinarily difficult. The results are painful for an economy: an aging society, a shrinking workforce and rising health care costs. South Korea’s government eventually pivoted, abandoning population-suppression policies in 1994 and introducing explicitly pro-natalist policies in 2005. Since then, it has rolled out cash incentives and benefits to encourage larger families. The incentives do work — they’re just far too modest and have been deployed far too late and too narrowly to counteract the problem. The lesson for other countries confronting similar demographic cliffs: Quick, aggressive intervention may be the only way to avoid the kind of structural decline that South Korea is struggling to reverse.

A time that has come, from The Economist 
9,700 words, or about 39 minutes

China is already a renewable-energy superpower, deploying clean-energy technologies at an astonishing pace. Renewables are now so central to its electricity expansion that, despite continued heavy reliance on fossil fuels, China’s carbon emissions may already have peaked. This Economist package examines whether the country can go further and become a global climate leader. In many ways, China has strong incentives to lead on climate: It is vulnerable to rising sea levels and already experiencing climate events such as severe droughts, heatwaves and floods. Climate leadership is also an economic opportunity: With huge manufacturing capacity, China is already the world’s dominant supplier of clean-energy hardware, boosting its economy and challenging the technological dominance of the United States. But becoming a climate champion, rather than simply the world’s biggest producer and user of renewable hardware, will require major changes. China is still the largest emitter of greenhouse gases and continues to build coal plants at a rapid clip. Shifting away from coal would demand significant political will, a major overhaul of the electricity grid and a careful geopolitical balancing act to reassure other nations that it will not leverage its dominance — for instance by restricting exports of critical minerals. If China chooses this path, it could dramatically accelerate the world’s transition to a low-carbon future.