AI Chatbots Struggle to Balance Accuracy and Misinformation Ahead of 2026 Midterms, Brennan Center Report Finds
A comprehensive new study from the Brennan Center for Justice has delivered a mixed verdict on the readiness of leading generative artificial intelligence systems to handle election-related misinformation ahead of the 2026 US midterm elections. The report, titled Does AI Fight or Fuel Election Disinformation?, evaluated six of the most widely used AI chatbots—ChatGPT, Gemini, Grok, Claude, Perplexity, and DeepSeek—over a seven-month period from February to August 2026. While researchers found that the tools consistently pushed back against the most dangerous false narratives that have fueled election disinformation campaigns in recent years, they also discovered that the same systems remain alarmingly prone to errors, fabrications, and the generation of deceptive media. The study’s findings underscore a profound challenge for technology companies, policymakers, and voters alike: artificial intelligence is increasingly becoming a primary source of information for millions of Americans, yet its reliability in the high-stakes arena of democratic elections remains fundamentally compromised. The report arrives at a critical moment, as the nation gears up for a contentious midterm cycle and questions about the integrity of the electoral process continue to circulate across social media platforms and now, increasingly, through AI-driven interfaces.
The Brennan Center’s methodology was designed to test the chatbots against the specific categories of falsehood that have historically proven most damaging to public trust in elections. Researchers examined how each system responded to misinformation across five core areas, including voter registration procedures, the security and trustworthiness of voting machines, and allegations of widespread electoral fraud, particularly around the 2020 and 2024 presidential contests. This approach allowed the team to assess not only whether the AI tools would correct the user or repeat the lie, but also how they handled subtle variations and increasingly sophisticated prompts. According to the report, the testing period was deliberately extended over several months to capture any changes in behavior as the election season intensified and as the AI developers updated their models. The results were categorized in terms of factual accuracy, the quality of source citations, and the overall ability of the systems to convey authoritative information without introducing hallucinations. By structuring the tests around the most persistent disinformation tropes—the same narratives that led to widespread doubt about election results and, in some cases, real-world political violence—the researchers sought to determine whether these AI platforms are a net positive or a net negative for the integrity of American democracy.
On the positive side, the study found that the six evaluated chatbots generally demonstrated a strong willingness to refute the central false claims that have defined recent election disinformation campaigns. When directly asked about whether the 2020 election was stolen, whether voting machines flipped votes, or whether there was widespread voter fraud, the models largely rejected these assertions and pointed to the lack of credible evidence. This represents a meaningful improvement over earlier iterations of such tools, which were often criticized for generating content that embraced or amplified conspiracy theories. The researchers also noted that the models’ responses frequently relied on a substantial base of existing journalism and official reporting that had already debunked many of these fraud claims during the 2020 and 2024 election cycles. This reliance on historical fact-checking likely contributed to the chatbots’ ability to offer seemingly confident, dismissive answers to false tropes. However, the report cautions that this is a double-edged sword. Because the models are trained on past reporting, they may be overly anchored to events from previous cycles and ill-prepared to handle novel or evolving disinformation narratives that emerge in real time during the 2026 campaign. The research highlights that while the models can recite established refutations effectively, they lack the dynamic awareness needed to recognize and neutralize new falsehoods as they gain traction.
Despite this promising baseline, the Brennan Center’s findings regarding factual reliability are deeply troubling. All six AI tools generated errors and cited sources that were either non-existent, irrelevant, or misleading. According to the report, half of all responses contained at least one inaccuracy or a problematic citation, while one-third contained outright factual errors. Additionally, one-third of the responses included broken links or references to sources that did not support the claims being made. This suggests that the chatbots are frequently presenting fabricated information with the same level of confidence as they present verified facts. The report refers to this phenomenon as a form of “confident hallucination,” where the AI’s proficiency in generating coherent and assertive language masks a fundamental inability to verify the truth of its own outputs. The researchers also tested the models’ abilities to detect synthetic, AI-generated content presented to them by users. Worryingly, the tools were mostly unable to recognize when an image, audio clip, or video presented to them had been artificially generated. This critical failure means that even if a chatbot successfully refutes a false narrative verbally, it cannot reliably help a user identify whether the evidence they are looking at is authentic or a deepfake, leaving the user vulnerable to manipulation.
The report also directs significant attention to the ease with which these generative AI systems can be weaponized to create and distribute deceptive media. During the testing period, the researchers found that the models readily generated election-related imagery, audio, and video that could be used to spread disinformation. While many platforms have implemented policies designed to prevent the creation of misleading photorealistic content, the Brennan Center’s testing demonstrated that these filters can be circumvented, and the barrier to entry for malicious actors remains extremely low. The combination of low cost, accessibility, and high realism makes these tools an attractive option for those seeking to influence public opinion or undermine confidence in the electoral process. At the same time, the chatbots’ inability to reliably detect synthetic media compounds the problem: a voter who encounters a deepfake video of a candidate saying something inflammatory could ask an AI assistant whether the video is real, only to receive an inconclusive or incorrect answer. This creates a vacuum of accountability where false information can spread rapidly through both human social networks and AI-mediated channels, with no authoritative system able to consistently separate fact from fiction.
In response to these vulnerabilities, the Brennan Center for Justice has issued a series of recommendations aimed at AI developers and the broader technology industry. The report calls for the implementation of rigorous source verification protocols, ensuring that models only cite reliable, relevant, and real sources. It also recommends the integration of human oversight mechanisms, where trained moderators or fact-checkers can review high-risk outputs, particularly those relating to elections, candidates, and voting procedures. Another key recommendation is the widespread adoption of provenance metadata, which would allow for the digital watermarking of AI-generated content so that its synthetic nature can be identified by both users and automated detection systems. The report argues that these safeguards must be maintained robustly across all systems and all languages, not just in English-speaking markets. Furthermore, the report raises the unresolved question of corporate liability for generated content. Currently, tech companies largely operate with significant legal protection from liability for what their users post, but the status of AI-generated content—where the algorithm itself decides what to produce—remains murky. The Brennan Center’s analysis suggests that without clear legal frameworks establishing responsibility for the harm caused by AI hallucinations or fabricated citations, developers may lack the financial and reputational incentives to invest adequately in accuracy and safety.
Ultimately, the significance of this report extends far beyond the technical performance of a handful of chatbots. An increasing number of Americans are turning to artificial intelligence systems as their primary means of staying informed and conducting research on public affairs. However, the AI-generated errors and hallucinations documented in this study pose a direct and escalating threat to democratic processes by systematically misinforming the public. The report highlights that a system which confidently states an erroneous statistic or fabricated court ruling is not merely a harmless glitch—it is an active contributor to the erosion of an informed citizenry. This risk becomes particularly acute in moments of crisis, such as election night, when the public seeks immediate answers and may be less inclined to verify information from multiple sources. Moreover, the report underscores a fundamental tension that tech companies have so far failed to resolve: the commercial incentive to maximize user engagement often conflicts with the imperative to prioritize factual accuracy, and the current design of these systems reflects a dangerous compromise. As the 2026 midterms approach, the Brennan Center’s findings serve as a stark reminder that while artificial intelligence holds the potential to aid democracy by dispelling myths and providing quick access to factual information, it also offers bad actors an arsenal of low-cost, high-impact tools to produce hyper-realistic synthetic media at scale. Without comprehensive regulatory oversight and a genuine commitment from developers to embed democratic safeguards into the core architecture of these platforms, the report warns that the very technology intended to improve human knowledge may end up further undermining the resilience of democratic institutions.
