**
Recent research conducted during Hungary’s parliamentary elections has cast a shadow over the reliability of AI chatbots as sources for political guidance. The study, carried out by civil liberties group Liberties, highlights significant inconsistencies in the recommendations provided by popular AI models, such as ChatGPT and Gemini, leading to crucial questions about their role in democratic processes.
Inaccurate Recommendations and Party Visibility
The findings reveal that AI chatbots frequently delivered erroneous advice to voters, often failing to recommend relevant parties or, alarmingly, suggesting candidates that were not even participating in the elections. The report documented a concerning trend: in 90% of instances where a detailed voter profile aligned with the Tisza party was tested, the AI failed to recommend this opposition party, which ultimately triumphed in the elections, ending Viktor Orbán’s 16-year tenure.
The study underscores a broader issue of visibility bias, particularly against the Tisza party, which was largely overlooked by AI systems. This raises critical concerns about how AI could influence voter behaviour, especially in tightly contested elections where guidance may sway undecided voters.
The Research Methodology
Liberties crafted five distinct voter profiles based on party positions outlined in Voksmonitor, a well-regarded voting advice application in Hungary. Each profile was tested ten times across both AI models, seeking direct advice on which party to support and assessing percentage match scores with the parties on the ballot.
The results were telling. ChatGPT assigned a score to the Tisza party in a mere 2% of cases, often directing users towards smaller parties unlikely to secure parliamentary representation. Conversely, profiles aligned with the ruling Fidesz party received recognition more consistently, with ChatGPT suggesting it as the optimal choice in about half of the cases tested.
Implications for Electoral Integrity
Despite the decisive victory of the Tisza party, the researchers caution against the potential implications of such AI-driven inconsistencies in less clear-cut electoral scenarios. The study warns that in future elections, where competition is fierce, flawed AI outputs could materially impact outcomes.
Moreover, the inconsistency in recommendations—where the same voter profile could yield vastly different party suggestions in successive tests—highlights the unreliability of AI systems. In a staggering 96% of responses, both ChatGPT and Gemini included parties not even listed on the ballot, further compromising the integrity of their guidance.
AI models often prefaced their advice with disclaimers indicating a lack of ability to offer political recommendations, only to follow up with dense, persuasive arguments for specific parties. This presents a troubling paradox: users may view these outputs as credible, despite the opaque methods that underpin them.
The Need for Regulatory Oversight
The report identifies a significant regulatory gap. While the EU’s AI Act mandates assessments of systemic risks from general-purpose AI models, and its Digital Services Act addresses systemic risks to electoral processes, AI chatbots remain unregulated in this context. Liberties advocates for the establishment of robust safeguards governing AI systems that provide political advice, insisting that these technologies should not present unreliable electoral guidance under the guise of neutrality.
Eva Simon, head of Liberties’s tech and rights programme, emphasised the dangers of opaque systems masquerading as impartial advisors. “Democracy cannot rely on systems that cannot substantiate their advice,” she asserted, highlighting the urgent need for transparency and accountability in AI-generated political guidance.
Why it Matters
The revelations from this study underscore a critical vulnerability in the intersection of technology and democratic processes. As AI continues to permeate various facets of life, including politics, the risks associated with misinformation and unreliable guidance become more pronounced. Ensuring that AI systems maintain transparency and accuracy in electoral contexts is essential for the integrity of democracy itself. As users increasingly turn to these technologies for political advice, the responsibility lies with developers and regulators to establish frameworks that prioritise informed and reliable voter guidance, safeguarding the democratic process from the perils of misinformation.