**
Recent research has uncovered significant flaws in the guidance provided by AI chatbots to voters, particularly during Hungary’s parliamentary elections earlier this year. The findings raise grave concerns about the reliability of general-purpose AI systems when it comes to political advice, as these tools often misidentify parties, fail to mention relevant options, and deliver inconsistent responses.
Flawed Recommendations in Election Contexts
A study conducted by the civil liberties organisation Liberties scrutinised the performance of AI models, specifically ChatGPT and Gemini, in providing electoral advice. The investigation was based on the Hungarian elections held in April, where Péter Magyar’s opposition Tisza party achieved a significant victory, ending Viktor Orbán’s lengthy tenure as Prime Minister.
Researchers created five distinct voter profiles aligned with different parties registered for the elections, using the respected Hungarian voting advice app Voksmonitor as a basis. Each profile was tested multiple times to see how well the AI tools could recommend suitable political parties. Alarmingly, the results indicated that ChatGPT failed to recommend Tisza in 90% of cases when provided with a profile that aligned with its political stance.
Inconsistent and Unreliable Outputs
The research highlighted that the AI systems not only misclassified party affiliations but also recommended parties that were not on the election ballot. ChatGPT and Gemini produced materially different recommendations for identical voter profiles, with Tisza receiving a mere 2% match in percentage-based tests. Conversely, profiles aligned with the ruling Fidesz party were consistently identified as primary voting options around 50% of the time.
Researchers noted that these AI models often began their responses with disclaimers stating their inability to provide political advice. However, they contradicted themselves by offering detailed and persuasive recommendations that could mislead users into believing the outputs were credible.
This inconsistency poses a risk, especially in more competitive electoral scenarios where voters may be less certain about their choices. The study did not claim that the election results were directly influenced by the AI outputs, but it cautioned that in different contexts, such misleading information could sway voter decisions.
The Regulatory Gap in AI Oversight
The researchers pointed out a significant regulatory gap concerning the use of AI in electoral processes. According to the EU’s AI Act, providers of general-purpose AI systems are required to assess systemic risks, while the Digital Services Act addresses “systemic risks” to electoral integrity. However, AI chatbots currently operate in a grey area, lacking adequate oversight.
Eva Simon, head of Liberties’ tech and rights programme, emphasised the need for stringent safeguards. She argued that AI systems should not offer personalised electoral recommendations unless they can ensure transparency, accuracy, and consistency. “Democracy cannot rely on opaque systems that claim neutrality while delivering unexplainable, unverified advice,” Simon asserted.
Why it Matters
The implications of this study are profound for the future of democratic processes in Europe and beyond. As reliance on AI systems for political guidance grows, the potential for misinformation increases, jeopardising the integrity of elections. This highlights the urgent need for regulatory frameworks that ensure accountability and transparency in AI tools used for electoral purposes. Without such measures, the promise of technology could inadvertently undermine the very tenets of democracy it aims to support.
