Before the Ballot: Data Poisoning Risks to Elections

Data poisoning as an election integrity threat differs structurally from conventional disinformation: rather than targeting voters directly, it targets the AI systems that voters increasingly use to inform themselves. The Pravda network — publishing content at scale across many domains to seed false or manipulative material into the sources AI models learn from or retrieve from — is the clearest documented example of this approach in an electoral context. This INSS report focuses the analysis on the Israeli case, drawing on evidence of Russian and Iranian influence activity, but the threat model it describes is not Israel-specific.

The finding that no systematic audit has tested how leading AI models answer Israeli election questions in Hebrew and Arabic is a significant gap. It reflects a broader problem in the field: the AI literacy required to assess data poisoning risks is not yet standard in electoral monitoring organisations, and the AI companies whose models are being targeted have limited visibility into the upstream content ecosystem. The report’s recommendations — regular model audits, stronger official machine-readable information, public awareness, and legal frameworks for monitoring — are proportionate to the evidence and appropriately cautious about claims that outrun the current empirical base.