Israel’s election overseer on how to stop AI distorting votes

A relative of mine recently consulted an AI chatbot about which of the dishes his mother had prepared he should eat. I pointed out that he ate her cooking every week, whereas the chatbot had never tasted it. He replied that he consulted the chatbot about everything. Apparently, having no taste buds was no obstacle to becoming the family’s food critic.

The exchange was amusing. Its implications for my work are less so. As interim head of the Israeli Central Elections Committee, I am preparing for our parliamentary election on October 27th. AI companies should use the time remaining to address a question that reaches far beyond Israel: what happens when „What should I eat?” becomes „Who should I vote for?”

De redactie van NRC selecteert de beste artikelen uit The Economist voor een breder perspectief op internationale politiek en economie.

Chatbots can make politics easier to understand. They can explain electoral systems, make sense of manifestos and help citizens interrogate competing claims. Those benefits matter.

But a voting recommendation involves judgments about values, credibility and trade-offs. Presenting it as the personalised output of an apparently impartial machine can conceal the assumptions that produced it.

The danger need not begin with malicious intent. A model may draw on stale information, favour popular viewpoints or give different answers to similar questions. In 2025 the Dutch data-protection authority warned that chatbots gave distorted voting advice, repeatedly favouring two parties, sometimes despite users’ stated preferences. Research published in Nature found that models instructed to advocate particular candidates could shift participants’ preferences. That shows persuasive capacity, not that ordinary chatbot use has changed an election result.

Malicious actors introduce another risk. They can plant misleading material in sources that AI systems retrieve, hoping to have propaganda delivered in the reassuring voice of an assistant. Demos, a British think-tank, has documented leading AI services amplifying narratives from a Russian influence operation. An attacker need not control the model if they can contaminate its information supply.

For the voter, the author of that intervention may be invisible. A campaign leaflet announces its allegiance; an answer from a familiar chatbot can feel like disinterested advice from someone who knows you. The trust earned through hundreds of mundane exchanges can carry across into a decision with public consequences.

To the companies building these systems, I propose five commitments.

First, decline to recommend people vote one way or another, including through party rankings or purported compatibility scores. Explain policies and compare verifiable positions using clear criteria, while leaving the judgment to the voter.

Brazil’s electoral court has already banned AI providers from recommending or ranking candidates and parties, even when users ask them to do it. Companies should adopt this principle wherever they operate. Factual comparison and robust political discussion should remain available. Companies should publish the boundary clearly, so a safeguard for voter autonomy does not become an excuse to suppress inconvenient questions or criticism.

Second, make election information verifiable. Questions about eligibility, polling places, procedures and results should draw on current sources from the responsible electoral authorities, with links users can inspect. Political claims require clearly attributed sources and visible uncertainty. A party’s promise should be identified as merely a promise. Where reliable information is unavailable, the system should say so.

Third, defend the sources as well as the model. Detect co-ordinated attempts to flood retrieval systems. Check provenance. Avoid treating repetition across copied websites as independent corroboration. Establish rapid channels with election authorities and independent researchers to investigate errors and manipulation. Corrections must be made while voters can still use them.

Fourth, keep personalisation from becoming political profiling. Information entrusted to an assistant about someone’s finances, health or anxieties should never become material for steering their vote. Companies should explicitly prohibit using conversational memory or inferred vulnerabilities for electoral persuasion.

Fifth, make safeguards testable. Publish election-specific evaluations, disclose material failures and give independent researchers meaningful access under privacy protections. Test repeated conversations and attempts to evade restrictions. In Israel, that means checking Hebrew, Arabic, Russian and English. A promise of neutrality is inadequate without evidence that the safeguards work across languages and political viewpoints.

Election officials must provide clear information. Citizens must exercise judgment. Companies, however, control the systems through which that judgment is increasingly formed. Asking them to preserve voters’ autonomy is a practical condition for public trust. They should publish and follow through on these commitments before polling day, with named teams accountable for delivery.

My relative can return to his mother’s table next week and reconsider the chatbot’s culinary advice. An election offers no such weekly correction. AI should help citizens understand their choices. The choice itself must remain theirs.

Dean Livne is the acting director-general of the Israeli Central Elections Committee.

© 2026 The Economist Newspaper Limited. All rights reserved.

Kies NRC als voorkeursbron