Main finding

Artificial intelligence in elections no longer means a single convincing deepfake. It reduces the cost of producing text, voice and video, increases the volume of campaign content, accelerates coordinated distribution and extends political competition into chatbot answers. Israel’s 2026 campaign offers an unusually data-rich warning: roughly one in eight monitored election-related social-media posts in recent months was created or altered with AI. That is approximately 12.5 percent.

Georgia’s most plausible risk is not one flawless video. It is the combined effect of synthetic audio, short emotional clips, anonymous pages, bot-like accounts, coordinated comments and apparently neutral chatbot answers. Such a system can be built and scaled faster than journalists, fact-checkers or election authorities can assess it. Preparedness therefore has to be measured by response speed, source transparency and the ability to deliver reliable information to voters before a decision is made.

The central recommendation is an election trust infrastructure built on four layers: clear disclosure and sponsor identification, an hours-not-days response channel, continuous monitoring of inauthentic networks, and practical voter media literacy. This is a BTUAI analytical framework, not an adopted Georgian or international standard.

Israel as an early warning

The Wall Street Journal’s 20 September 2026 analysis describes Israel’s 27 October election as a campaign in which generative AI has moved from isolated experimentation into ordinary political infrastructure. Senior politicians circulate synthetic videos, campaigns generate cheap and fast variations, and regulators try to adapt rules while the campaign is already under way.

Citing a research initiative at the Open University of Israel, the report says roughly one in eight election-related posts reviewed in recent months was created or altered with AI. Likud was associated with 46 percent of those AI posts. This does not mean that 46 percent of all election-related posts were Likud AI content. Multiplying the exact one-in-eight equivalent, 12.5 percent, by 46 percent produces 5.75 percent, but this is only a conditional arithmetic illustration. It is not an independently observed estimate unless the two percentages use matching samples and denominators.

The deeper significance lies in campaign economics. Expensive political video once imposed a practical limit on volume. Generative tools allow campaigns to create dozens of versions within hours, alter tone and target audience, synthesize different voices and test reactions immediately. Advantage can shift from the actor with the strongest message to the actor able to produce and distribute the most variations.

What the numbers show

The response speed gap

Israel’s Central Elections Committee conducted dozens of crisis simulations with security agencies. One scenario involved a false election-day message claiming that polling stations in a city had closed because of a security emergency. The scenario captures the central operational problem: a claim can become viral in minutes while an official rebuttal requires coordination, evidence and an authorized institutional voice.

A rule announced in late July requires AI-generated election photos, video and audio to be labelled. Yet a live dashboard operated by the Brandeis Institute identified 166 undisclosed AI posts in one month. The institute estimated that 98 percent of content it flagged as violating AI disclosure rules never reached the election committee. Expressed in reverse, only about 2 percent reached the committee. This is not a complete performance evaluation, but it reveals a large loss between detection and formal review.

Synthetic media is only one layer. In a domestic campaign attacking the legitimacy of the election committee, bots and bot-like accounts represented 17 percent of posts, equivalent to roughly one in 5.9 posts. Such networks can amplify genuine messages, manufacture the appearance of consensus or deepen suspicion about an election result. Content authentication and behavioral network analysis therefore cannot be separated.

Competition for chatbot answers

The campaign frontier now extends beyond the social feed. According to the WSJ report, at least one Israeli politician was using a generative engine optimization consultant to improve how chatbots answered questions about him. Traditional search optimization seeks a higher ranking in a list of links. Generative optimization seeks to shape the information environment a model retrieves and summarizes.

The political risk is the answer’s appearance of neutrality. A voter may ask about a candidate, party or voting rule and receive a concise, confident response without seeing the full source trail. If the web is flooded with coordinated and repetitive material, a generative system may treat an artificially enlarged footprint as a widely established position. Georgian public institutions should therefore publish election information in timely, structured, citable and easily retrievable Georgian-language formats, rather than relying on scattered announcements alone.

Georgia is not starting from zero

Georgia already had an online influence infrastructure before generative AI. ISFED’s social-media monitoring regularly documents political advertising, coordinated campaigns and inauthentic accounts. Its 2026 monitoring page, for example, includes recurring reviews of political ads on Meta platforms and research into a network of more than 1,500 fake Facebook accounts. These cases do not establish that every coordinated campaign uses AI. They show that the distribution channels already exist and that generative tools can expand their production capacity.

A one-year CRRC-Georgia study provides another measure of the researchable local environment. From 29 November 2024 to 29 November 2025, researchers conducted in-depth analysis of 4,017 posts from 16 selected public Facebook pages and groups. The monitoring covered Georgian-, Armenian- and Azerbaijani-speaking audiences and classified distorted context, conspiracy framing, visual manipulation and the use of AI in misleading images and videos.

The CRRC dataset is not an election-specific census and cannot establish the prevalence of AI-generated election content. Its value is different: it demonstrates that a multilingual, visually manipulative and platform-based information environment is already a measurable reality in Georgia. Election readiness must therefore operate in Georgian, Armenian and Azerbaijani segments; a rebuttal delayed in one language may allow a false claim to continue spreading in another.

A four-layer readiness framework

Transparency and accountable sponsors

Synthetic or materially altered political content should carry a visible and machine-readable disclosure covering images, video, audio and edits that substantially change the meaning of an event. A label saying AI is not enough: voters should also know who commissioned, funded or first distributed the material. The EU AI Act offers a relevant reference point. Its transparency rules, applicable from August 2026, require clear labelling for deepfakes and certain AI-generated public-interest content. Any Georgian legal model would still require a separate rights-based and proportionality assessment.

A unified rapid response channel

Before an election, the election administration, relevant public bodies, platforms, newsrooms, fact-checkers and civil monitors should agree on an escalation network. High-risk cases need an initial public status within hours: verified, probably manipulated, or under investigation. Precise acknowledgement of uncertainty is more credible than silence that leaves an information vacuum to the manipulator.

Election-day claims need a specific protocol: polling-station closure, changed voting hours, ballot invalidation, candidate withdrawal, violence or a security incident, and allegations that results were falsified. Each category should have an authorized confirming source and a pre-arranged distribution path.

Monitoring inauthentic networks

Monitoring should not stop at asking whether one image is generated. It should examine account creation dates, synchronized activity, repeated text, coordinated comments, advertising sponsors and major targeting parameters. A searchable election-ad archive should ideally include the buyer, campaign period, expenditure, core targeting attributes and AI-use status. This allows researchers to see a campaign network rather than evaluate isolated posts.

Practical voter media literacy

Voters should not be expected to perform technical deepfake forensics. They need a short action rule: do not share only because a clip is emotionally powerful; locate the full recording; verify the original source and publication time; check whether the responsible authority confirms the claim; and preserve the link or screenshot when reporting a concern. Journalists also need formats that debunk suspicious content without repeatedly broadcasting the manipulation to a larger audience.

What Georgia should measure

Preparedness should become measurable. Useful indicators include minutes from detection of a high-risk claim to the first public response; average platform response time; the share of flagged cases that reach the competent authority; the number of repeat uploads; whether corrections reach the original audience; and equal response speed in Georgian, Armenian and Azerbaijani.

These metrics do not decide which political claim is true. They test the operating system: whether the country can see the problem, preserve evidence and deliver verified information before voters decide. Political expression, satire and technological experimentation should remain protected; disclosure and response measures should be narrow, proportionate and appealable.

Conclusion

AI-enabled elections are no longer a distant scenario for Georgia. Production barriers are low, distribution networks already exist and chatbots are emerging as new intermediaries of political information. Israel’s experience suggests that a system which sees a violation only after a complaint may lose most of the relevant content when production is massive and distribution takes minutes.

Georgia should not seek to prohibit all synthetic political content. It should ensure that voters know when they are seeing manufactured evidence, who is behind it and where they can verify the facts before deciding. That requires all four layers together: transparency, rapid response, network monitoring and media literacy. Together, they form an infrastructure for electoral trust.

Data and principal sources

The Wall Street Journal, How AI Is Reshaping Election Campaigns in Israel, 20 September 2026 — https://www.wsj.com/world/middle-east/how-ai-is-reshaping-election-campaigns-in-israel-f5a2e79a

CRRC-Georgia, Anti-Western Rhetoric on Facebook From November 29 2024 to November 29 2025 — https://crrc.ge/en/report-anti-western-rhetoric-on-facebook-from-november-29-2024-to-november-29-2025/

CRRC-Georgia, Anti-Western Rhetoric and Visual Manipulation on Facebook — https://crrc.ge/en/blog-anti-western-rhetoric-and-visual-manipulation-on-facebook-one-year-monitoring-findings/

ISFED, Social Media Monitoring — https://isfed.ge/geo/sotsialuri-mediis-monitoringi

European Commission, AI Act regulatory framework — https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai

Prepared by the academic team of Business and Technology University and the BTUAI Research Team, Tbilisi, Georgia.

Legal notice

This material is analytical and educational. It does not constitute financial, investment, tax or legal advice. Consult an appropriate specialist before making a specific decision.