AI が選挙に与える影響、国際会議で軽視される懸念

Last month, all 193 United Nations member states gathered in Geneva for the inaugural Global Dialogue on AI Governance. For two days they debated how the world should govern AI, yet paid little attention to one of AI’s most consequential governance challenges: its impact on elections. The omission risks undermining progress on every other aspect of the AI debate: safety, trustworthiness, capacity building, societal implications, and human rights.

Elections are not simply another sector affected by AI; they are the main mechanism that makes every other democratic right enforceable. Yet when we think about AI and elections, many of us think about disinformation. The preliminary report from the U.N.’s Independent International Scientific Panel on AI — intended to inform the Dialogue — does the same, mentioning AI in elections only in the context of information integrity and deepfakes targeting candidates.

Voter registers, biometric devices, identity databases, results-management software, and cloud infrastructure increasingly incorporate AI-enabled features.”

But one of the clearest lessons of 2024’s “year of elections” was that the impact of deepfakes was significantly overstated. This framing mistakes the loudest risk for the largest one. The more important question is not what happens when AI generates a fake video of a candidate, but what happens when it is embedded in the systems that administer elections themselves — in procurement decisions, biometric voter-verification systems, and data-sharing arrangements. That is where AI’s impact on democracy may prove most profound.

Consider India’s record. In 2015, the country’s election commission launched a program linking voters’ photo ID records to Aadhaar, the biometric national identity database. Aadhaar is run by the Unique Identification Authority of India, a statutory body under executive control, not an independent institution like the election commission. It trained a machine learning algorithm to cross-reference the two databases and flag duplicate and dead voters for deletion. Electoral data, meanwhile, flowed into a state-run data system, giving government authorities direct access to records the election commission alone was constitutionally meant to hold.

The results were staggering: Roughly 5.5 million voters were deleted from rolls in two states, many without any notice or path to reinstatement, before India’s Supreme Court halted the program. Right to Information disclosures later revealed that the matching algorithm’s failure rate was as high as 93%. Despite this, Aadhaar-voter linkage was revived nationally in 2021, bringing the same structural risk to all of India, and leaving a clearly flawed machine learning model to play a major role in deciding who retains the right to vote.

There is a deeper concern beyond any single algorithmic failure: the gradual transfer of electoral authority away from public institutions and into systems they do not fully control. Across much of Africa, Asia, and Latin America, election technologies are designed, supplied, and hosted by private vendors that operate beyond the reach of national electoral oversight.

Voter registers, biometric devices, identity databases, results-management software, and cloud infrastructure increasingly incorporate AI-enabled features, and voters are turning to chatbots for advice and information before going to the polls. This creates serious risks. Sensitive voter data may leave national jurisdictions, proprietary systems may be impossible to audit, and hidden algorithmic errors may exclude thousands of voters without ever being challenged, or even detected.

Chatbots such as OpenAI’s ChatGPT and Google’s Gemini, now fully incorporated into the search function, may hallucinate incorrect information, or reinforce bias, particularly in contexts where digital information is limited and linguistic diversity is high. This exacerbates existing inequities by providing poor-quality information to communities that are already least represented online.

Opacity and complexity

These risks are not unique to AI, but AI’s opacity and complexity — as well as the scale at which the technology operates — compound them exponentially. Absent clear governance standards, election authorities may find themselves accountable for decisions embedded in technologies they did not design, cannot fully inspect, and often lack the authority to regulate.

Election authorities may find themselves accountable for decisions embedded in technologies they did not design, cannot fully inspect, and often lack the authority to regulate.”

It is only by foregrounding elections in global AI governance discussions that governments, civil society, and the private sector can help ensure that AI does not undermine the institutional foundation of democracy. Key questions must be asked: Who controls electoral data? Who can audit these systems? And who is accountable when a machine learning model affects a citizen’s right to vote?

We can learn from the approach adopted by African institutions. In April, the African Union’s Peace and Security Council said that Africa must “shape and control” its AI ecosystem because a corrupted voter register is not only a rights violation, it is a trigger for conflict. The Council mandated the creation of the African Union Advisory Group on AI in peace, security, and governance to develop guidance for electoral commissions on the AI systems they procure. It deemed data localization, technology transfer, and disclosure of source code as nonnegotiable conditions. The AU’s approach does two things the U.N. Global Dialogue has thus far failed to do: It foregrounds elections in AI governance, and anchors that governance in regional realities. It is a regional model worth replicating on the global stage.

While concern about AI’s impact on health, food security, climate, and other sectors is growing, attention to AI governance in elections is equally necessary. Without the ability to elect or remove a government, citizens lose the principal peaceful means of holding power to account when leaders mismanage a drought, misallocate healthcare spending, or deny them their rights. By leaving elections out of the AI governance agenda, the Global Dialogue effectively removed the accountability mechanism for AI applications in every sector.

Electoral integrity deserves a dedicated track when the Global Dialogue reconvenes in New York in May 2027, with the same institutional weight given to safety, human rights, and the digital divide. Until elections are recognized as central to AI governance, global efforts to make AI safe, trustworthy, and rights-respecting will rest on a dangerous illusion — that democracy can be protected without protecting the institutions that make democracy possible.