2026 年選挙に関する情報と安全対策

2026 is the world’s second major election year since generative AI became widely available, and we’re continuing to build on the foundation we laid in 2024 to help protect elections in countries and territories around the world.

Our focus is to build and responsibly deploy groundbreaking products in ways that:

• Surface reliable information about voting and results
• Support cyber infrastructure defenders
• Increase transparency around AI-generated content
• Combat misuse by bad actors
• Monitor bias in our models to keep ChatGPT’s responses politically neutral

Surfacing reliable information

People already use ChatGPT to ask practical questions in their preferred languages about civic events: how to register, where to vote, what deadlines apply, what’s happening with a developing news event, or where to find official election results.

Building on our efforts in 2024, we are working with partners to direct people to reliable sources of information about voting. Beginning this fall in the United States and Brazil, OpenAI will provide live vote counts from The Associated Press⁠(opens in a new window) as results roll in on election night. In the US, we will also partner with Democracy Works⁠(opens in a new window) to display reliable information about voting and registration processes, including voting locations and other election logistics, when people ask about those topics. Globally, we will continue to refine the way web search surfaces helpful information with source links.

Since 2024, we have continued to improve the quality of information people get when they ask ChatGPT about election topics and breaking news. ChatGPT can search the web to provide stronger answers with source links so people can go deeper.

Supporting cyber infrastructure defenders

This is an important moment for cyber defenders across industries, and we believe AI plays a critical role in hardening digital infrastructure — including systems that support elections. OpenAI is committed to building resilience across the infrastructure stack, including in ways that support election execution.

We recently announced Daybreak⁠, our effort to change the way software is built and defended. Daybreak includes a number of programs intended to make software safer and more resilient, including Codex Security, which automatically identifies, validates, and helps remediate vulnerabilities in developers code. For more advanced defense, our Trusted Access for Cyber ⁠(TAC) program provides verified individuals access to frontier models for cyber defense. We have offered Codex Security as well as TAC access to registered voting system manufacturers⁠(opens in a new window) in the US, and we are engaging the National Association of Secretaries of State (NASS) and the National Association of State Election Directors (NASED) to ensure state election authorities are briefed on the latest cyber capabilities and our tools for defenders.

Increasing transparency

People are increasingly using AI tools to create content they then share on social media, messaging apps, and the web. To help combat misleading “deepfakes”, we are investing in a multi-layered provenance approach that will equip people to verify whether content they’re seeing has been created or modified with AI.

Last week, we announced a partnership to bring SynthID ⁠digital watermarks to images generated through ChatGPT⁠, Codex, or the OpenAI API. This builds on our ongoing commitment to using the C2PA (Coalition for Content Provenance and Authenticity) standard, which uses metadata and cryptographic signatures to help information about an image securely travel with the content itself. These two approaches are complementary: SynthID embeds an invisible watermarking layer that can hold up through screenshots and other transformations, while metadata can provide more information than a watermark alone.

We’re also previewing a public verification tool⁠ that lets people check whether an image they encounter off-platform was generated using OpenAI tools. The site can detect whether a SynthID watermark originating from OpenAI is present, as well as surface C2PA metadata when it is found.

Provenance tools aren’t a complete solution to election-related deception, but they are an important part of a broader integrity framework. And since most people encounter AI content on social media and messaging platforms, we are happy to partner with social media companies as they consider steps to protect elections by using provenance markers as a relevant signal in deciding which civic content they recommend and distribute in people’s feed.

And we are excited to support legislation that would advance transparency:

• Protect Elections from Deceptive AI Act (S. 1213)⁠(opens in a new window) would add new prohibitions from knowingly distributing materially deceptive AI-generated content involving federal candidates when used for federal election activity.
• Preparing Election Administrators for AI Act (S. 2346)⁠(opens in a new window) would require the Election Assistance Commission to issue a report with voluntary guidelines for election offices on the use and risks of AI technologies in election administration. These guidelines are especially important given increased cyber capabilities of frontier models.

Combatting misuse

OpenAI will continue to do our part to combat election interference. We enforce Usage Policies⁠ prohibiting users from deploying our tools for election interference, demobilization, or deception about the origin of AI-generated content.

Our detection and enforcement systems have continued to improve since 2024. We monitor and enforce⁠ policies with privacy safeguards in place and clear review processes. And we act when someone tries to use our tools to interfere with the election process, discourage participation, or mislead people about the origin of AI-generated content, including as part of coordinated deceptive activity. When we identify abuse, we take enforcement action, including restricting or terminating access to our services.

Our Intelligence and Investigations, Safety, Security, and Integrity teams work to detect misuse patterns and investigate attempts to use our tools for deceptive or election-related abuse, including covert influence operations. Since February 2024⁠, we’ve regularly⁠ published⁠ reports⁠ with⁠ our⁠ findings⁠ and will continue to do so.

While the public, policymakers, campaigns, and companies work to develop norms for AI use in politics, we continue to prohibit the use of our products to create or distribute scaled campaign messaging for or against a candidate, political party, or ballot measure. We do think political campaigns should be able to use our tools for responsible, human-directed work, like drafting internal briefings, planning, everyday writing, translation work, compliance, and administrative tasks. These use cases are allowed under our usage policies.

AI tools offer real opportunities for people to engage others in new ways that expand and deepen civic communities by expanding access across languages, education-levels, and backgrounds. We believe it’s important that people can use AI systems to learn about, explore, and discuss political issues, and we will continue to enable that, while avoiding misuse by bad actors.

Consistent with our rules against using our services for scaled campaign advocacy, we will not allow advertisers to run political ads on our platform this cycle. You can read more about how we are approaching advertising on ChatGPT.⁠

Monitoring political bias in our models

We know people come to ChatGPT for more than election logistics. They use ChatGPT to explore ideas, test arguments, and gain more understanding of issues they care about. That’s why we’re committed to combatting political bias in the way ChatGPT responds to questions.

Our Model Spec principle Seeking the Truth Together⁠(opens in a new window) outlines our commitment to keeping ChatGPT objective by default, with the user in control, including when users engage in political conversations. ChatGPT is a tool for people to engage with big ideas and develop their own perspective on issues. So the model is designed to avoid behavior that would manipulate a user or to conceal relevant facts or viewpoints to nudge them in a particular direction. Last fall, we announced ⁠a new political bias evaluation we use to test our models’ ability to remain objective. This important work has made OpenAI a leader in political bias evaluations that test how models respond to political topics.

Learning while we do our part

We will continue to learn from partners, update our safeguards, and adapt. Our goal is to support people’s ability to participate freely in elections and make their own decisions — with reliable information, transparency, and effective safeguards. We look forward to deepening the ways we can support our users, civil society, and voters as they participate in elections this year and beyond.