2026年选举信息与安全保障措施
在全球选举来临之际，我们致力于帮助人们获取信息、支持网络安全防御者，并提升人工智能的透明度
2026年是自生成式人工智能广泛普及以来全球第二个重大选举年，我们将继续在2024年奠定的基础上，帮助保护全球各国和地区的选举安全。
我们的重点是以以下方式构建并负责任地部署突破性产品：
- 呈现关于投票和结果的可靠信息
- 支持网络基础设施防御者
- 提升人工智能生成内容的透明度
- 打击恶意行为者的滥用行为
- 监控模型中的偏见，以保持ChatGPT回复的政治中立性
呈现可靠信息
人们已经在使用ChatGPT，用他们偏好的语言询问关于公民活动的实际问题：如何登记、在哪里投票、有哪些截止日期、正在发生的新闻事件进展如何，或者在哪里可以找到官方选举结果。
在2024年工作的基础上，我们正与合作伙伴共同努力，引导人们获取关于投票的可靠信息来源。从今年秋季开始，在美国和巴西，OpenAI将在选举之夜结果陆续出炉时，提供来自美联社的实时计票数据。在美国，我们还将与Democracy Works合作，当人们询问相关话题时，显示关于投票和登记流程的可靠信息，包括投票地点及其他选举后勤事宜。在全球范围内，我们将继续优化网络搜索呈现有用信息并附带来源链接的方式。
自2024年以来，我们持续改进了人们向ChatGPT询问选举话题和突发新闻时所获信息的质量。ChatGPT可以搜索网络，提供更可靠的答案并附带来源链接，方便人们深入了解。
支持网络基础设施防御者
当前是各行业网络安全防御者的重要时刻，我们相信人工智能在加固数字基础设施（包括支持选举的系统）方面发挥着关键作用。OpenAI致力于在整个基础设施栈中构建韧性，包括以支持选举执行的方式。
我们近期宣布了 Daybreak 计划，旨在改变软件的构建与防御方式。Daybreak 包含多项旨在提升软件安全性与韧性的项目，其中包括 Codex Security——该工具可自动识别、验证并协助修复开发者代码中的漏洞。在更高级的防御层面，我们的可信网络访问（TAC）计划为经过验证的个人提供前沿模型访问权限，用于网络防御。我们已向美国注册投票系统制造商提供了 Codex Security 及 TAC 访问权限，并与全美州务卿协会（NASS）及全美州选举主管协会（NASED）保持沟通，确保州选举机构能够及时了解最新的网络能力及我们为防御者提供的工具。
提升透明度
人们越来越多地使用 AI 工具创作内容，并将其分享到社交媒体、即时通讯应用及网络上。为帮助打击具有误导性的“深度伪造”内容，我们正在投资构建多层次的内容溯源方案，使人们能够验证所看到的内容是否由 AI 创建或修改。
上周，我们宣布了一项合作，将 SynthID 数字水印引入通过 ChatGPT、Codex 或 OpenAI API 生成的图像。这建立在我们持续采用 C2PA（内容来源与真实性联盟）标准的基础之上，该标准利用元数据和加密签名，使图像的相关信息能够安全地随内容本身一同传递。这两种方法互为补充：SynthID 嵌入了一层不可见的水印，能够经受截图及其他形式的变换；而元数据则能提供比单纯水印更丰富的信息。
我们还预览了一款公开验证工具，让人们能够检查在平台之外看到的图像是否由 OpenAI 工具生成。该网站能够检测是否存在源自 OpenAI 的 SynthID 水印，并在发现 C2PA 元数据时将其呈现出来。
溯源工具并非解决选举相关欺骗行为的完整方案，但它们是更广泛诚信框架的重要组成部分。由于大多数人是在社交媒体和即时通讯平台上接触到 AI 生成内容的，我们很高兴能与社交媒体公司合作，在他们考虑采取措施保护选举时，将溯源标记作为一项相关信号，用于决定在用户信息流中推荐和分发哪些公民内容。
我们也很高兴支持那些旨在提升透明度的立法：
- 《保护选举免受欺骗性 AI 影响法案》（S. 1213）将新增禁令，禁止在联邦选举活动中故意传播涉及联邦候选人的、具有实质欺骗性的 AI 生成内容。
- 《帮助选举管理者应对 AI 法案》（S. 2346）将要求选举援助委员会发布一份报告，为选举办公室提供关于在选举管理中 AI 技术的使用及风险的志愿性指南。考虑到前沿模型不断增强的网络能力，这些指南尤为重要。
打击滥用行为
OpenAI 将继续尽己所能打击选举干预行为。我们执行《使用政策》，禁止用户利用我们的工具进行选举干预、削弱投票意愿，或对 AI 生成内容的来源进行欺骗。
自 2024 年以来，我们的检测与执行系统持续改进。我们在设有隐私保护措施和明确审查流程的前提下，对政策进行监控和执行。当有人试图利用我们的工具干预选举过程、阻碍参与，或就 AI 生成内容的来源误导公众（包括作为协同欺骗活动的一部分）时，我们会采取行动。一旦发现滥用行为，我们将采取执行措施，包括限制或终止其对我们服务的访问权限。
我们的情报与调查、安全、安保及诚信团队致力于检测滥用模式，并调查试图利用我们的工具进行欺骗性或与选举相关的滥用行为，包括隐蔽的影响力行动。自 2024 年 2 月以来，我们定期发布调查报告，并将继续这样做。
在公众、政策制定者、竞选活动以及企业共同为人工智能在政治领域的应用制定规范的同时，我们持续禁止使用我们的产品来创建或分发针对候选人、政党或公投提案的大规模竞选宣传内容。但我们认为，政治竞选活动应当能够使用我们的工具来开展负责任、由人类主导的工作，例如起草内部简报、制定计划、日常写作、翻译工作、合规事务以及行政任务。这些使用场景在我们的使用政策中是允许的。
人工智能工具为人们提供了真正的机会，能够以新的方式与他人互动，通过跨越语言、教育水平和背景的障碍，来扩展和深化公民社群。我们认为，人们能够使用人工智能系统来了解、探索和讨论政治议题至关重要，我们将继续支持这一点，同时防止恶意行为者的滥用。
根据我们禁止将服务用于大规模竞选宣传的规定，在本选举周期内，我们将不允许广告商在我们的平台上投放政治广告。您可以阅读更多关于我们如何处理 ChatGPT 上广告投放的信息。
监控我们模型中的政治偏见
我们知道，人们使用 ChatGPT 的目的不仅仅是获取选举相关信息。他们利用 ChatGPT 来探索想法、检验论点，并加深对他们所关心议题的理解。这就是为什么我们致力于消除 ChatGPT 在回答问题时的政治偏见。
我们的模型规范原则“共同寻求真相”概述了我们的承诺：默认保持 ChatGPT 的客观性，由用户掌控，包括当用户参与政治对话时。ChatGPT 是一个让人们接触宏大思想并形成自己对问题看法的工具。因此，该模型的设计旨在避免操纵用户的行为，或隐瞒相关事实与观点以引导用户走向特定方向。去年秋天，我们宣布了一项新的政治偏见评估方法，用于测试我们的模型保持客观的能力。这项重要工作使 OpenAI 在测试模型如何回应政治话题的政治偏见评估领域处于领先地位。
在履行自身职责的过程中不断学习
我们将继续向合作伙伴学习，更新我们的安全防护措施，并不断调整适应。我们的目标是支持人们自由参与选举并自主做出决定——通过提供可靠信息、保持透明度以及实施有效的安全防护。我们期待在今年及未来的选举中，进一步深化对用户、公民社会及选民的支持方式。
Election information and safeguards in 2026
Ahead of global elections, we’re helping people access information, supporting cyber defenders, and increasing AI transparency
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 as results roll in on election night. In the US, we will also partner with Democracy Works 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 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) 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) 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 regularlypublishedreportswithourfindings 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 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.