# OpenAI 公布 EU AI Act 下的文本溯源方案,推出 textGrain 文本水印
> 原标题:Our approach to EU text provenance rules
- 来源:OpenAI:官网动态
- 发布时间:2026-10-05T15:00:00.000Z
- AIHOT:https://aihot.news/items/xx5mgdemrqmqw5zw410sbawcz
- 原文:https://openai.com/index/eu-text-provenance
## 摘要
OpenAI 公布应对 EU AI Act 的文本水印方案,发布在模型词选择中加入不可见统计信号的 textGrain 技术,API 客户即日起可对部分模型选择性开启水印,未来数周将在欧盟地区为 ChatGPT 和 Codex 输出添加隐形水印,检测器暂只向获批的研究者和专家机构开放。
## 推荐理由
原文给出 textGrain 的检测数据和局限边界,读者可以据此了解文本水印目前的实际能力和适用条件。
## 正文 · 中文译文
## 我们在欧盟文本溯源规则上的做法
在当今技术条件下促进透明度。
加载中…
[音频 1](https://openai.com/index/eu-text-provenance)
分享
[内容溯源](https://openai.com/index/advancing-content-provenance/)帮助人们了解内容的来源、创建或编辑方式,以及其中是否包含与我们模型相关的信号。我们已公开提供工具,用于识别由我们模型生成的图像和音频。今天,我们分享针对欧盟《人工智能法案》的文本水印方案,以及它如何融入我们更广泛的工作。
欧盟《人工智能法案》要求生成式 AI 提供商以机器可读的方式使生成的文本可被识别。文本水印和检测仍属早期技术,存在显著局限,关于其益处和负责任使用的看法仍在发展中。我们的分阶段方案既反映了欧盟《人工智能法案》的要求,也反映了该技术的局限,并强调对文本水印能告诉人们什么、不能告诉人们什么的透明度:
- 从今天起,全球 API 客户将能够为部分模型选择启用文本水印。文本水印在 API 中默认保持关闭。
- 在未来几周内,我们将为欧盟境内符合条件的 ChatGPT 和 Codex 文本输出添加不可见水印。
- 我们开放文本水印检测器的申请。访问权限最初将仅限于经批准的研究人员和专家组织,他们可以帮助我们评估和改进这项技术。
- 以上仅适用于文本溯源——我们用于音频和图像的验证工具,包括我们的 [openai.com/verify](https://openai.com/verify) 网页工具和我们的 [Content Provenance API(在新窗口中打开)](https://developers.openai.com/api/docs/guides/content-provenance),将继续向希望了解图像或音频文件是否由我们某个系统生成的组织公开开放。
## 我们的水印如何工作及表现如何
我们的文本水印技术 textGrain 会向模型的用词选择中添加一种不可见的统计信号。我们的检测器会寻找该信号,以评估一段文字是否包含 OpenAI 水印。有关 textGrain 工作原理的更多细节可在我们的[技术报告(在新窗口中打开)](https://cdn.openai.com/pdf/e9508624-d767-41b6-a26d-e34ca798ada6/textgrain-entropy-calibrated-watermarking-for-language-model-text.pdf)中找到,该报告将在未来几周内更新更多细节。我们还计划将该技术开源,以便他人在此基础上继续开发。
在我们的评估中,textGrain 达到或超过了我们测试过的其他方法的表现,包括用于文本的 SynthID。即便如此,理想条件下的强劲表现并不能保证在日常使用中可靠检测。
检测器可能犯两类错误:它们可能在不存在水印时报告存在水印——即假阳性——或漏掉存在的水印——即假阴性。以下评估说明了其中一些挑战:
- **更短或更受限的文本更难检测。**在 1% 的目标假阳性率下,对于心理学等内容,我们的检测器在约 80% 的 200 token 段落中识别出水印,而在 400 token 段落中约为 95%。对于数学等用词灵活性较低的内容,检测率则大幅降低。
- **编辑会削弱水印。**在一项针对 400 token 段落的评估中,将 10% 的词语替换为同义词使检测率从约 92% 降至 66%。替换 25% 的词语则使其降至 17%。
这些局限性促使我们决定仅向经过批准的研究人员和专家组织提供初始检测器访问权限,他们可以帮助我们评估可靠性和负责任的使用方式。
该图表展示了针对数学和心理学问题的带水印响应的结果,这些问题来自 [ELI5 数据集(在新窗口中打开)](https://huggingface.co/datasets/rexarski/eli5_category),目标假阳性率为 1%。检测效果随文本长度增加而提升,但对于用词灵活性较低的内容(如数学),整体检测率显著更低。
编辑会大幅削弱水印信号。该图表展示了替换段落中 10% 或 25% 的词语如何影响检测。结果基于对来自 [ELI5(在新窗口中打开)](https://aclanthology.org/P19-1346/) 问题的带水印英文响应。
### 水印对输出质量的影响
在用于评估 Astra(我们最新的前沿模型)的各项基准测试中,我们未发现加水印与不加水印之间存在有意义的性能差异。
| **基准测试** | **无水印文本(Astra,max)** | **带水印文本(Astra,max)** |
| --- | --- | --- |
| Artificial Analysis Intelligence Index | 49.57 分 | 49.76 分 |
| AutomationBench | 34.09% | 34.86% |
| DeepSWE v1.1 | 72.80% | 71.68% |
| Terminal-Bench 4.0 | 53.90% | 56.06% |
| Terminal-Bench Science 0.1 | 56.90% | 60.00% |
| BrowseComp | 87.92% | 87.35% |
| HealthBench Professional | 64.27% | 64.60% |
| GPQA Diamond | 94.44% | 93.94% |
## 文本水印无法告诉你什么
文本水印只能提供关于我们的系统在段落中所起作用的有限信号。理解检测结果能得出和不能得出什么结论非常重要。
- **水印不能衡量人类的贡献。** 它可以表明 OpenAI 系统生成或处理了段落的某部分,但无法说明其中包含多少人类判断、编辑或创造力。
- **水印不能确立所有权或责任。** 它不能确定文本归谁所有、其使用是否合法、是否需要披露,或由谁负责。
- **水印不能识别用户。** 它不会将个人、组织、账户、提示词或对话与文本关联起来。
- **水印不能验证准确性。** 它不能告诉你段落是否真实、具有误导性、有害,或是否呈现于正确的语境中。
- **未检测到水印并不能证明是人类创作。** 使用 OpenAI 工具生成的文本可能因过短、经过编辑或翻译而无法可靠检测。它也可能来自不受支持的模型、早于水印功能推出,或由其他公司的工具生成。
## 我们的下一步
- **在欧盟推出文本水印。** 在未来几周内,我们将仅在欧盟面向所有套餐的符合条件的 ChatGPT 和 Codex 用户引入文本水印。我们不会在发布时将文本水印设为全球默认。这种区域性做法让我们有空间从实际使用和反馈中学习。
- **为 API 客户启用可选水印。** 从今天起,全球的 API 客户将能够为部分模型选择启用带水印的文本输出。这让客户可以决定水印如何契合其透明度义务以及他们为用户提供的体验。我们还在与云合作伙伴合作,在未来几周内为通过其服务访问的 OpenAI 模型输出提供水印。
- **为研究人员和专家组织提供检测器访问权限。** 经批准的研究人员和专家组织可从今天起[申请](https://openai.com/form/content-provenance-api/)。根据[实践准则(在新窗口中打开)](https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content),访问权限最初将按个案授予,以支持文本来源的评估和改进。该工具将报告是否检测到 OpenAI 水印,而不会识别用户或披露其提示或对话。鉴于水印漏检和误报的风险,我们在发布时不会将其公开提供。
随着技术、标准和证据的发展,我们预计会重新审视这一方法的每个部分。
## 我们在内容来源方面的更广泛方法
没有任何单一来源技术本身足够,因此我们采用分层方法,结合开放标准、持久水印和验证工具。
我们为支持的图像输出添加 Content Credentials,[符合 C2PA(在新窗口中打开)](https://c2pa.org/conformance/),在支持的图像和音频中嵌入不可见的 [SynthID(在新窗口中打开)](https://deepmind.google/models/synthid/) 水印,并通过 [openai.com/verify](https://openai.com/verify) 和我们的 [Content Provenance API(在新窗口中打开)](https://developers.openai.com/api/docs/guides/content-provenance) 提供图像和音频验证。这些技术相互补充:Content Credentials 可以记录文件的来源和历史,而不可见水印可以在元数据被移除时保留信号。
正如我们在[选举保障工作](https://openai.com/index/election-safeguards-2026/)中所述,来源信息可以帮助人们和平台评估可能具有误导性的 AI 生成内容,包括深度伪造。我们将这些信号与政策、滥用检测、报告、调查和执法相结合,并与研究人员、标准机构、平台和民间社会合作,使来源信息在更广泛的生态系统中发挥作用。
将来源信息扩展到文本需要考虑其被重写、翻译或编辑的容易程度。正如我们在[六月](https://openai.com/index/supporting-eu-trustworthy-ai-ecosystem/)所讨论的,我们的方法必须反映当前技术的实际局限。我们将继续改进检测,研究水印如何经受编辑和翻译,并探索更有意义地区分 AI 辅助与 AI 创作的方法。我们将随着证据、标准和监管要求的发展调整我们的方法,并在我们认为结果可以被负责任地解读时扩大检测器的访问权限。
如需了解更多信息,请访问我们的[帮助中心文章(在新窗口中打开)](https://help.openai.com/en/articles/8912793-provenance-signals-content-credentials-synthid-in-openai-generated-content)。
## 正文 · 原文
## Our approach to EU text provenance rules
Promoting transparency within the limits of today’s technology.
Loading…
[Audio 1](https://openai.com/index/eu-text-provenance)
Share
[Content provenance](https://openai.com/index/advancing-content-provenance/) helps people understand where content came from, how it was created or edited, and whether it contains signals associated with our models. We’ve already made tools publicly available to identify images and audio generated by our models. Today, we’re sharing our approach to text watermarking in response to the EU AI Act, and how it fits into our broader work.
The EU AI Act requires generative AI providers to make generated text identifiable in a machine-readable way. Text watermarking and detection remain early technologies with significant limitations, and views about their benefits and responsible uses are still developing. Our phased approach reflects both the EU AI Act requirements as well as the technology’s limitations, with an emphasis on transparency about what a text watermark can and cannot tell people:
- Starting today, API customers globally will be able to opt in to text watermarking for select models. Text watermarking will remain off by default in the API.
- Over the coming weeks, we will add an invisible watermark to eligible ChatGPT and Codex text output in the European Union.
- We’re opening applications to access our text watermark detector. Access will initially be limited to approved researchers and expert organizations that can help us evaluate and improve the technology.
- The above only applies to text provenance—our verification tools for audio and images, including our [openai.com/verify](https://openai.com/verify) web tool and our [Content Provenance API(opens in a new window)](https://developers.openai.com/api/docs/guides/content-provenance), will continue to be publicly accessible to organizations looking to understand whether an image or audio file was generated by one of our systems.
## How our watermarking works and performs
Our text watermarking technology, textGrain, adds an invisible statistical signal to the model’s word choices. Our detector looks for that signal to assess whether a passage contains an OpenAI watermark. More details about how textGrain works can be found in our [technical report(opens in a new window)](https://cdn.openai.com/pdf/e9508624-d767-41b6-a26d-e34ca798ada6/textgrain-entropy-calibrated-watermarking-for-language-model-text.pdf), which will be updated with additional details in the coming weeks. We also plan to make the technology available in open source so that others can build on it.
In our evaluations, textGrain matched or exceeded the performance of other approaches we tested, including SynthID for text. Even so, strong performance under ideal conditions does not guarantee reliable detection in everyday use.
Detectors can make two kinds of errors: they can report a watermark where none is present—a false positive—or miss a watermark that is present—a false negative. Our evaluations below illustrate some of the challenges:
- **Shorter or more constrained text is harder to detect.** At a target false positive rate of 1%, our detector identified watermarks in about 80% of 200-token passages, compared with about 95% of 400-token passages, for content such as psychology. Detection rates were substantially lower for content such as mathematics, where there is less flexibility in word choice.
- **Editing can weaken the watermark.** In an evaluation of 400-token passages, replacing 10% of words with synonyms reduced detection from about 92% to 66%. Replacing 25% of words reduced it to 17%.
These limitations contribute to our decision to provide initial detector access only to approved researchers and expert organizations, who can help us evaluate reliability and responsible uses.
This chart shows results for watermarked responses to mathematics and psychology questions from the [ELI5 dataset(opens in a new window)](https://huggingface.co/datasets/rexarski/eli5_category)at a target false positive rate of 1%. Detection improves with text length, but is substantially lower overall for content where there is less flexibility in word choice, such as mathematics.
Editing can substantially weaken the watermark signal. This chart shows how replacing 10% or 25% of the words in a passage affects detection. Results are based on watermarked English responses to questions from [ELI5(opens in a new window)](https://aclanthology.org/P19-1346/).
### Impact of watermarking on output quality
Across the benchmarks we use to assess Astra, our latest frontier model, we do not see meaningful performance differences with and without watermarking.
| **Benchmark** | **Unwatermarked text (Astra, max)** | **Watermarked text (Astra, max)** |
| --- | --- | --- |
| Artificial Analysis Intelligence Index | 49.57 points | 49.76 points |
| AutomationBench | 34.09% | 34.86% |
| DeepSWE v1.1 | 72.80% | 71.68% |
| Terminal-Bench 4.0 | 53.90% | 56.06% |
| Terminal-Bench Science 0.1 | 56.90% | 60.00% |
| BrowseComp | 87.92% | 87.35% |
| HealthBench Professional | 64.27% | 64.60% |
| GPQA Diamond | 94.44% | 93.94% |
## What a text watermark doesn’t tell you
Text watermarks provide a limited signal about the role our systems played in a passage. It’s important to understand what can and cannot be concluded from a detection result.
- **A watermark does not measure human contribution.** It can indicate that an OpenAI system generated or processed part of a passage, but not how much human judgment, editing, or creativity went into it.
- **A watermark does not establish ownership or responsibility.** It does not determine who owns the text, whether its use was lawful, whether disclosure was required, or who is responsible for it.
- **A watermark does not identify the user.** It does not associate a person, organization, account, prompt, or conversation with the text.
- **A watermark does not verify accuracy.** It does not tell you whether a passage is true, misleading, harmful, or presented in the right context.
- **The absence of a detected watermark does not prove human authorship.** Text generated with OpenAI tools may be too short, edited, or translated for detection to work reliably. It may also come from an unsupported model, predate watermarking, or have been generated by another company’s tools.
## Our next steps
- **Rolling out text watermarking in the EU.** Over the coming weeks, we will introduce text watermarking to eligible ChatGPT and Codex users across all plans in the EU only. We are not making text watermarking a global default at launch. This regional approach gives us room to learn from real-world use and feedback.
- **Enabling opt-in watermarking for API customers.** Starting today, API customers around the world will be able to opt in to watermarked text outputs for select models. This lets customers decide how watermarking fits their transparency obligations and the experiences they provide to users. We are also working with cloud partners to make watermarking available for OpenAI model outputs accessed through their services in the coming weeks.
- **Providing detector access to researchers and expert organizations.** Approved researchers and expert organizations can [apply](https://openai.com/form/content-provenance-api/) starting today. In accordance with the [Code of Practice(opens in a new window)](https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content), access will be initially granted on a case-by-case basis to support evaluation and improvement of text provenance. The tool will report whether it detects an OpenAI watermark, without identifying the user or revealing their prompts or conversations. Given the risk of missed watermarks and false positives, we are not making it publicly available at launch.
We expect to revisit each part of this approach as the technology, standards, and evidence evolve.
## Our broader approach to content provenance
No single provenance technique is enough on its own, so we take a layered approach that combines open standards, durable watermarking, and verification tools.
We add Content Credentials to supported image outputs, are [C2PA conformant(opens in a new window)](https://c2pa.org/conformance/), embed invisible [SynthID(opens in a new window)](https://deepmind.google/models/synthid/) watermarks in supported images and audio, and make image and audio verification available through [openai.com/verify](https://openai.com/verify) and our [Content Provenance API(opens in a new window)](https://developers.openai.com/api/docs/guides/content-provenance). These techniques complement one another: Content Credentials can record a file’s origin and history, while invisible watermarks can preserve a signal when metadata is removed.
As we described in our[election safeguards work](https://openai.com/index/election-safeguards-2026/), provenance can help people and platforms assess potentially misleading AI-generated content, including deepfakes. We pair these signals with policies, abuse detection, reporting, investigations, and enforcement, and work with researchers, standards bodies, platforms, and civil society to make provenance useful across the wider ecosystem.
Extending provenance to text requires accounting for how easily it can be rewritten, translated, or edited. As we discussed in[June](https://openai.com/index/supporting-eu-trustworthy-ai-ecosystem/), our approach must reflect the practical limits of current technology. We’ll continue improving detection, studying how watermarks withstand editing and translation, and exploring ways to distinguish AI assistance from AI authorship more meaningfully. We’ll adapt our approach as evidence, standards, and regulatory requirements evolve, and expand detector access when we believe results can be interpreted responsibly.
For more information, please visit our [help center article(opens in a new window)](https://help.openai.com/en/articles/8912793-provenance-signals-content-credentials-synthid-in-openai-generated-content).