AI监管

字节与 MPA 签署版权协议 国内智能体数字护照落地

ByteDance-MPA Copyright Framework China's Agent Digital Passports

从起诉到合作,好莱坞与 AI 公司的第一份版权框架协议。从内容到行为,智能体监管开始管 AI 的「手」了。

From lawsuits to cooperation — Hollywood's first AI copyright framework deal. From content to behavior — agent regulation starts policing what AI actually does.

No.039 2026年8月20日 约 5 分钟阅读 ~5 min read

AI 版权和智能体监管,在同一天迎来了两个标志性事件。

一个在西边:8 月 17 日,字节跳动与美国电影协会(MPA)签署谅解备忘录,就 Seedance 视频和 Seedream 图像生成模型的知识产权保护建立合作框架。这是好莱坞与 AI 公司之间的第一份版权"休战协议"。

一个在东边:国内多部门密集出台智能体监管文件,提出要给 AI Agent 发"数字护照"——全生命周期身份登记,完整留存推理和决策日志,厘清责任主体。

一西一东,指向同一个方向:AI 从"野蛮生长"进入"有法可依"的时代

从起诉到合作:版权博弈的转折点

字节和 MPA 的这份协议,结束了双方长达半年的对峙。

今年 2 月,Seedance 2.0 推出后在好莱坞炸开了锅——用户用它生成好莱坞明星打斗视频、知名电影角色的二创内容,在 TikTok 上疯传。MPA 代表迪士尼、派拉蒙、索尼、环球、华纳兄弟等片商,向字节跳动发出停止侵权函。迪士尼更是直接指责 Seedance 涉及"星球大战""漫威"等旗下角色。

半年后,双方从对簿公堂的边缘,走到了谈判桌前。虽然协议的具体技术细节和财务条款没有披露,但从"对峙"到"合作"的姿态转变,本身就是行业的标志性事件。

这意味着什么?意味着 AI 公司和版权方都意识到,打官司不是出路,合作才是。完全禁止 AI 用版权内容训练,不现实,也挡不住技术进步。但完全放任 AI 随便用,又会摧毁内容产业的根基。中间一定有一条路——在训练、分发、溯源三个层面,划出清晰的边界。

字节的思路很明确:一边做 IP 保护机制,一边做商业化合作。比如 6 月推出的 AI 版权商业化平台,让周星驰经典电影桥段可以被用户基于官方授权模板进行二创。既保护了版权,又释放了 AI 创作的活力。

智能体"数字护照":监管从内容到行为

如果说版权监管管的是"AI 生成了什么",那智能体监管管的就是"AI 做了什么"。

国内最近出台的一系列文件,核心逻辑只有一个:让智能体的行为可追溯、可问责。具体做法包括:

把"可控性、算法正当程序"法定嵌入智能体底层架构(Harness),要求完整留存每一轮推理、决策、执行的全流程日志。探索智能体"数字护照"——依托全生命周期身份登记,厘清单智能体与多智能体协同场景下的责任主体。

说白话就是:AI 替你下的单、调的接口、发的消息,以后都得留痕可查。出了问题,能查到是哪个智能体干的、谁让它干的、中间决策过程是什么。不能再说一句"是 AI 干的"就完事了。

这个方向非常关键。因为以前的 AI 监管,大多停留在"内容层面"——生成的文本合不合规、图片有没有问题。但 Agent 不一样,它不只是生成内容,它会执行操作。会发邮件、会订机票、会转账、会控制物理设备。行为的后果比内容严重得多。

如果说过去的 AI 监管是管"嘴",那现在的监管开始管"手"了。

合规会成为新的护城河

监管收紧听起来是个坏消息,但对行业头部玩家来说,可能反而是好事。

因为合规是有成本的。要做水印、要做溯源、要做日志留存、要做责任划分,这些都需要投入技术资源和法务资源。小公司可能做不起,但大公司可以。监管越严,头部玩家的合规优势就越明显。

以版权为例,字节能和 MPA 签协议,是因为字节有体量、有技术、有商业合作的基础。小的 AI 视频创业公司呢?它们既拿不到 IP 授权,也做不起版权防护系统,在监管收紧的环境下,生存空间会越来越小。

智能体监管也是一样。要做"数字护照"、要留全流程日志,这些都会拉高 Agent 产品的门槛。那些"套个壳就叫 Agent"的产品,会最先被淘汰。

所以别只看到监管带来的限制,也要看到它带来的行业洗牌。每一次监管收紧,都是头部玩家的机会

AI 行业的青春期快结束了。成年的标志就是——不能再想干什么就干什么了。

明天见。

AI copyright and agent regulation hit two landmark milestones on the same day — one in the West, one in the East.

In the West: on August 17, ByteDance and the Motion Picture Association (MPA) signed a memorandum of understanding establishing a framework for IP protection across the Seedance video and Seedream image generation models. It's the first copyright "truce" between Hollywood and an AI company.

In the East: Chinese regulators issued a flurry of new agent oversight documents, proposing "digital passports" for AI agents — full lifecycle identity registration, complete reasoning and decision logs, and clear accountability frameworks.

One west, one east — both pointing in the same direction: the AI industry is moving from "wild growth" into a regulated era.

From Lawsuits to Cooperation: A Copyright Turning Point

The ByteDance-MPA deal ends six months of confrontation between the two sides.

Back in February, Seedance 2.0 exploded onto the scene and caused an uproar in Hollywood — users were generating videos of Hollywood stars in fight scenes, derivative content of famous movie characters, and it was all going viral on TikTok. The MPA, representing Disney, Paramount, Sony, Universal, Warner Bros. and others, sent cease-and-desist letters to ByteDance. Disney specifically called out Seedance for infringing on Star Wars, Marvel, and other flagship IP.

Six months later, both sides have moved from the brink of litigation to the negotiating table. The specific technical details and financial terms of the agreement haven't been disclosed, but the shift from confrontation to cooperation is itself an industry landmark.

What does this mean? It means both AI companies and copyright holders have realized that fighting in court isn't the way forward — cooperation is. Completely banning AI from training on copyrighted content isn't realistic and won't stop technological progress. But letting AI use anything it wants would destroy the content industry's foundation. There must be a middle path — drawing clear boundaries at three levels: training, distribution, and provenance.

ByteDance's approach is clear: build IP protection mechanisms on one side, and commercialization partnerships on the other. For example, the AI copyright commercialization platform launched in June lets users create derivative works from officially licensed templates — like Stephen Chow's classic movie scenes. It both protects copyright and unlocks the creative energy of AI.

Agent "Digital Passports": Regulating What AI Does

If copyright regulation governs "what AI generates," agent regulation governs "what AI does."

The core logic of the recent Chinese regulatory documents is singular: make agent behavior traceable and accountable. Specific measures include:

Legally mandating "controllability and algorithmic due process" in the underlying agent architecture (the Harness layer), requiring complete logs of every round of reasoning, decision-making, and execution. And exploring agent "digital passports" — full lifecycle identity registration to clarify responsibility across single-agent and multi-agent scenarios.

In plain language: the orders an AI places, the APIs it calls, the messages it sends — all of it must be logged and auditable. When something goes wrong, you can trace which agent did it, who instructed it, and what its decision process looked like. You can't just say "the AI did it" and walk away.

This direction is critical. Previous AI regulation mostly stopped at the "content layer" — is the generated text compliant, are the images problematic? But agents are different. They don't just generate content — they execute actions. They send emails, book flights, transfer money, control physical devices. The consequences of behavior are far more serious than those of content.

If past AI regulation was policing AI's "mouth," the new regulation is starting to police AI's "hands."

Compliance Becomes the New Moat

Tighter regulation sounds like bad news, but for industry leaders, it might actually be an opportunity.

Because compliance costs money. Watermarking, provenance tracking, log retention, liability frameworks — all require technical and legal resources. Small companies might not afford it; large ones can. The stricter the regulation, the bigger the compliance advantage for incumbents.

Take copyright as an example. ByteDance can strike a deal with the MPA because it has scale, technology, and a foundation for commercial partnership. What about small AI video startups? They can't get IP licenses and they can't afford copyright protection systems. As regulation tightens, their survival space shrinks.

Same with agent regulation. Implementing "digital passports" and retaining full process logs raises the bar for agent products. Those "slap a wrapper on it and call it an agent" products will be the first to go.

So don't just see the constraints of regulation — see the industry shakeout it brings. Every regulatory tightening is an opportunity for the top players.

The AI industry's adolescence is ending. The sign of adulthood? You can't just do whatever you want anymore.

See you tomorrow.

Sources · 信源 Sources