一句提示词挡不住抄袭A Prompt Line Won't Stop Cloning
Anthropic 往 Claude Code 的提示词里加了两句话,禁止复刻现有游戏,官方确认属实。这不算治理,只是给一台没有闸门的机器贴了张纸条。
真正的动作在别处:独立开发者取消公开试玩。《Drunken Goddess Reflux》在 Steam 上有两万五千个愿望单,几千人等着在十月的 Next Fest 试玩,开发方 Alliance Arts 和 Who You 直接退出;Digital Sun 把《ReVamp》的公开试玩改成 Discord 上的受控测试。这就是差额被看见的地方——本来可以公开、可以让几千个真玩家上手的东西,现在要收进小房间,因为隔壁有人会把它整个搬走。这不是游戏行业的创作问题,是 Violence = Potential − Actual 里那个 Actual 被硬生生压低了一截。
压低它的不是 Claude。Claude 是工具,工具不承担道德,承担道德的是按下按钮的人和收钱的公司。Anthropic 的措辞值得看一眼:「用 Claude 侵犯 IP 违反我们的使用政策」。重点是「我们的使用政策」,不是法律,不是开发者的权利。翻译过来是:这件事让我们有合规风险,所以我们把它从提示词里挡掉。至于怎么判断「原创素材」和「抄来的素材」,报道说规则不明,被误判的用户要联系工作人员才能继续干活——一个连判定标准都不公开的机制,谈不上保护谁,只是把责任推回给被抄的人去自证。
更根本的是,这条规则写在提示词里,不写在能力里。报道自己说了:这类安全限制可以被绕过,vibe-coder 转头就能去用开放权重的模型,那些模型的护栏一拆就掉。于是这条更新的实际效果,是把抄袭从 Anthropic 的服务器上赶出去,赶到不需要为抄袭负责的地方。这是文化暴力最标准的操作:制造一个「我们在管」的叙事,让监管看起来发生了,而入口一个没堵。
谁在这场博弈里赢了?Anthropic 赢了公关,vibe-coder 换个模型继续赢,被抄的独立开发者赢了什么?赢到一句写在别人系统提示词里的话,和一次被迫取消的公开试玩。
那两句话不是没有价值——至少它承认了复刻别人的游戏是个问题。但承认问题和堵住入口是两件事。真正的入口在模型能力和分发平台上:Steam 让 AI 游戏淹没 Next Fest 的时候做了什么,itch.io 两个月翻倍的页面谁来审,开放权重模型的护栏由谁负责。这些名字在这条新闻里一个都没出现。
Anthropic added two sentences to Claude Code's prompt barring users from recreating existing games. The company confirmed it. This is not governance. It is a note taped to a machine that has no gate.
The real action is elsewhere: indie developers canceling public playtests. Drunken Goddess Reflux had 25,000 Steam wishlists, thousands of fans waiting to try the demo during October's Next Fest, and its developers Alliance Arts and Who You pulled out entirely. Digital Sun moved ReVamp's public playtest to a controlled test on Discord. That is where the gap becomes visible. Something that could have been public, that thousands of real players could have touched, now has to be moved into a small room, because someone next door will carry the whole thing away. This is not a creativity problem in the games industry. It is the Actual in Violence = Potential − Actual being pushed down by force.
What pushes it down is not Claude. Claude is a tool. Tools carry no morals; the people who press the button and the company that gets paid do. Look at Anthropic's wording: using Claude to infringe IP violates 「our usage policy.」 The emphasis is on our usage policy — not the law, not the developers' rights. Translated: this creates compliance risk for us, so we block it at the prompt. As for how it tells original assets from copied ones, the report says the guidelines are unknown, and a user who believes they have been wrongly accused must contact a staff member to clear things up and keep working. A mechanism that does not publish its own criteria is not protecting anyone. It hands the burden back to the person being copied, to prove they are innocent.
More fundamentally, the rule lives in the prompt, not in the capability. The report says it outright: safety limits on LLMs can be circumvented, and vibe-coders can simply turn to open-weight models whose guardrails are easy to remove. So the actual effect of this update is to push copying off Anthropic's servers and toward places that answer to no one for it. This is cultural violence in its standard form: manufacture the narrative that someone is minding the store, let regulation appear to happen, and leave every entrance open.
Who wins this game? Anthropic wins the press release. Vibe-coders switch models and keep winning. What do the copied indie developers win? A sentence in someone else's system prompt, and a public playtest they had to cancel.
The two sentences are not worthless — at least they admit that recreating someone's game is a problem. But admitting a problem and closing an entrance are two different things. The real entrances are model capability and distribution platforms: what Steam did while AI games flooded Next Fest, who reviews the doubled pages on itch.io, who is responsible for the guardrails on open-weight models. None of those names appear in this story.