作战室里没有选民No Voters in the War Room
“作战室”这个词本身就是一次叙事升级。2018年 Meta 起这个名,是为了应对“选举相关的虚假信息扩散”,现在 OpenAI 和 Anthropic 各开一间,盯的是深度伪造视频和 AI 搜索结果。名字起得像打仗,打的却不是任何一方的仗——是平台自己的公关战:出了事,要有一个“快速、既定的分诊流程”。
注意它们在防什么。Anthropic 发言人的原话是“监测我们平台上的滥用”。滥用。不是“我们的模型会不会把错误信息包装得比真相更可信”,不是“用户想看到什么,AI 搜索就回给他什么”——报道自己写了这条:AI 搜索工具有把用户想看的答案反射回去的倾向。这条被写成“让工作更困难”的外部条件,可它恰恰是产品本身的设计后果。作战室防的是别人滥用平台,不防平台按用户口味生产现实。
报道里还有一个更老实的例子。上周 Google 的 AI Mode 拒绝回答“2020年大选谁赢了”,公司说正在修。一个连基本事实都答不出来的系统,被说成是“正在修复的故障”。可这个故障的方向值得记一下:不是答错,是拒答;修的方向是让它答,还是让它答得让某些人不生气?报道没写。谁的名字没出现,谁的声音缺席,这里就是材料。
加尔通的暴力三角里,这一层是文化暴力:制造“什么是事实”的解释权。平台把解释权攥在自己手里,再开一间屋子专门看守它,然后告诉公众,选举信息的准确性由我们来保障。可选举不是平台的事故,是公众的事。作战室里坐着的是公司的工程师和公关,没有一个选民,也没有一个被错误信息实际伤害到的人。
Katie Harbath 说作战室是“为了处理我们不一定能预测会出错的事”。这句话对,也不对。能预测的那部分——模型会迎合用户、搜索会反射偏好、拒答会被当成故障来修——是产品结构里长出来的,不需要作战室,需要改产品。作战室处理的是不能预测的那部分,也就是平台最不想为它负责的那部分。
所以这条新闻里的差额很清楚:Potential 是平台本可以让选举信息的流通更透明、更可审计、把模型的行为公开给公众检验;Actual 是两间不对外公开的屋子,和一个“快速响应”的承诺。差额就是暴力,只不过这次它穿着“我们很负责”的制服。
还有一件事报道没提:这两间作战室由谁监督,出错的判定标准由谁定,事后会不会公开。如果答案都是“平台自己”,那这间屋子守的不是选举,是平台在选举叙事里的定价权。
"War room" is itself an upgrade in narrative. Meta coined it in 2018 to handle the spread of election-related misinformation; now OpenAI and Anthropic each get one, watching for deepfake videos and AI search results. The name sounds like combat, but it isn't anyone's combat except the platform's own PR: if something goes wrong, there needs to be a "quick, established process in terms of triaging."
Look at what they're guarding against. Anthropic's spokesperson says the war room monitors "our platform for misuse." Misuse. Not "will our model package falsehood more convincingly than truth," not "AI search reflects back what users want to see" — the report states that last one itself: AI search tools tend to reflect back the results users want. It's framed as an external condition that makes the job harder, when it is precisely a consequence of the product's design. The war room guards against others misusing the platform; it does not guard against the platform manufacturing reality to taste.
There's a more honest example in the piece. Last week Google's AI Mode wouldn't answer who won the 2020 election, and the company said it was working to fix it. A system that can't answer a basic factual question is described as a malfunction under repair. Note the direction: not a wrong answer, a refusal; and the fix — is it to make it answer, or to make it answer in a way nobody gets angry about? The report doesn't say. Whose name is missing, whose voice is absent — that absence is the material.
In Galtung's triangle this sits at the cultural layer: control over what counts as fact. The platforms hold that power, open a room to guard it, and tell the public that election-information accuracy is now in their hands. But the election is not the platform's incident; it is the public's. The people in the war room are engineers and communications staff. No voter is in there, and no one actually harmed by misinformation either.
Katie Harbath says war rooms are "built for dealing with the things we can't necessarily predict going wrong." True and not true. The predictable part — models flatter users, search mirrors preferences, a refusal gets treated as a bug to fix — grows out of the product's structure. That part doesn't need a war room; it needs a product change. The war room handles the unpredictable part, which is the part the platform least wants to be responsible for.
So the gap here is clear. Potential: the platforms could make election information transparent, auditable, and open the model's behavior to public scrutiny. Actual: two closed rooms and a promise of quick response. The difference is violence, wearing a uniform that says "we take this seriously."
One thing the report leaves out: who oversees these rooms, who defines what counts as an error, and whether any of it will be disclosed afterward. If the answer is "the platform," then the room isn't guarding the election. It's guarding the platform's pricing power over the election narrative.