Skip to content

自行核实

如何识别深度伪造

入门阅读时间 8 分钟机器翻译

本页为机器翻译,正在等待人工审校。英文版本为正式记录版本。 阅读英文正式记录版本

You usually cannot spot a good deepfake by staring at it — detection by eye is losing to the technology. The reliable method is to check a clip's provenance before its content: where it came from, when it appeared, and whether any original source exists. Suspicion of timing beats scrutiny of pixels.

本文内容

不要再试图凭肉眼识别

关于深度伪造最常见的建议——留意异常眨眼、数手指、观察牙齿——正悄然变得过时。这些特征只能识别粗糙的伪造,但技术每个月都在进步,而在选举中真正重要的伪造早已超越了这些特征。最清楚的证明是:迄今为止最有效的政治深度伪造是音频,不是视频,而音频根本没有手指、牙齿或眨眼可供检查。

因此,首先要放弃这样一种假设:你可以与现代伪造进行一场盯视较量。通常你做不到,而尝试这样做只会培养错误的自信。好消息是,你很少需要这样做,因为伪造最薄弱的环节几乎从来不是媒体本身——而是媒体周围的一切。这与如何核查任何主张的教训相同:判断来源和语境,而不只是判断载体。

来源比像素更重要

可靠的问题不是“这看起来真实吗?”,而是“它来自哪里?”深度伪造可以完美伪造面孔和声音;但它无法伪造可信、可追溯的来源。真实事件会留下痕迹——完整录音、新闻活动、以及愿意为其信誉背书的媒体报道。伪造内容通常以短小、裁剪过、低质量的片段形式出现,来自一个你从未听说过的账号,背后没有任何原始材料。

核查来源并不吸引人——搜索最早版本、寻找一手来源、查看还有谁在报道——但这才是真正有效的方法,而且无论像素或波形多么逼真,它都有效。

时间陷阱

选举中的深度伪造之所以被武器化,更多是靠时机而非质量。其套路是在最后时刻发布伪造内容——理想情况下是在投票前的最后几天,这样就没有时间辟谣;而在一些国家,竞选静默期还会在法律上限制最有能力回应的人发声。

Slovakia 2023年议会选举就是一个参考案例。大约在投票开始前48小时,伪造音频声称捕捉到一名政党领袖与一名记者讨论操纵选票,在选前禁宣期内传播开来,而此时反驳受到限制。该政党最终以微弱差距落败。诚实的结论不是伪造内容决定了选举——这一点无法证明——而是它的时机本身就是全部设计。因此,一个实用规则是:当一段片段最具指控性、并且恰好出现在最关键时刻时,尤其要保持怀疑。

视觉与音频特征(以及它们为何正在消失)

谨慎使用时,旧有迹象对低质量伪造仍有一定价值:

MediumPossible tellCaveat
Video不自然的眨眼、异常的手部/牙齿、发际线处模糊、口型同步漂移在高质量伪造中正迅速消失
Video光线或阴影与场景不符现在很容易做对
Audio情绪平淡、呼吸异常、机械式节奏常被低质量所掩盖
Any过短、裁剪过度、缺乏语境最持久的信号

有两点警告。第一,出现某种特征提示可能是伪造,但没有特征并不能证明任何事——最好的伪造往往没有这些特征。第二,不要落入镜像陷阱:深度伪造的存在会让人们把真实录音也当作伪造来否定。这种反射式判断——说谎者红利——意味着“这大概是深度伪造”对你或对被录下的人都不是免死金牌。

工具有帮助,但不能作出决定

自动化检测工具确实存在,WITNESS 等组织也推动更广泛地获取这些工具。它们值得使用——但只能作为提示,不能作为裁决。检测能力始终落后于生成能力,而且这些工具会产生假阳性和假阴性,因此单一工具的输出应与来源和语境并列参考,绝不能取代它们。

真正持久的防线是选举中的深度伪造所讨论的人为防线:预先辟谣,让受众预期到最后时刻的伪造;快速、可信的反驳;以及在问“这看起来真实吗?”之前,先问“它来自哪里,为什么是现在?”的习惯。把这个顺序弄对,你就已经领先于伪造——而不必与单个像素进行盯视较量。

Checking a suspicious clip, step by step

  1. 1

    Pause before you react

    A clip built to go viral is built to bypass your judgement with a jolt of shock or anger. The moment you notice that jolt, stop. Do not share it while you are still feeling it — that reaction is the delivery mechanism, not evidence.

  2. 2

    Check where it came from

    Find the earliest version you can. Who posted it first, from what account, and when? A dramatic clip that surfaces from an anonymous or brand-new account, with no traceable origin, should be treated as unverified until proven otherwise.

  3. 3

    Look for a primary source

    If a public figure really said or did this, there should be an original — a full video, a press event, a reputable report. Search for it. A clip that exists only as a short, cropped, low-quality fragment with no original behind it is a warning sign.

  4. 4

    Cross-check trusted outlets

    See whether independent, credible outlets are reporting the same thing. A genuine scandal caught on tape gets covered; a fake often lives only in partisan or fringe channels. Silence from serious sources is meaningful.

  5. 5

    Weigh the timing and the beneficiary

    Ask when it appeared and who gains. A bombshell dropped in the final days before a vote — especially inside a campaign-silence period — is a classic pattern designed to spread faster than it can be debunked.

  6. 6

    Use detection tools as a hint, not a verdict

    Deepfake-detection tools can flag suspicious media, but they lag the technology and produce false results in both directions. Treat a tool's output as one more clue alongside provenance and context — never as the final word.

常见问题

Can you tell a deepfake just by looking at it?

Increasingly, no. Good deepfakes no longer have obvious visual glitches, and deepfake audio — the most effective kind in elections so far — has no visual tells at all. Relying on your eyes is a losing strategy; checking where a clip came from is far more reliable than scrutinising how it looks.

What are the visual signs of a fake video?

Historically: unnatural blinking, odd hands or teeth, mismatched lighting, blurring where the face meets the hair, and lips slightly out of sync with audio. These still catch weak fakes, but the tells are disappearing fast — and their absence does not prove a clip is real.

Why is deepfake audio so hard to detect?

Audio has no faces, hands, or lighting to give it away, generating a convincing voice needs only minutes of real speech, and a deliberately phone-quality clip makes any flaws sound like a bad connection. It also spreads through private voice notes where no fact-checker can see it.

What should make me most suspicious that a clip is fake?

Timing and origin. A shocking clip that appears right before an election, in low quality, from an unclear source, and that just happens to damage one side, fits the deepfake playbook exactly — as the Slovakia 2023 case showed. Be most careful precisely when a clip feels most damning.

来源

  1. 1.Prepare, Don't Panic: Synthetic Media and DeepfakesWITNESS Media Lab
  2. 2.Slovakia: Deepfake audio of Denník N journalist offers worrying example of AI abuseInternational Press Institute (IPI) (2023)
  3. 3.Beyond the deepfake hype: AI, democracy, and 'the Slovak case'Harvard Kennedy School Misinformation Review (2024)
撰写者 VoteRights Editorial审阅者 VoteRights standards desk最后审阅

继续学习

数据库中的相关案例

分享链接会打开该网络自身的撰写器。此处不嵌入任何内容,因此不会有第三方得知您阅读了哪些页面。