What do the sources establish?
| Platform / Source | Core Stance / Quote | Engagement Metric | Timestamp |
|---|---|---|---|
| Pangram expands its AI detection tools to video and images | Reported context: [object Object] [object Object] [object Object] [object Object] | Not recorded | |
| Pangram releases Image 1.0 tool to sniff fake pics and videos | Reported context: [object Object] [object Object] [object Object] [object Object] | Not recorded |
THE SHORT VERSION
- Pangram, known for detecting AI-written text, has released image and video detection to subscribers, including a Chrome extension that flags synthetic pictures while people browse.以检测 AI 生成文字起家的 Pangram，现已向订阅用户开放图片与视频检测，并推出可在浏览网页时自动标记合成图片的 Chrome 扩展。
- The company says its video model exceeds 98 percent accuracy and its image model reaches 99.96 percent on its own evaluation set, with a false positive rate of 0.27 percent in testing.公司称其视频模型准确率超过 98%，图片模型在自家评测集上达到 99.96%，测试中误报率为 0.27%。
- Both tools carry limits the company states itself: results below 512 by 512 pixels come with a reduced-confidence warning, and phone camera postprocessing makes real photos harder to clear.两项工具都带着公司自己承认的限制：小于 512×512 像素的结果会附带置信度降低的提示，手机相机的后期处理也会让真实照片更难被判为真。
From text to pictures and clips从文字走向图片与片段
Pangram, which built one of the most widely used detectors for AI-written text, has released tools that flag AI-generated images and video. Subscribers can now upload a clip and get an assessment of whether AI was used to create or edit it. The image model that spent months as a research preview is generally available. Both run on one subscription, with a free tier for a few images.以检测 AI 生成文字起家的 Pangram，如今也推出了标记 AI 生成图片与视频的工具。订阅用户可以直接上传一段视频，得到“是否用 AI 生成或修改过”的判断。此前只作为研究预览开放的图片模型，现在已正式可用。两项功能同属一份订阅，免费账户仍可查少量图片。
The move matters because text was the easy case. A paragraph carries statistical fingerprints that a classifier can learn, while a picture or a clip has to be judged on pixels and frames. Pangram is arguing that the same approach works across all three, and that buyers of detection should not need a separate vendor for each medium. Text detection also has a longer track record behind it.这件事重要，是因为文字本来是最容易的一类。段落里带有分类器可以学习的统计指纹，而一张图片或一段视频只能靠像素与帧来判断。Pangram 的主张是，同一套方法对三种媒介都适用，需要做检测的买家不必为每种媒介另找一家供应商。
What the video check reports视频检测到底报什么
Engadget reports that the video tool distinguishes clips that are partially AI from those that are entirely generated, and that it produces a heatmap showing which parts of a frame are likely to be synthetic. That distinction is the useful part for editors, because most suspect footage now mixes real camera work with generated inserts rather than being fake end to end.Engadget 报道称，视频工具能区分“部分由 AI 生成”和“完全由 AI 生成”的片段，并生成一张热力图，标出画面中哪些区域可能是合成的。对编辑来说，这个区分才是有用的部分，因为如今可疑素材大多是真实拍摄夹杂生成片段，而不是从头到尾都是假的。
Pangram says the model behind it particularly excels at frontier AI video detection, and that one internal test put accuracy above 98 percent. Digital Trends reports the company's claim of 98.7 percent on the DF26 benchmark, a 29-point lead over the next best detector. Internal tests belong to the vendor, and no outside replication was published alongside the release. That is a wide margin to accept on the company's word.Pangram 表示，支撑该功能的模型“尤其擅长前沿 AI 视频的检测”，其中一次内部测试的准确率超过 98%。Digital Trends 报道称，公司声称在 DF26 基准上达到 98.7%，领先第二名 29 个百分点。内部测试属于厂商自己，发布时并没有附带任何外部复现。视频片段比文字更难验证，因为观者只能凭借画面本身来判断。
Where the accuracy numbers come from准确率数字的来源
The image claims are larger. Digital Trends reports that Pangram says its image detector identifies AI-generated pictures with 99.96 percent accuracy on its own evaluation set, while Engadget reports a false positive rate of 0.27 percent with more than 99 percent of AI images caught. Both figures come from the same internal testing, and the evaluation set is not public. Independent replication is the missing piece in both cases.图片那边的说法更大。Digital Trends 报道，Pangram 称其图片检测器在自家评测集上识别 AI 生成图片的准确率为 99.96%；Engadget 则报道，测试中误报率为 0.27%，同时捕获了超过 99% 的 AI 图片。两组数字都来自同一批内部测试，而评测集并未公开。换句话说，这些数字目前只能当作厂商声明来读。
Accuracy figures for detectors are notoriously sensitive to the benchmark. A set built from the newest generators favours whichever model trained on them, and a set of compressed social media images is harder than a set of clean exports. Neither source article says how the evaluation set was assembled or how old its samples are. A detector is only as good as the set it was measured on.检测器的准确率对基准极其敏感。用最新生成模型搭建的测试集，会偏向在它们上面训练过的模型；一堆经过压缩的社交媒体图片，也比干净的导出文件更难判断。两篇源报道都没有说明评测集如何构建，也没有说明样本有多旧。
Image detection leaves preview图片检测走出预览阶段
Pangram Image 1.0 is now on the web, in the company's API and in its Chrome extension, which scans images as a user browses and flags the ones likely to be AI. Digital Trends reports that the extension marks content on social platforms, shopping sites and rental listings, the places where a fake photo does immediate commercial damage to whoever trusts it.Pangram Image 1.0 现在可在网页端、公司 API 以及 Chrome 扩展中使用；扩展会在用户浏览时扫描图片，标出可能由 AI 生成的那些。Digital Trends 报道称，该扩展会在社交平台、购物网站与租房信息页上标记内容，而这些正是假照片会立刻造成商业损失的地方。这类页面上的图片，往往早已被压缩过许多次。
The company also lowered the minimum size it will judge. Digital Trends reports that images as small as 256 by 256 pixels can now be checked, down from a 512-pixel floor at launch, though results below 512 pixels carry a reduced-confidence warning. Small thumbnails are exactly what most people forward, so the limit matters in practice. That floor was a launch limitation rather than a technical ceiling.公司还降低了可判断的最小尺寸。Digital Trends 报道，现在可以检查小到 256×256 像素的图片，低于发布时 512 像素的门槛，不过小于 512 像素的结果会附带置信度降低的提示。小尺寸缩略图恰恰是大多数人转发的东西，所以这条限制在实际使用中很关键。
The limits the company admits公司自己承认的限制
Engadget reports a caveat the company states itself: below 512 by 512 pixels, results come with a reduced-confidence disclaimer. The outlet also notes that an image shared by chief executive Max Spero showed an unsure result on two pictures that were obviously not genuine photographs, which is a fair illustration of how the tool behaves at the edges. An unsure result is still a result that somebody will act on.Engadget 报道了公司自己说明的一项限制：小于 512×512 像素时，结果会附带置信度降低的免责提示。该媒体还提到，首席执行官 Max Spero 分享的一张示例中，扩展对两张明显不是真实照片的图片给出了“不确定”的结果，这倒很能说明工具在边界上的表现。
Spero said phone cameras add their own AI postprocessing to real photos, and that Pangram had to build a model robust to that while still separating AI edits from fully generated visuals. Engadget reports he told Pirate Wires that the image release was delayed to test on several iPhone 18 Pros after the model flagged their camera output. Camera output is now the hardest honest case for any detector.Spero 表示，手机相机本身会对真实照片做一定程度的 AI 后期处理，Pangram 因此必须构建一个对这种处理具有鲁棒性的模型，同时还要区分“AI 修图”与“完全生成”。Engadget 报道，他告诉 Pirate Wires，图片模型的发布之所以推迟，是因为模型起初把几台 iPhone 18 Pro 拍的照片判成了 AI。
The comparison table and its gaps对比表与它的空隙
The comparison table is where buyers will look. Digital Trends reports TruthScan at 96.33 percent on images, AI or Not at 91.38 percent and Sightengine at 88.33 percent, against Pangram's claimed 99.96 percent. On video it reports TruthScan at 84.2 percent and Sightengine at 81.8 percent, against Pangram's 98.7 percent. Every row is a vendor number. None of the three competitors was invited to confirm its own score.买家会先看对比表。Digital Trends 报道，图片方面 TruthScan 为 96.33%，AI or Not 为 91.38%，Sightengine 为 88.33%，而 Pangram 自称 99.96%。视频方面，该媒体给出的 TruthScan 是 84.2%，Sightengine 是 81.8%，Pangram 则是 98.7%。每一行都是厂商自己的数字。三家对手都没有被邀请确认自己的分数，表格由一家公司整理，排序自然也对它有利；这也是为什么这类对比更该被当作声明来读，而不是测量结果。
Gaps that wide invite suspicion, and the honest reading is that each company is quoting its own benchmark. Pangram says it also compared its detectors against models built by academic researchers and consistently outperformed them. Independent tests, run by someone with no product to sell, are the only way to settle the question. Until then the table is marketing with decimal points attached.差距这么大，反而让人起疑，最诚实的解读是：每家公司引用的都是自己的基准。Pangram 表示，它还把自己的检测器与学术研究者构建的模型做了比较，并持续胜出。真正的答案只能来自独立测试，由没有产品要卖的人来跑，而在这之前，表格只是带着小数点的营销材料。
What to watch next接下来看什么
The first question is whether outside evaluators reproduce the numbers, and the second is what happens when a detector is wrong in public. A false positive on a real photograph is a reputational problem for the person accused, and a false negative spreads a fake. Neither error is priced into a benchmark score. Both are cheap to make and expensive to correct.第一个问题是外部评测者能否复现这些数字，第二个是当检测器在公众面前出错时会发生什么。把真实照片误判为假，对当事人是名誉问题；把假图放过，则会让伪造内容扩散。这两种错误都不会体现在基准分数里。而两者都很容易发生，纠正起来却很贵。
The third is distribution. A Chrome extension that flags images while people shop puts detection in front of users who never asked for it, which is a different business from selling checks to newsrooms. Newsrooms pay for accuracy; shoppers pay for a warning they can ignore. This site has followed how detectors change writing and the shopping agents retailers now block; generated video is the next surface.第三个是分发。一个在人们购物时顺手标图的 Chrome 扩展，把检测推到了从未主动要求它的用户面前，这与其说是卖给新闻编辑室的生意，不如说是另一种生意。本站此前追踪过检测器如何改变写作与零售商开始封堵的购物智能体，而生成式视频就是下一个战场。
Sources & reporting limits来源与核查范围
Follow Products to find related stories on your next visit.关注产品新闻，下次访问时阅读相关文章。