人工智能ai内容阅读_用人工智能打击非法内容

人工智能ai内容阅读

“As the amount of user-generated content that platform users upload, continues to accelerate, it’s become impossible to identify & remove harmful content using traditional human-led moderation approaches at the speed & scale necessary”

“随着平台用户上传的用户生成内容的数量继续增加，使用传统的人为主导的审核方法以必要的速度和规模来识别和删除有害内容已成为不可能”

Protecting users from never-ending harmful and illegal online content, has brought about massive global deliberation on this very crucial subject. A central element of this debate has shone a light on the capabilities and role of Al’s and Machine Learning’s automated approaches. These are engineered to provide users with stronger protection from potentially toxic material, via more efficient online content moderation.

保护用户免受永无止境的有害和非法在线内容的侵害，已经在这个非常关键的主题上进行了大规模的全球审议。 这场辩论的核心内容阐明了Al和机器学习的自动化方法的功能和作用。 这些产品旨在通过更有效的在线内容审核为用户提供更强大的保护，使其免受潜在有毒物质的侵害。

但这并不像挥舞魔杖那么容易 (But it’s not as easy as waving a magic wand)

There are multiple challenges. According to the 2019 Cambridge Consultant’s report FCOM, entitled: ‘Use of AI in Online Content Moderation’:

有很多挑战。 根据2019年Cambridge Consultant的报告 FCOM，题为：``在在线内容审核中使用AI''：

“Effective moderation of harmful online content is a challenging problem. While many of these challenges affect both human and automated moderation systems, some are especially challenging for AI-based automation systems to overcome. There is a broad range of content which is potentially harmful, including, but not limited to: child abuse material, violent and extreme content, hate speech, graphic content, sexual content, cruel and insensitive material and spam content.”

有效地管理有害在线内容是一个具有挑战性的问题。 尽管这些挑战中的许多都会影响人为和自动化的审核系统，但是对于基于AI的自动化系统而言，要克服这些挑战尤其具有挑战性。 有很多可能有害的内容，包括但不限于：虐待儿童材料，暴力和极端内容，仇恨言论，图形内容，性内容，残酷和不敏感的内容以及垃圾内容。”

擦！ (The rub!)

Ascertaining whether content is harmful, is not as easy as people may think. And while some content can simply be identified using analysis, other content requires an understanding of the context around it. — This makes matters perpetually challenging for both automated systems and humans, simply because it necessitates an extremely broad understanding of cultural, societal, political and historical factors.

确定内容是否有害，并不像人们想象的那么容易。 尽管可以通过分析轻松识别某些内容，但其他内容则需要了解其周围的环境。 －这对自动化系统和人类来说都是永恒的挑战，仅因为它需要对文化，社会，政治和历史因素有极其广泛的了解。

在盒子外面看 (Looking outside the box)

If we really consider this issue, we become mindful of the fact that certain contextual thoughtfulness is essential due to the diversity of countless countries; what societies regard as acceptable; and the differences in national laws. To that end, in order to be effective, content moderation procedures should be culturally-specific and contextually aware.

如果我们真的考虑这个问题，我们就会意识到一个事实，即由于无数国家的多样性，某些语境上的周密性至关重要。 什么社会认为可以接受； 以及国家法律的差异。 为此，为了有效，内容审核程序应具有针对特定文化和上下文的意识。

Content on the web is delivered in various different formats. Some of the latter are far more difficult to moderate and analyse than others. Video content is an excellent example: it necessitates the combining of image analysis across multiple frames, with audio analysis. Memes is another good illustration, as it demands compounding image and text analysis with cultural and contextual understanding. Deepfakes (a blend of “deep learning” and “fake”), refer to counterfeit media, in which someone in an existing video or image, is replaced with another person’s likeness. Deepfakes are produced using machine learning in order to create false, but nonetheless, persuasive text, audio, video and images. They have the potential to be extremely harmful. Moreover, at this present time, they are not easy to detect, either by artificial intelligence or humans.

网络上的内容以各种不同的格式交付。 后者中的某些要比其他人难得多。 视频内容是一个很好的例子：必须将跨多个帧的图像分析与音频分析相结合。 模因是另一个很好的例证，因为它要求将图像和文本分析与文化和上下文理解相结合。 深度假货(“深度学习”和“伪造”的混合)是指伪造媒体，其中将现有视频或图像中的某人替换为另一人的肖像。 深造是 使用机器学习制作的视频，目的是创建虚假但有说服力的文字，音频，视频和图像。 它们有可能造成极大危害。 此外，目前，无论是通过人工智能还是人类，都不容易检测到它们。

Check out this fascinating YouTube video: ‘Everybody Can Make Deepfakes Now!’

观看这个有趣的YouTube视频：“每个人都可以制作Deepfakes！”

颠覆审核制度 (Subverting the moderation systems)

Other content could involve a live text chat or live video stream, which as you will understand, has to be monitored and analysed in real time. Clearly, this form of content is more problematic, due to the fact that the level of harm can rapidly intensify; and only the current and previous content elements are accessible for consideration. As time goes on, the format and language of online content will evolve at the speed of light; and a percentage of users will endeavour to change moderation systems by using various tactics, such as making adjustments to the current phrases and the words. To that end, moderation systems need to be adapt at keeping up with such modifications.

其他内容可能涉及实时文本聊天或实时视频流，如您所知，必须对其进行实时监视和分析。 显然，由于伤害程度会Swift加剧，这种形式的内容更成问题。 并且只有当前和以前的内容元素可供访问。 随着时间的流逝，在线内容的格式和语言将以光速发展。 并且一定比例的用户将努力通过使用各种策略(例如对当前词组和单词进行调整)来更改审核系统。 为此，调节系统需要适应这种修改。

赶上22 (Catch 22)

“AI-enabled content moderation systems are developed to identify harmful content by following rules & interpreting many different examples of content which is & is not harmful”

“开发了支持AI的内容审核系统，可以通过遵循规则并解释许多有害的内容和有害的内容来识别有害内容”

In order to mitigate the risk of a company gaining a bad reputation due to their users witnessing harmful material, online platforms can continually moderate any third-party content which they host.

为了减轻由于用户目睹有害材料而导致公司声誉不佳的风险，在线平台可以持续审核其托管的任何第三方内容。

However, this can result in a catch 22 situation, as:

但是，这可能导致catch 22的情况，如下所示：

“removing content which is not universally agreed to be harmful, can also result in reputational damage and undermine users’ freedom of expression.”

“删除并非普遍认为有害的内容，还会导致声誉受损，并损害用户的表达自由。”

Moreover, when it comes to their community standards, it is not easy for online platforms to define the content and behaviours that are not permitted on their platforms. This is simply because there are various types of harmful content. Further, automated systems can have quite a challenge interpreting a platform’s community standards, regardless of whether the content is actually harmful or not.

此外，在涉及其社区标准时，在线平台很难定义其平台上不允许的内容和行为。 这仅仅是因为存在各种类型的有害内容。 此外，无论内容是否有害，自动化系统在解释平台社区标准方面都可能面临很大的挑战。

“Clarity in platforms’ community standards is essential to enable the development & refinement of AI systems, so that standards are consistently implemented”

“平台社区标准的清晰性对于实现AI系统的开发和完善至关重要，因此标准得以一致实施”

Generally speaking, at this stage of the game, being able to automate 100% effective content moderation, is mission impossible. Indeed, human moderators’ input will still be needed to appraise highly contextual, subtle content, for the foreseeable future. This means there will be a huge expense for companies, particularly as the volume of uploaded content will continue to shoot up.

一般来说，在游戏的这个阶段，要能够100％有效地进行内容审核自动化是不可能的。 实际上，在可预见的将来，仍然需要人类主持人的输入来评估高度上下文相关的微妙内容。 这意味着公司将付出巨大的代价，尤其是随着上载内容的数量将继续激增。

“AI can only help relieve some of the emotional toll of online content moderation by humans”

“人工智能只能帮助缓解人类在线内容审核带来的情感损失”

“Google and YouTube have content moderators working around the clock to scrub the platforms of violent and disturbing content. However, the job has some serious consequence.”

“ Google和YouTube的内容审核员全天候工作，以清理暴力和令人不安的内容平台。 但是，这项工作有一些严重的后果。”

Check out this CBS News YouTube Video: “Emotional toll of online content moderation”

观看此CBS新闻YouTube视频：“在线内容审核的情感代价”

AI正在加紧 (AI is stepping up)

Artificial Intelligence-based content moderation systems can help to cut down human moderators.

基于人工智能的内容审核系统可以帮助减少人工审核员。

Indeed, a large percentage of companies have an online content moderation workflow strategy which employs moderation at one or both points in the workflow:

实际上，很多公司都有在线内容审核工作流策略，该策略在工作流中的一个或两个点都采用了审核：

1: Pre-moderation (when the uploaded content is moderated before publication. This normally involves automated systems).

1：预审核(在发布之前审核上传的内容时，通常涉及自动化系统)。

2: Post or reactive-moderation (when the content is moderated post publication; when it has been flagged as potentially harmful by automated processes or other users; or content which had already been removed, but upon appeal, necessitates a second review.

2：发布后或被动审核(当发布后内容受到审核时；自动化流程或其他用户将其标记为可能有害的内容时；或者已被移除但已提出上诉的内容，则需要进行第二次审核)。

AI技术可能影响在线内容和内容审核工作流程的3种方式 (The 3 ways AI tech could impact online content & content moderation workflow)

Number 1: Artificial Intelligence can add value to the pre-moderation phase by flagging content than should be reviewed by staff, thus improving moderation accuracy. For example, useful tools comprise: fairly simple procedures including ‘hash matching’ (when an image fingerprint is compared with a database showing notable harmful images; as well as ‘keyword filtering’, in which words that signal possible harmful content are utilised to flag-up content. However, there are limitations.

数字1：人工智能可以通过标记内容(超出工作人员应审核的内容)来为审核之前的阶段增加价值，从而提高审核的准确性。 例如，有用的工具包括：相当简单的过程，包括“哈希匹配”(将图像指纹与显示显着有害图像的数据库进行比较时)以及“关键字过滤”，其中使用表示可能有害内容的单词进行标记内容，但是有局限性。

Sleuthing the subtleties of language, including the use of emojis, sarcasm and slang terms, is not easy. Moreover, languages, particularly slang terminology, evolve over time. Current tech research is putting a lot of weight on techniques for ‘sentiment analysis’ and ‘natural language understanding;’ and as such, these facets are becoming more efficient. In the same vein, there have been advancements in ‘scene understanding’ and ‘object detection,’ both of which are regarded as crucial aptitudes for moderating intricate video and image content. Of note, ‘recurrent neural networks’ is a particularly impressive AI approach. This is due to the fact it can empower a more refined video content analysis, as it is especially demanding to moderate, as video frames have to be relative to others.

掩盖语言的微妙之处，包括使用表情符号，讽刺和语，并不容易。 此外，语言(尤其是语)会随着时间而发展。 当前的技术研究非常重视“情感分析”和“自然语言理解”的技术； 因此，这些方面正变得越来越有效。 同样，“场景理解”和“对象检测”也有了进步，它们都被认为是调节复杂的视频和图像内容的关键能力。 值得注意的是，“循环神经网络”是一种特别令人印象深刻的AI方法。 这是因为它可以实现更精细的视频内容分析，因为视频帧必须相对于其他视频帧，因此特别要求适度。

“In practice, the most harmful content is generated by a minority of users, & so AI techniques can be used to identify malicious users & prioritise their content for review”

“实际上，最有害的内容是由少数用户生成的，因此AI技术可用于识别恶意用户并确定其内容的优先级以供审核”

These days, when it comes to using AI moderation techniques to check the sentiment around content, difficulties and complexities arise. ‘Metadata’ is able to encode certain context linked to content moderation decisions. This includes: data on the user’s true identity such as location or age; the number of followers or friends; and their history on the site. Of note, many online interactions have a historical and cultural context, and this can shed light on previous discussions and interactions from preceding content. On the downside, however, as the available metadata alters according to the type of content posted, and the actual platform; platform-agnostic moderation tools are not able to take full advantage of the metadata, thus making decisions quite difficult.

如今，在使用AI审核技术来检查围绕内容的情绪时，出现了困难和复杂性。 “元数据”能够对链接到内容审核决策的某些上下文进行编码。 这包括：有关用户真实身份的数据，例如位置或年龄； 追随者或朋友的数量； 以及他们在网站上的历史记录。 值得注意的是，许多在线互动都具有历史和文化背景，这可以阐明先前的讨论以及先前内容的互动。 但不利的一面是，由于可用的元数据根据发布的内容类型和实际平台而变化； 与平台无关的审核工具无法充分利用元数据，从而使决策非常困难。

Artificial Intelligence techniques including ‘scene understanding’ and ‘object detection’, are crucial automated systems components. This is because they can distinguish various levels of potentially toxic content, such as material on child abuse — something which primarily necessitates examining a video or image. Conversely, pinpointing content involving bullying, normally demands considering both the content itself, and the context of the user’s interactions. This is due to the fact that bullying content features are not so clearly defined. To that end, the challenges and costs to companies developing such specific AI moderating architecture, are substantially high.

包括“场景理解”和“对象检测”在内的人工智能技术是关键的自动化系统组件。 这是因为它们可以区分各种程度的潜在有毒成分，例如有关虐待儿童的资料，而这主要是必须检查视频或图像的内容。 相反，查明涉及欺凌的内容通常需要同时考虑内容本身和用户交互的上下文。 这是由于没有如此明确地定义欺凌内容功能。 为此，开发这种特定的AI调节架构的公司面临的挑战和成本很高。

Number 2: Implementing Artificial Intelligence can synthesise training data in order to ameliorate pre-moderation performance generative AI methods. These include GANs (generative adversarial networks), which can produce original and new text, audio, video and images. At its most rudimentary level, GANs denotes a system which pits two AI systems against each other in order to boost the quality of their results. However, such a concept can be employed to generate harmful content including violence and nudity. Indeed: “these images can supplement existing examples of harmful content when training an AI-based moderation system. Further, GANs can apply ‘style transfer’, changing the style of one image into another.” In addition, when it comes to developing Artificial Intelligence-based moderation systems: “generating more data out of a limited dataset, is a valuable approach for augmenting training datasets. This permits them to make more accurate decisions and reduces their dependence on the availability of large datasets containing suitably anonymised content.”

第2点：实施人工智能可以综合训练数据，以改善调节前性能生成的AI方法。 其中包括GAN(生成对抗网络)，可以生成原始和新的文本，音频，视频和图像。 在最基本的层次上，GANs表示将两个AI系统相互竞争以提高其结果质量的系统。 但是，可以使用这种概念来生成有害内容，包括暴力和裸露。 确实：“这些图像可以在训练基于AI的审核系统时补充有害内容的现有示例。 此外，GAN可以应用“样式转换”，将一个图像的样式更改为另一个图像。” 此外，在开发基于人工智能的审核系统时：“从有限的数据集中生成更多数据，是增强训练数据集的一种有价值的方法。 这使他们能够做出更准确的决策，并减少了对包含适当匿名内容的大型数据集可用性的依赖。”

Number 3: Artificial Intelligence can help human moderators step up their productivity levels. Moreover, the former can ameliorate the humans’ efficiency, by prioritising the content which needs to be reviewed, according to the degree of uncertainty from the automated moderation stage, or the grade of harmfulness detected within the content. Further, as AI provides top quality translations, it thus cut downs the challenge of moderating content in different languages.

第三点：人工智能可以帮助人类主持人提高生产力水平。 此外，前者可以根据自动审核阶段的不确定性程度或在内容中检测到的有害程度，通过对需要复查的内容进行优先级排序来改善人类的效率。 此外，由于AI提供了高质量的翻译，因此它减少了以不同语言审核内容的挑战。

中庸与言论自由？ (Moderation vs freedom of expression?)

With regard to controlling the massive quantity of online content: present-day policy arguments have revived the focus on certain proactive measures. These include: the automated detection of potentially illegal content, and upload filtering. Moreover, tech industry players and policy makers frequently put Artificial Intelligence forward as: “the solution to complex challenges around online content, promising that AI is a scant few years away from resolving everything from hate speech to harassment, to the spread of terrorist propaganda.” Yet, there is a very important piece of the jigsaw missing — and this is the recognition that: “proactive identification and automated removal of user-generated content raises problems beyond issues of ‘accuracy’ and over-breadth: problems that will not be solved with more sophisticated AI.”

关于控制大量在线内容：当前的政策论点使人们对某些主动措施的关注重新恢复。 其中包括：自动检测潜在的非法内容，以及上传过滤。 此外，科技行业的参与者和政策制定者经常将人工智能提出为：“解决在线内容周围复杂挑战的解决方案，并承诺人工智能距离解决仇恨言论，骚扰和恐怖宣传的方方面面还需要几年的时间。 ” 然而，拼图中仍然存在一个非常重要的部分-这是一种认识，即：“主动识别和自动删除用户生成的内容所带来的问题超出了“准确性”和过度使用的范围：这些问题无法解决拥有更先进的AI。”

The Joint Declaration of Freedom of Expression and the Internet, was issued way back in 2011, by the four International Mechanisms for Promoting Freedom of Expression.

2011年，四个促进言论自由的国际机制发布了《言论自由和互联网联合声明》。

This incorporated the statement that:

它包含以下声明：

“Content filtering systems imposed by a government or commercial service provider which are not end-user controlled, are a form of prior censorship and are not justifiable as a restriction on freedom of expression.”

“由政府或商业服务提供商实施的，不受最终用户控制的内容过滤系统是一种事先审查制度，不能作为限制言论自由的理由。”

These human rights law specialists were mindful of the substantial threat to free expression that was imposed by filtering; and although it has to be said that tech which automatically evaluates and detects content has become more advanced over time, such technical progress does not take into account many of the rudimentary issues of filtering.

这些人权法专家意识到过滤对自由表达的巨大威胁。 尽管必须说自动评估和检测内容的技术随着时间的推移变得越来越先进，但是这种技术进步并未考虑到许多基本的过滤问题。

The legal experts, note that:

法律专家请注意：

“Filtering acts as a prior restraint on speech, regardless of the ‘accuracy’ of the tool being used. [Moreover], content hosts that deploy filters must recognize the human rights risks inherent in their content moderation systems and work to mitigate them.”

“无论所使用工具的“准确性”如何，过滤都是对语音的事先限制。 [此外，部署过滤器的内容托管者必须认识到其内容审核系统固有的人权风险，并努力减轻这种风险。”

翻译自: https://medium.com/@marcinborecki/fighting-illegal-content-with-artificial-intelligence-82964f11ade3

人工智能ai内容阅读