Integrating user experience data into image algorithms to mitigate online harm

这项提议是数十年来学术研究和商业应用的巅峰之作，它在算法如何考虑最终用户体验方面做出了一步改变。图像具有不同程度的情感‘价态’，也就是唤起观者情感的能力。这种唤醒会让我们感到高兴、悲伤或愤怒，并因此导致大笑、哭泣或大喊大叫。传统上，这一点在视觉艺术中得到了利用，通过绘画、电影、电视、广告和宣传。与此同时，计算机仍然难以处理语义，这是嵌入在社交媒体图像内容馈送中的那种深刻而丰富的含义。不幸的是，这意味着许多为减轻在线危害而开发的算法工具同时删除了无害的内容，但却未能过滤出令人震惊的内容。我们在广告研究中知道“重复就是说服”，然而，随着时间的推移，消费图片的累积影响也是无法计算在内的。这导致了可怕后果的反复出现，因为算法在无意中自动升级了伤害，完全没有考虑到人为因素；在这种情况下，图像内容的情感价位及其系统性和网络化消费。此外，这些测试集通常由发展中国家的受剥削工人手动标记，因此他们也广泛暴露在最糟糕的互联网内容中。再一次有效地将个人和社会问题与业务问题一起离岸。这项建议是为了优化一个功能齐全的图像过滤工具原型，该工具考虑到用户在减轻在线危害方面的体验，自动识别高价有害内容。这将使该公司开始从服务转向产品，这是该地区数字经济中公认的瓶颈，并向面向私人投资的SaaS模式提供初步应用。与即将出台的在线危害法案相结合，该工具将能够为创意产业内容创作者提供有关其图像馈送中价位的量化数据，以减轻面对不断上升的负债和成本时不必要的社会和行为后果。

The culmination of decades of academic research and commercial application, this proposal offers a step change on how algorithms account for the end user experience. Images hold within them varying degrees of emotional 'valence', that is capacity to arouse feelings in the viewer. That arousal can make us feel happy, sad or angry and lead to laughing, crying or shouting as a result. This has traditionally been leveraged in the visual arts, through paintings, films, television, advertising and propaganda.Meanwhile, computers still struggle to process semantics, that is the kind of deep rich meaning embedded in social media image content feeds. Unfortunately, this means that many algorithmic tools developed to mitigate online harm simultaneously pull out innocuous content yet fail to filter out appalling material. We know in advertising research that 'repetition is persuasion', yet the cumulative effects of consumption of an image feed over time is also unaccounted for. This has lead to repeated instances of dire outcomes as algorithms unwittingly automate the escalation of harm, by failing entirely to account for human factors; in this case, the emotional valence of image content and its systematic and networked consumption. Furthermore, these test sets are often tagged manually by exploited workers in developing countries, who are thus extensively exposed to the worst of internet content too. Effectively once again, just offshoring the personal and social problem alongside the business one.This proposal is to optimise a fully functioning prototype image filter tool that accounts for users experience towards mitigating online harms, automatically recognising harmful content with high valence. It will allow the company to start the move from service to product, a well recognised bottleneck in the digital economy in the region, and deliver a preliminary application towards a SaaS model for private investment. Aligned with the forthcoming Online Harms Bill, the tool will be able to offer creative industries content creators with quantitative data on the level of valence in their image feeds to mitigate unwanted social and behavioural outcomes in the face of escalating liabilities and costs.