Beyond the usefulness of artificial intelligence, it’s easy to draw a line in terms of its limitations; whilst AIs process language at a basic level, its intricacies are impossible to program into a machine.

Context, irony, humour and artistic licence are only a few aspects pertaining to the evolution of language (and humans in general); detecting them requires more than an immense vocabulary – it requires human intuition.

On social media, moderating a large volume of content is an impossible endeavour, especially in real time. Employing AIs for this gargantuan task might seem logical, in terms of spotting genuine threats, actual slurs directed at others or incitement to violence. Unfortunately, due to the ever-expanding scope of detecting unacceptable language (as deemed by each platform), false positives are very common, attracting sanctions based on misunderstandings. Among the most common are the following:

  • Quotes and references, interpreted as the user’s actual language or opinion;
  • Jokes and harmless banter;
  • Satire and sarcasm;
  • Cultural references which make no sense when taken ad litteram etc.

In certain cases, sanctions for hateful language were incurred by users whose message was the exact opposite of what the platform had deemed it to be through this system.

Though it might seem superfluous to mention, apps facilitating discussions are referred to as “social” media precisely because users seek to interact with other people, assuming the context or subtext will be understood as intended every time, at least by most.

The awareness of such moderation compels users to filter their language not only through the potential interpretations of others, but also through the limited one of an AI which takes everything literally. Undoubtedly, that makes self-expression more difficult or makes it off-putting to engage in conversation altogether.

In 2016, an app known as Candid took things further, declaring its purpose of training an AI to “detect emotions” (according to its CEO), followed by labelling users through the attitudes they were generally “found” by the AI to have. Oddly, it was promoted as a free speech app, where users could remain anonymous in every thread, while simultaneously feeding the AI with data enabling it to establish criteria for future censorship.

The issue many have taken with this phenomenon is not related to the protection of genuinely hateful speech; instead, it is consternation in the face of non-human interaction, the rules and results of which cannot be predicted. The fact that people have been targeted for quoting opinions they were criticising in real time, or for satire, is absurd and places them in an unnatural situation, worth avoiding altogether.