Algorithms meant to spot hate speech online are far more likely to label tweets “offensive” if they were posted by people who identify as African-American.

AI systems meant to spot abusive online content are far more likely to label tweets “offensive” if they were posted by people who identify as African-American.

The news: Researchers built two AI systems and tested them on a pair of data sets of more than 100,000 tweets that had been annotated by humans with labels like “offensive,” “none,” or “hate speech.” One of the algorithms incorrectly flagged 46% of inoffensive tweets by African-American authors as offensive. Tests on bigger data sets, including one composed of 5.4 million tweets, found that posts by African-American authors were 1.5 times more likely to be labeled as offensive. When the researchers then tested Google’s Perspective, an AI tool that the company lets anyone use to moderate online discussions, they found similar racial biases.

A hard balance to strike: Mass shootings perpetrated by white supremacists in the US and New Zealand have led to growing calls from politicians for social-media platforms to do more to weed out hate speech. These studies underline just how complicated a task that is. Whether language is offensive can depend on who’s saying it, and who’s hearing it. For example, a black person using the “N word” is very different from a white person using it. But AI systems do not, and currently cannot, understand that nuance.

We tend to think of AI as a monolithic entity, but it has actually developed along multiple branches. One of the main branches involves performing traditional calculations but feeding the results into another layer that takes input from multiple calculations and weighs them before performing its calculations and forwarding those on. Another branch involves mimicking the behavior of traditional neurons: many small units communicating in bursts of activity called spikes, and keeping track of the history of past activity.

Each of these, in turn, has different branches based on the structure of its layers and communications networks, types of calculations performed, and so on. Rather than being able to act in a manner we would recognize as intelligent, many of these are very good at specialized problems, like pattern recognition or playing poker. And processors that are meant to accelerate the performance of the software can typically only improve a subset of them.

That last division may have come to an end with the development of Tianjic by a large team of researchers primarily based in China. Tianjic is engineered so that its individual processing units can switch from spiking communications back to binary and perform a large range of calculations, in almost all cases faster and more efficiently than a GPU can. To demonstrate the chip’s abilities, the researchers threw together a self-driving bicycle that ran three different AI algorithms on a single chip simultaneously.

The research and development of neural networks is flourishing thanks to recent advancements in computational power, the discovery of new algorithms, and an increase in labelled data. Before the current explosion of activity in the space, the practical applications of neural networks were limited.

Much of the recent research has allowed for broad application, the heavy computational requirements for machine learning models still restrain it from truly entering the mainstream. Now, emerging algorithms are on the cusp of pushing neural networks into more conventional applications through exponentially increased efficiency.