It has happened again. Facial recognition technology is in hot waters, this time for flagging off Black men as primates. The New York Times reported an incident where Facebook users watched a video of Black men on a British tabloid were prompted with “Would you like to keep seeing videos of primates?”. This led to the company to immediately disable the AI feature that provided this recommendation, apologize for this ‘unacceptable error’ and explore why this happened.

A question we’re all asking too. The commonly expected response is the lack of a varied dataset. Conventionally, datasets especially for visual algorithms pertaining to applications like Computer Vision and Facial Recognition lack sufficient diverse representation. There is a predominance of Caucasian faces and skin tones, which often makes the machine ‘discriminate’ against other races such as African Americans and Asians. The first study to substantiate algorithmic bias was by the National Institute of Standards and Technology (NIST) – in 2003, it was found that machines found it harder to recognize female subjects than male ones. In 2018, researchers from MIT and Microsoft released a report that revealed gender classification algorithms (different from facial recognition algos) had error rates of just 1% for white men, but almost 35% for dark-skinned women. NIST found that Asians, African Americans, and American Indians generally had higher false positive error rates than white individuals, children and the elderly had higher false positive rates than middle aged adults.

But why would a machine classify a human being as an animal? One possible explanation is that history has long alluded to referring to African Americans are primates and apes. Machines have learnt this prejudice, embedding it further into their digital DNA. In 2015, Google categorized a freelance web developer called Jacky Alcine as a gorilla. Google Photos sent a notification to Alcine’s phone following image categorization of him and his friends. The ones where Alcine is seen with his Black friends, the phone categorized those pics as ‘Gorilla’. Google supposedly resolved this issue within 14 hours of Alcine tweeting at the company. Image sharing site Flickr too has slipped up with image categorization, calling people ‘animals’ and concentration camps as ‘jungle gyms’.

Noted AI scientist Yoshua Bengio says these kinds of incidents will keep happening. “We would have to have people tell the machine why it makes specific mistakes. But you would need not just like two or three examples, you would need thousands or millions of examples for the machine to catch all of the different types of errors that we think exists. The system can make mistakes and you have to deal with the fact that there will be mistakes.” This means, a machine needs to be taught how complex a human-centric society really is.

After the Facebook-primate incident last week, the company spokesperson Dani Lever said: “As we have said, while we have made improvements to our A.I., we know it’s not perfect, and we have more progress to make. We apologize to anyone who may have seen these offensive recommendations.”

So clearly, this problem is here to stay. But the solution to this problem lies with one fundamental building block of AI – data. A study by Dimensional Research in 2019 stated that 96% of companies surveyed ran into training-related problems with data quality, labelling required to train the AI and building overall model confidence. Here are some more compelling findings from this report:

• 81% admit that training AI with data is harder than expected
• 96% encounter challenges with training data and quality
• Almost 78% of AI/ML projects stall at some stage before deployment.
• Companies often resort to a mix of data labelling approaches including labelled/annotated data, pre-labelled data and contracted labelling
• AI teams have less than 10 members 

These are some troubling facts and compel us to ponder over the nature of data cleaning, parsing and annotation. Purchasing datasets is also an expensive proposition, and conventionally, the tech giants like Amazon, Facebook and Google have a monopoly over this as well. While smaller companies and upcoming AI startups notably struggle in their early days to get relevant and clean data, bigger companies do have resources and budgets available for these largescale endeavours. Moreover, GPUs have made impressive advancements in recent times to develop complex Deep Convolutional Neural Networks (DNNs) – which is where facial recognition software are run.

While it is possible to concur that regulations always play catch-up to tech, cultural and social norms have to reflect in the very technology we use to make our lives better. By sidelining evolved perspectives on modern society, technology isn’t a true extension of the society and times we live in but a means to reinforce jaded tropes, misinformation and centuries-old biases. With technology, we can enable the shift in narrative that is being seen worldover today, not send us back in time.