How to ensure fair AI throughout the supply chain

For some time now, there has been talk about how leaders developing AI applications need to build “fair AI”; this should be unbiased and equitable, ideally improving the quality of life of everyone it touches.

However, most of the thinking around ethical AI has been focused around models, explainability, technical teams and data governance.

But how about considering basic fairness in the AI development lifecycle? This lifecycle is immense. It entails the data collection and annotation contractors, the companies and individuals managing the data, the technology specialists building the AI models, the go-to-market experts building the AI applications, and the businesses and individuals using AI-powered products and services.

Such an approach is the only way we can ensure that technology continues to make the world a better place.

When an AI product is deployed in the real world, it must work as expected, deliver equitable results for all intended beneficiaries under all circumstances, and not harm anyone physically, mentally or emotionally. That’s a tall order. And it starts with building an unbiased and comprehensive data set.

While this may seem obvious, it’s easy for development teams to put unbiased data to the bottom of the pile and instead focus on achieving results as quickly as possible. However, doing so creates a long-term responsibility that developers can easily forget to maintain.

Unfairness in the form of bias can appear in AI in multiple ways.

Consider criminal risk assessment algorithms that use behavioural and demographic data to determine the risk of reoffending.

One recent study found one such algorithm to be racially biased. For example, an 18-year-old black woman was charged with petty theft, having stolen an $80 bicycle. Despite having only one prior juvenile misdemeanour, she was rated as having a higher risk of reoffending than a 41-year-old white male who was charged with a similar crime but had several prior offenses.

Moving forward in time, the woman committed no other offenses while the male is now serving a prison term for the theft of thousands of dollars of electronics. Clearly, the algorithm was based on poor data.

Unfair AI also lives inside everyday technology. If a company sells cars equipped with speech recognition in multiple countries but trains the product using only native male speakers for each language, the system may struggle to understand women or anyone with a different accent.

This could lead to drivers being taken to the wrong destination. Worse still, it could cause distracted driving, leading to accidents. In addition to being unfair to some users, biased data can saddle solution providers with substandard products that can damage their reputation.

Developing a comprehensive and unbiased dataset requires data diversity and breadth. This ensures the product is trained in every situation it is likely to encounter in real life, such as all of the accents, voice tones and languages that a car’s speech recognition system may encounter in its target markets.

Achieving this means working with people who resemble the entire customer profile to collect, annotate and validate the AI model training data.

It also means working with a diverse team on the model building itself. A computer-based annotation system can’t do it alone when it comes to interpreting complex situations and catching subtle errors that could have real-life consequences.

For example, a human annotating images or a video for a self-driving car application could interpret that a person with a certain posture walking between two cars may be pushing a buggy that will appear in traffic before the person does.

Even the best computer-based annotation systems would struggle to make this interpretation. Similarly, a human reading a product review is much more likely to detect sarcasm than a machine is.

The people behind the data

Leaders committed to fair AI must include another important link in the AI development lifecycle when building global AI products or services: the millions of people who collect and label the data. Engaging these people in a fair and ethical way is mission-critical and should be part of every organisation’s responsibility charter.

Fair treatment means committing to fair pay, flexible working hours, including people from any and all backgrounds, respecting privacy and confidentiality, and working with people in a way that they feel heard and respected.

Leaders should also inspire their contractors in a way that instils pride in working on the most impactful technology used by the global economy.

Why does fair AI matter, beyond the obvious?

Quite simply, it’s good for society, and it’s good for business. Product teams, for example, are inspired when they’re building products that have a positive impact on their market and the world. But what else do fair products do?