Ask Dr. Paola: Detecting & Battling Biases in Artificial Intelligence & Machine Learning (Part 2)

Topics: Artificial Intelligence, Data Analytics, Efficiency, Harvard Law School, Law Firms, Legal Innovation, Midsize Law Firms Blog Posts, Women’s Leadership Blog Posts

Artificial Intelligence

Each month, Dr. Paola Cecchi-Dimeglio, a behavioral scientist and senior research fellow for Harvard Law School’s Center on the Legal Profession and the Harvard Kennedy School, answers questions about how law firms and legal service firms can navigate a dramatically changing legal environment by using data analytics and behavioral science to create incentives for their lawyers and others to change their behavior. (You can follow Paola on Twitter at @HLSPaola.)

Ask Dr. Paola: You spoke last time about the inherent biases that can be found in data, and how artificial intelligence that works with that data can spread those biases forward. How can organizations detect those biases and combat them?

Dr. Paola Cecchi-Dimeglio: You have to remember that with many legal organizations, the data they are looking at is either what is publicly available or data they have gathered from working with their clients. And when artificial intelligence (AI) starts working with this data, it can be a very positive thing for a law firm, for example, allowing it to make better decisions about jurisdiction and judges, and client matters in comparable situations. The firm can take all the past experience and distill it into likely scenarios in order to help the firm leadership make better choices — that is a positive aspect of AI here.

But problems arise, especially problems with biases, when the organization isn’t careful about where it’s taking its data from or about what portion of data it’s using and not using. Because if you start out with a biased history, you’re going to have biased results.

Ask Dr. Paola

Dr. Paola Cecchi-Dimeglio

Worse yet, you’re continuing to perpetuate the biases because the algorithms you’re using to search through the data have lost the context and thus take at face value what the data says. So, you don’t see the true pattern of the data for the population at large. And we know that this biased data can create feedback loops where it just sends down what it’s expected to find, rather than objectively looking at what is out there.

It’s important to understand, and I’ve said this before, that AI — through utilizing both machine learning and deep learning — has the ability to mimic, to some extent, human decision-making. However, if you don’t pay attention to input and outcomes, then the human decision-making process can evolve into one with inherent biases. Many people think (and I strongly agree) that AI’s machine learning and deep learning give us a rare opportunity to reverse this process and in fact, it can become a way for us to be less biased as humans and not perpetuate our current biases.

Organizations need to understand this dilemma as almost a “cycle of bias”, asking themselves how they can keep from passing that bias down to the machines. The first step in that, of course, is to examine and understand how biased your current data is. Analyze the data to assess if there are group members who are disadvantaged by the very nature of the dataset you’re using. You can also look at the output and see if the output is biased or not.

The second point is to look at who built those algorithms and what factors are taken into account when the algorithms were built.

Finally, I think the organization’s leadership and its professionals have to make a commitment toward fixing this problem. They should clearly state, “We accept that fixing the problem means not only a commitment now, but it may need more fine-tuning over time. This is not the end of it.” Because there may be biases that you may not realize are evident in your software or your algorithms right away, but the more AI will interact and be fed with it, the more the biases will appear.

So, an organization must ask itself several questions about its data and about how its algorithms access that data in order to truly begin the process of ridding it of biases. That’s why there is a big push overall in AI, machine learning, and deep learning for more transparency and accountability, and for more ethical rules.

And to get there, I think there are a few things that people should address when pushing for more transparency. Any organization that is looking at lessening the potential biases inherent in its data needs to ask itself “Who has access to the people (the subjects of the data), who has access to the algorithms, and who has designed and built them?”

You need to establish a better architecture of the team and understanding where the biases exist before you can ask the question, “Are we looking at what is accurate and inaccurate? What would a more rigorous testing of the system and its input/outcomes show us about our biases?”

Identifying biases in algorithms is a task that requires commitment and the ability to be resilient but if done well, over time you can greatly reduce and remove these biases.

In the end your algorithms need to provide information that is unbiased in order to help human beings make unbiased decisions.

Next time, I will examine ways in which you can reduce biases in recruitment and analyze how firms can ensure a more level playing field for their candidates and prospects.