Here are 3 Tips to Reduce Bias in AI-Powered Chatbots


AI-powered chatbots that use natural language processing are on the rise across all industries. A practical application is providing dynamic customer support that allows users to ask questions and receive highly relevant responses. In health care, for example, one customer may ask “What’s my copay for an annual check-up?” and another may ask “How much does seeing the doctor cost?” A smartly trained chatbot will understand that both questions have the same intent and provide a contextually relevant answer based on available data.

What many people don’t realize is that AI-powered chatbots are like children: They learn by example. Just like a child’s brain in early development, AI systems are designed to process huge amounts of data in order to form predictions about the world and act accordingly. AI solutions are trained by humans and synthesize patterns from experience. However, there are many patterns inherent in human societies that we don’t want to reinforce — for example, social biases. How do we design machine learning systems that are not only intelligent but also egalitarian?

Social bias is an increasingly important conversation in the AI community, and we still have a lot of work to do. Researchers from the University of Massachusetts recently found that the accuracy of several common NLP tools was dramatically lower for speakers of “non-standard” varieties of English, such as African American Vernacular English (AAVE). Another research group, from MIT and Stanford, reported that three commercial face-recognition programs demonstrated both skin-type and gender biases, with significantly higher error rates for females and for individuals with darker skin. In both of these cases, we see the negative impact of training a system on a non-representational data set. AI can learn only as much as the examples it is exposed to — if the data is biased, the machine will be as well.

Bots and other AI solutions now assist humans with thousands of tasks across every industry, and bias can limit a consumer’s access to critical information and resources. In the field of health care, eradicating bias is critical. We must ensure that all people, including those in minority and underrepresented populations, can take advantage of tools that we’ve created to save them money, keep them healthy, and help them find care when they need it most.

So, what’s the solution? Based on our experience of training with IBM Watson for more than four years, you can minimize bias in AI applications by considering the following suggestions:

  • Be thoughtful about your data strategy;
  • Encourage a representational set of users; and
  • Create a diverse development team.
1. Be thoughtful about your data strategy

When it comes to training, AI architects have choices to make. The decisions are not only technical, but ethical. If our training examples aren’t representative of our users, we’re going to have low system accuracy when our application makes it to the real world.

It may sound simple to create a training set that includes a diverse set of examples, but it’s easy to overlook if you aren’t careful. You may need to go out of your way to find or create datasets with examples from a variety of demographics. At some point, we will also want to train our bot on data examples from real usage, rather than relying on scraped or manufactured datasets. But what do we do if even our real users don’t represent all the populations we’d like to include?

We can take a laissez-faire approach, allowing natural trends to guide development without editing the data at all. The benefit of this approach is that you can optimize performance to your general population of users. However, that may come at the expense of an underrepresented population that we don’t want to ignore. For example, if the majority of users interacting with a chatbot are under the age of 65, the bot will see very few questions about medical services that apply only to an over-65 population, such as osteoporosis screenings and fall prevention counseling. If bots are only trained on real interactions, with no additional guidance, it may not perform as well on questions about those services, which disadvantages older adults who need that information.

In order to combat this at my company, we create synthetic training questions or seek another data source for questions about osteoporosis screenings and fall prevention counseling. By strategically enforcing more distribution and representativeness in our training data, we allow our bot to learn a wider range of topics, without unfair preference for the interests of the majority user demographic.

Read the source article in VentureBeat.