Ready to learn Artificial Intelligence? Browse courses like Uncertain Knowledge and Reasoning in Artificial Intelligence developed by industry thought leaders and Experfy in Harvard Innovation Lab.
From facial recognition for unlocking our smartphones to speech recognition and intent analysis for voice assistance, artificial intelligence is all around us today. In the business world, AI is helping us uncover new insight from data and enhance decision-making.
For example, online retailers use AI to recommend new products to consumers based on past purchases. And, banks use conversational AI to interact with clients and enhance their customer experiences.
However, most of the AI in use now is “narrow AI,” meaning it is only capable of performing individual tasks. In contrast, general AI – which is not available yet – can replicate human thought and function, taking emotions and judgment into account.
General AI is still a way off so only time will tell how it will perform. In the meantime, narrow AI does a good job at executing tasks, but it comes with limitations, including the possibility of introducing biases.
AI bias may come from incomplete datasets or incorrect values. Bias may also emerge through interactions overtime, skewing the machine’s learning. Moreover, a sudden business change, such as a new law or business rule, or ineffective training algorithms can also cause bias. We need to understand how to recognize these biases, and design, implement and govern our AI applications in order to make sure the technology generates its desired business outcomes.
Recognize and evaluate bias – in data samples and training
One of the main drivers of bias is the lack of diversity in the data samples used to train an AI system. Sometimes the data is not readily available or it may not even exist, making it hard to address all potential use cases.
For instance, airlines routinely apply sensor data from in-flight aircraft engines through AI algorithms to predict needed maintenance and improve overall performance. But if the machine is trained with only data from flights over the Northern Hemisphere and then applied to a flight across sub-Saharan Africa, the conditions will provide inaccurate results. We need to evaluate the data used to train these systems and strive for well-rounded data samples.
Another driver of bias is incomplete training algorithms. For example, a chatbot designed to learn from conversations may be exposed to politically incorrect language. Unless trained not to, the chatbot may start using the same language with consumers, which Microsoft unfortunately learned in 2016 with its now-defunct Twitter bot, “Tay.” If a system is incomplete or skewed through learning like Tay, then teams have to adjust the use case and pivot as needed.
Rushed training can also lead to bias. We often get excited about introducing AI into our businesses so naturally want to start developing projects and see some quick wins.
However, early applications can quickly expand beyond their intended purpose. Given that current AI cannot cover the gamut of human thought and judgement, eliminating emerging biases becomes a necessary task. Therefore, people will continue to be important in AI applications. Only people have the domain knowledge – acquired industry, business, and customer knowledge – needed to evaluate the data for biases and train the models accordingly.
Diversify datasets and the teams working with AI
Diversity is the key to mitigating AI biases – diversity in the datasets and the workforce working day to day with the models. As stated above, we need to have comprehensive, well-rounded datasets that can broadly cover all possible use cases. If there is underrepresented or disproportionate internal data, such as if the AI only has homogenous datasets, then external sources may fill in the gaps in information. This gives the machine a richer pool of data to learn and work with – and leads to predictions that are far more accurate.
Likewise, diversity in the teams working with AI can help mitigate bias. When there is only a small group within one department working on an application, it is easy for the thinking of these individuals to influence the system’s design and algorithms. Starting with a diverse team or introducing others into an existing group can make for a much more holistic solution. A team with varying skills, thinking, approaches and backgrounds is better equipped to recognize existing AI bias and anticipate potential bias.
For example, one bank used AI to automate 80 percent of its financial spreading process for public and private companies. It involved extracting numbers out of documents and formatting them into templates, while logging each step along the way. To train the AI and make sure the system pulled the right data while avoiding bias, the bank relied on a diverse team of experts with data science, customer experience, and credit decisioning expertise. Today, it applies AI to spreading on 45,000 customer accounts across 35 countries.
Consider emerging biases and preemptively train the machine
While AI can introduce biases, proper design (including the data samples and models) and thoughtful usage (such as governance over the AI’s learning) can help reduce and prevent them. And, in many situations, AI can actually minimize bias that would otherwise be present in human decision-making. An objective algorithm can compensate for the natural bias that a human might introduce such in approving a customer for a loan based on their appearance.
In recruiting, an AI program can review job descriptions to eliminate unconscious gender biases by flagging and removing words that may be construed as more masculine or feminine, and replacing them with more neutral terms. It is important to note that a domain expert needs to go in and make sure the changes are still accurate, but the system can recognize things that people could miss.
Bias is an unfortunate reality in today’s AI applications. But by evaluating the data samples and training algorithms and making sure that both are comprehensive and complete, we can mitigate unintended biases. We need to task diverse teams with governing the machines to prevent unwanted outcomes. With the right protocol and measures, we can ensure that AI delivers on its promise and yields the best business results.
Originally appeared in Information Management