Five Keys to Trusted AI

By Jitendra Gupta, Director, Ops Decision Science, Wolters Kluwer

Artificial intelligence (AI) has quickly evolved from an emerging technology to on its way to becoming an established one. Since 2017, adoption has more than doubled. Globally, the market for AI is slated to top $1 trillion by 2028. But as more companies and industries reap the benefits of AI, they also need to ensure the technology they’re using is trustworthy.

While Deloitte, NIST, and Gartner have all released guidance regarding AI trustworthiness, I’ve compiled my own checklist based on years of successfully training and refining my company’s model, which has a proven track record of success. Here are five main components of trustworthy AI.

  1. Data

AI is only as trustworthy as the data it’s trained on. The greater the volume and variety of data, the more trustworthy the AI will be. By ingesting clean, refined, and (ideally) structured data that reflects a wide swath of scenarios, a model will be able to offer more accurate predictions and recommendations.

Volume and variety are also important to reduce bias. On the one hand, AI often promises to reduce bias by minimizing the role of fallible, subconsciously prejudiced humans. On the other, bias can be inherent in certain sets of data, increasing the risk that the AI will produce biased results. Training AI with a high volume of varied data mitigates this risk, as does the use of a feedback loop to monitor for bias on an ongoing basis.

  1. Models

Trusted AI is transparent AI. “Black-box algorithms”—AI models in which there is no visibility into how the decision is made—are at the heart of distrust in AI. Therefore, AI must be transparent. Interpretive models that clearly explain the algorithm’s output are important to building trust in AI.

Users must be able to understand how the AI came to its conclusions. Then, they can assess those conclusions and feed input back into the system, helping make the AI even more accurate, intelligent, and trusted.

  1. Partners

Carefully vetting the partners you use for AI solutions means carefully vetting the models they build. In addition to ensuring your partner has a large, diverse data set on which to train AI, also analyze whether they have a solid business model to support continual development. There are countless startups touting game-changing AI, but the quality of their models is irrelevant if they go out of business a few months (or even a few years) down the road.

I suggest choosing partners with the most usage, as usage translates to trustworthiness in most instances. Ask potential partners how many customers are using their models, how long they have been building their models or how long those models have been in operation, and whether they have the skills and assets to maintain high-quality models for years to come.

  1. Quality

AI models must be updated and calibrated continually to ensure trustworthiness. Feedback loops from internal users are the bread-and-butter of continual quality control, as they help ensure that models are up-to-date and accurate.

I also recommend having a third-party audit any AI you use. At our company, we have an independent team audit our models regularly, assessing everything from data collection to model building. While quality controls are built into our models, it’s reassuring to have an external team apply its own framework for assessing reliability and accuracy.

Having the stamp of approval from an independent source also helps improve the confidence of customers and partners looking to use our AI. It has helped us increase usage, refine our model, and improve trustworthiness.

  1. People

Perhaps the biggest misconception about AI is that it’s going to eliminate the need for humans in workflows altogether. This is a myth. Keeping a human in the loop is crucial to maintaining trustworthiness.

While a machine can sort through a large amount of data much more quickly than a human, any decision recommended by AI must still be vetted and approved by someone with the proper expertise. Indeed, it is their expertise and input that makes AI work more efficiently and effectively. Essentially, people keep the system honest while using their expertise and experience to improve the overall quality of AI generated results.

Trust in AI cannot be built overnight, but it can be cultivated and, ultimately, achieved. It must be established and fostered on an ongoing basis, through diverse data sets, sufficient feedback loops, great partners, careful implementation, and knowledgeable human beings. Put all of those together, and you’ll have a trustworthy AI system that works for you.