How to Build Trust In AI

It's hard to make it through a tech-conference these days without hearing a pitch for an idea that sounds a lot like this: 

"Do you have an important business problem that no one knows how to solve? Don’t worry we will build an AI that will solve the problem for you".

With so much hype around AI it’s hard to tell the difference between a promise someone can fulfill and the more hopeful, aspirational goals meant simply to excite. However, if we really want AI to have a positive impact on the world we need more than excitement, we need AI we can trust.

Trust is a complicated thing; you might think we don’t trust AI because we’re afraid it will make a mistake. But I have many friends and colleagues that I trust even if they make mistakes. In this article I will explore why AI is different and share some of my experiences building AI that people trust. This article is based in part on my keynote presentation at the 2017 Quest for Quality conference in Dublin, and several discussions I had there around the topic. 

Beware of what you call AI

Today, when someone says they are using AI they are most likely talking about a very narrow subset of AI: machine learning. 

Machine learning (ML) algorithms use thousands or millions of carefully labeled examples to learn patterns. These patterns can then be used to make predictions about new data. This way of learning does not conform very well to what you or I would consider intelligence. Yet, somehow we’re ok with calling ML algorithms intelligent. This is because intuitively we expect intelligence to be required for the problems ML is able to solve. 

The danger here is that the term “intelligence” creates high expectations, even among experts. This matters when building trust because people trust things that behave as they expect them to. And high expectations and excitement in the short term can erode trust in the long term.

Keep a human In the loop

If machine learning is not intelligence, should we still be excited? Yes, and the reason is that machines can process data faster than humans. That means machine learning can help us automate tedious or laborious tasks. As machine learning methods become more advanced, the range of tasks we can automate expands to include video and audio processing, marketing, transportation, and much more.

However, there are very few examples where this automation has delivered an end-to-end service. In most cases a human still needs to make a final decision. When it comes to building trust, this is actually not a bad thing. When it comes to trust, users prefer humans over computer systems. Humans are still far better than machines at adapting to unforeseen situations.

Some current AI products anthropomorphize their AI to bridge that gap. But, unless you’re making a chat-bot, you probably don’t need to name your AI. Your deep learning model probably looks more like a black box than a sassy personal assistant. (Not to mention the fact that all the sassy chat-bots out there get their sass from hard-coded rules made by developers, not from AI). Computer scientists already observed in 1966 that if you give your system human characteristics, your users will likely overestimate its capability. Instead, why not just be proud of your actual humans?

Plan for failure

There have been a many recent headlines trumpeting how machine learning algorithms outperformed humans in visual recognition, speech recognition, and games like Go. And while these headlines have captured the imagination of the general public, the reality is that these examples are more the exception than the rule. Many business/user problems are more complex or do not have large labeled data sets to train machine learning algorithms, resulting in low accuracy. 

This, along with the fact that machines are not very good at adapting to unforeseen inputs, means that when you deploy AI in the wild, it's just a matter of time before the system fails. More important than avoiding failure is what you do when failure happens. 

One way to address this is to think of your AI not as a decision maker, but as a recommender system. For example, consider product recommendations in online stores. An AI algorithm doesn't have to predict everything a person might want to buy, it just has to find a few high-confidence examples.

Other approaches include fallback models, anomaly detection, or building directly into the AI how they should handle unexpected inputs.

Build a kick-ass product

It's good to have inspiring goals, but it's also important to deliver a good product. All the marketing in the world won’t turn a bad product into a good one. Excitement sells, but it’s not a viable long-term strategy on its own.

One recent example in the web-design world promised “websites that design themselves”. It did not take long for the cracks to start showing. After a year of development, it was described by users as "shoddy and expensive”. A clear mismatch between expectation and reality. Too bad for a technology that has the potential to save hours of time for a designer (without putting them out of a job).

Conclusion

All this adds up to a simple message: Talk less about AI, and just delight your users with a terrific product. If your product is powered by AI, great. If not, also great. There are as many customers that will get excited about AI as will be turned-off by the idea, and all of them will be disappointed if you fail to deliver on your promises. It is too hard to regain trust once lost.