Microsoft’s Tay and the Catch-22 of Rigid Boundaries

Microsoft recently released an ambitious “chatbot” on Twitter called Tay. Tay was designed as a way to “conduct research on conversational understanding.” Of course, there’s no better way to test conversation at scale than on Twitter.

People quickly began repeatedly baiting Tay into negative conversations, eventually leading to the chatbot generating a series of offensive tweets. Microsoft shut Tay down within just a few days of the launch.

Many have since argued that Microsoft should have built stronger safeguards against this behavior; meanwhile, the market continues to demand more dynamic, adaptive A.I. that can respond to any question or comment you throw at it.

In short, the technology is stuck in a catch-22. We want to assure that conversational A.I. – chatbots, intelligent virtual assistants (IVAs) and the like – converse in as human a way as possible, but the margin for error is razor-thin. In the next few years, we hope to see more research that helps us improve conversational A.I. systems; nevertheless, it’s important that businesses understand how we deliver conversational A.I. that works today.

Certainly safeguards are important, but what does that mean?

At Next IT, we believe there should be human oversight of the responses an IVA gives to a user. In Tay’s case, Microsoft had some degree of human oversight in place, but when you test on a platform as large and fast moving as Twitter, things can quickly get out of control. That’s why, in addition to human oversight, Next IT sets rigid boundaries and process flows that pass users off to human agents whenever an IVA lacks confidence in its ability to offer an accurate response.

Safeguards help eliminate the chance of something going wrong, but that’s not enough: We also want to increase the odds that our IVAs will deliver a great experience. This is why every IVA we deliver is designed to execute goal-based conversations.

Goal-based conversation guides dialogue toward predictable endpoints, which could include helping a user reset a password or supplying a link to an internal document. If the IVA isn’t sure what its goal is, it will ask additional questions to clarify user intent. Tay wasn’t designed to conduct goal-based dialogue, which is why users were able to manipulate her responses. If she were designed with this capability, the scope of possible conversational topics would be narrowed sufficiently to allow for effective human oversight.

So why should goal-based conversation matter to you?

Because goals are valuable to your business. Whether you aim to help travelers book a flight or reduce the cost of your service and support operations, when you set goals, you define an appropriate path to their realization. If you create goal-based dialogue with your conversational A.I., you’ll protect against the unexpected while delivering a world-class user experience.

One day we might develop an A.I. without firm boundaries that can freely converse with users, but today’s most successful IVAs have boundaries designed to realize your goals and drive business value.