Magritte's the Treachery of Images

A few days ago, several of my friends shared a meme on Facebook that showed how artificial intelligence / machine learning (AI/ML) had tagged a perfectly innocent picture with a caption that was, shall we say, not suitable for polite company. The caption was technically accurate, but the algorithm that produced it was completely unaware it could be read another way. ML is not yet able to cope with unintended double-entendres.

While it caused a few seconds amusement, the example did get me thinking. Is machine learning ready for prime time? On the surface it would seem so. At a recent online industry conference I attended nearly every speaker mentioned AI in some form or other. Conversations with customers are peppered with discussions around how they can use machine learning to meet the exponential growth in demand for applying metadata to their content.

Yet along with those conversations I have also heard an equal amount about trust and validation of the results of implementing an AI/ML solution. I think these sorts of concerns can be addressed by taking a systematic approach to thinking about why you want an AI/ML solution in the first place, and how you will actually use it. The following six questions can help you develop that approach:

1. What Do You Need an AI/ML Solution For?

Consider why you are considering implementing an AI/ML solution in the first place. I hope the answer isn’t just because it sounds cool or the CIO saw a fancy demo. You should have a real need it will address. An AI initiative is not something to be undertaken just because everyone else is doing it. Without a clear goal it will become an expensive failure that will more than likely set up a road-block when a real need arises.

Related Article: Use AI Thinking to Improve Your Customer Experience

2. What Business Problem Are You Trying to Solve?

In general you could say that the underlying drive for AI/ML is the fact that it can do what humans can do but at scale and speed with a greater consistency. It’s ideal for reducing or eliminating or reducing repetitive tasks. It’s also great for scaling up. But all that needs to be done to solve an actual problem where you have clear objectives and measurable results that positively impact the business.

3. What Type of Models Are You Using?

There are several AI/ML services available that provide data and insights based on public data sets. These, by their very nature, tend to deliver generic results. In many cases these are good enough — at least to get you started. But to really derive the most benefit you need an AI/ML solution that you can build models for that are specific to your business needs. Is it good enough for you to know that something in an image is a bicycle, or do you need to know that it’s a specific brand of mountain bike?

Related Article: What Do Machine Learning Engineers Do Anyway?

4. What Are You Teaching It?

When setting up your AI/ML models consider the language you are teaching it. The ML needs to deliver results that can be used across the business. You need to ensure that all the business systems that are using the output are speaking the same language. If you do not already have one, you need a well defined taxonomy to ensure that the same item is referenced in the same way across the organization. This will also help with later search and findability. Is that bicycle a mountain bike, an electric bike, a hybrid bicycle, a touring bicycle, or a tandem?

5. What Are You Doing With the Results?

It’s no good getting good results from your AI/ML if you don’t do anything with them. From a content perspective, these results should be considered as metadata to be applied to the source material to make it more findable and useful, especially for things like personalization and omnichannel publishing. In fact the mass creation and application of metadata for growing content is often a driving need for AI/ML.

Related Article: The Pandemic Is Changing How Consumers Feel About AI in CX

6. What Validation Processes Do You Have in Place?

When is a bicycle not a bicycle? How about when it's a painting of a bicycle icon in a photo of a bicycle lane. Just how will your AI/ML tag that? 

This is where we need to make sure that AI/ML is a tool used alongside your subject matter and context experts. AI/ML will tag or label what it sees based on the algorithms and models. What it will not understand is context. You need to ensure that you have a human-in-the-loop validation program in place to avoid the sort of problems I mentioned at the start of this article.

If you work your way through these discussion points and realize that artificial intelligence and machine learning is a tool to work alongside you existing human expertise, you can go a long way to addressing any trust and validation concerns.