Usable AI: User Experience Designers have a simple answer to the AI “True Understanding” problem

We are in the Age of Wonder when it comes to technology that “understands” everyday spoken and written human language. To move on from this we only need to go back to well-established software design practices so we’re clear on what this understanding is useful for.

It started late 2017 and is in full-force now with marvels like Talk To Tranformer and HuggingFace Transformer and Allen AI’s Grover, all using language technology called “general purpose language models” that have given everyone a boost in better quality “language understanding”.

The NLP community at the moment has a massive amount of energy invested in ways to “understand language”. There are over 200 NLP “tasks” identified such as summarisation, translation, question answering and a bunch of technical use cases such as “co-reference resolution” and “named entity recognition”.

Again and again I see discussion of what these systems “understand”.

“Understanding” is a human term. Technology will never “understand”. We have anthropomorphised technology, eagerly projecting human-like labels at the merest hint of promising behaviour. We can’t help it. I caught myself saying that my Eufy robot vacuum cleaner “wanted to escape again” today because of its tendency to seem to wander out of the room I’m trying to get it to clean.

We love humanising technology. Our Eufy robovac “behaves like a young toddler” and our household empathises when it seems to get lost, or manages to climb on top of my son’s train set and gets stuck, gets confused by party balloons, or somehow, innocently of course, manages to get tangled up in some cables and pulls a laptop and an ipad off a bench onto the stone floor! Aww isn’t cute.

Enter ISO-9241–11

The fact is, not even humans have “true understanding”: understanding all a matter of degree (see my fun example).

So what do we do? Let’s say you see a competitive opportunity for using NLP and after a while it becomes really hard to understand which technologies and methods will give you your understanding.

Usefully, this is not a new problem. There is even a very sound ISO standard for software design: ISO-9241 part 11, which says:

Usability is the extent to which a product can be used by users to perform tasks in order to achieve goals, with effectiveness, efficiency and satisfaction in a specified context of use.

So Usable AI then, has the key. You design a system for specific tasks, and within a context of use and goal, you measure:

• How efficient the software makes people using it
• How effective they are at the task (how often do they actually complete the tasks)
• How satisfying: software that is a joy to use can be a significant competitive advantage.

Next time someone wants “to build some conversational AI”, or “build an NLP interface to our system”, it becomes clear that the shape of the vocabulary you use is usefully defined by the context, and the accompanying goal, the tasks that achieve that goal, and the nature and sophistication of the dialog necessary for that domain.

Assuming that, how do you ensure that this probabilistic technology (ie only delivering features a certain percentage of the time) meets usability goals?

Like this post if you’re curious and if I have more than 20 people like it I’ll do the next one sooner rather than later.