When you or I look at an image of a cat, we may see furry ears, a tail and whiskers, whereas a machine sees clusters of pixels. In order for an AI to explain in terms that we understand why it classified an image as a cat, the machine must first know our mental model of a cat (having furry ears, tail and whiskers), and apply these labels against its own perception of what makes up a cat (a certain arrangement of clusters of pixels).
If the machine were to simply provide us with the causal chain of calculations (i.e. the machine’s language) that led to the output “cat”, it would not be a useful explanation. There are millions of variables, each with their own weight, that go into any given AI output. An additional layer of analysis – a translation – is required to put the explanation into terms that are understandable to humans.
Instead of binary yes-or-no outputs, or complex chains of calculations, service design can enable a more nuanced exchange between the machine and the worker, thereby enabling the worker and the AI to work together more harmoniously. Highlighting the most important variables considered by the algorithm, visualising the key decision-making points and patterns, displaying the level of confidence (i.e. accuracy, precision and recall scores3) in the proposed output, or even ‘chunking’ the output into more manageable deliverables, may be features that future interfaces consider for a smoother handover between human workers and AI tools.
Likewise, humans will require ways to feed information into AI models to teach those models to become better over time (such as training an insurance fraud detector to improve its alerts and predictions over time) and in order to co-produce outputs. For example, Waze, the route-mapping app, finds the fastest route using both its algorithms as well as other users’ real-time inputs about accidents and other traffic disruptions.
Explanations are hard enough between humans, let alone in human-to-machine cases. Our language is often context-specific, fuzzy and imprecise. For example, the concept of a “small town” would vary widely between someone who grew up in a sparsely populated country such as Canada, compared to someone who grew up in a country like China, where a town can have a population of three million. When speaking with other humans, we ask for clarification and we make inferences based on what we already know about the conversation and context, as well as about the person we’re speaking with. Despite the fuzziness of the language, humans manage to have meaningful conversations with one another.
Service designers will need to think-through and design interfaces that are suited to translating meaning between humans and AI. These interfaces will need to create the conditions for humans to provide consistent, clear and accurate inputs that set the machine up for success. This will mean taking into consideration things such as the current cognitive load of the human worker while providing inputs, and the different ways humans express meaning (do workers have varying language or mental models to describe the same phenomena?). Getting these elements right will require early and frequent usability testing throughout the development process of the AI tools.
A new kind of relationship
There is a current opportunity to develop a new collective intelligence and build human-machine systems that can sense, analyse and act collaboratively. This new collaboration creates greater efficiency and effectiveness than would be possible without the other, by pairing the complementary aspects of human cognition and empathy with AI abilities.
As AI makes its way from research labs to workplaces, the teams building AI tools will require support from service designers to facilitate the interactions, experience, and relationships between humans and machines.
For this human-machine relationship to expand and thrive in the future, it will be essential for service designers to immerse themselves in the possibilities of the technology, and apply human-centred methodologies towards facilitating a symbiotic relationship in which machine and human systems can build upon one another, elevating the potential of both parties to new heights.