It’s 2023 and we are in the early days of artificial intelligence design.
We’re certainly not “there” yet (i.e., you’re not running into Commander Data at the bar) but we do have a substantial cadre of narrow AI capabilities sprinkled throughout our daily lives. You can get a computer to schedule meetings for you, answer your HR questions, or help you parallel park. More seriously, you can also get a computer to do critical thinking on your behalf — thinking that dramatically impacts peoples’ lives — like recommending salary increases, qualifying loan recipients, and informing prison sentences.
During my career as a design researcher, a design manager, and a design principal delivering non-AI and AI software, I have observed the development of a unique “AI Design” skill set in myself and on my teams. I think it is a skill set worth capturing. Last year I sat down with a couple of colleagues at IBM who also work in AI design, Adam Cutler and Milena Pribic, and together we defined a unique set of skills for “AI Design.”
Designing for AI means designing for a human-machine relationship that is in flux. This is in contrast to non-AI design where the relationship is static.
While with non-AI machines, the interactions do not change, with AI machines, the interactions evolve over time. The machine is learning, and therefore, changing. In parallel, the human is learning, and therefore, changing. This co-learning creates a dynamic feedback loop — human and machine trading information back and forth, learning as they go. Creating the conditions for good communication, in this dynamic human-machine relationship, is the formidable task of the AI designer.
In order to facilitate this rich back-and-forth relationship between humans and AI, data takes center stage. Before AI, data transfer looked like a human pressing start, and a machine turning on. With AI, data is now the raw material of learned behaviors: data inputs are about providing the machine with facts, not directives. And machine outputs, not pre-defined as they were before AI, require explanations about how the data generated a response. An AI designer is responsible for designing these interactions across the data lifecycle, from data capture to data output, through data explanation, and back again.
Consider for example designing a record player vs. designing Spotify.
The ideal behavior or “output” of a record player is predictable. Place a record on the platter, drop the needle, and it should play that record exactly the same way, every time, for as long as it can.
Contrast this with Spotify which uses machine learning to recommend new songs and create custom playlists (among other things.) In this case, the Spotify outputs get better with use. The software improves by capturing information about you, overtly or covertly, and uses that input to make decisions about what songs to serve you. When Spotify makes a decision about what you want to hear, it can give you varying levels of insight into the machine’s “thought process” and varying levels of control over how you can manipulate the machine’s outcomes.
In both of these cases, the basic principles of design apply — things like simplicity and consistency are essential. Underneath those basics, though, are two different kinds of technology — one fundamentally static and one fundamentally dynamic. What used to be a one-way street — human pushes a button, machine responds as programmed — is now a two-way street, creating a cycle of change. As a designer, that means you’re now in charge of how well (read: effectively, easily, safely) the machine and human trade information back and forth. In AI design, human-machine interaction is not about designing for day 0 and inevitable degradation, but designing the parameters of an evolving and ongoing relationship.
The short-term, practical value of defining AI design skills is straightforward: designers need to know their responsibilities. Designers working on AI technology need to know what is expected of them, so they can meet those expectations or grow as needed. In concert, non-design team members need to know what they can expect from a designer.
This is reason enough for defining AI design skills, but in addition to this near-term functionality, is the valuable start of a long-term conversation.
Through the narrow AI available today, we are getting a taste of what it means to create machines that learn, reason, sense, and act like humans. These machines need to be engineered (i.e., made to work) but they also need to be designed (i.e., made with purpose.) Defining AI design skills while we’re still in the sandbox is prudent. This way, as machines evolve, our understanding of how to create quality AI can evolve with them. This is useful for creating highly functional AI, but it becomes even more important as we begin to outsource human decision-making capabilities to machines. Defining these skills gives shape to a profession that will be responsible for designing not just functional, but ethical and safe, AI.
The scope of AI design skill is not limited to what is visible “on the glass” (i.e., putting pixels on the screen). Instead, designing for this dynamic human-machine interaction changes the end-to-end process of design. AI design starts when a team chooses a problem to solve (and defines what data they need to solve that problem) and never stops even after the design is released (when learning must be tracked and revisions need to be made.) This new end-to-end process means new collaborative partners, new market strategies, new ethical considerations, and new technical knowledge.
An AI designer should have proficiency in the following five skills.
Understanding Data Science and AI Terminology and Techniques
AI designers should have a foundational understanding of the domain of artificial intelligence technology. They should know enough to be able to influence the development of AI and to be held accountable for how AI is used. At the highest level, an AI Designer should be able to have an informed and lengthy conversation about data science.
Observable Behaviors
- Can explain the basics of how machine learning works
- Can explain Data Engineering concepts (e.g., Data Collection, Data cleaning, Data curation, Data Modeling)
- Can explain AI techniques (e.g., NLP, NL understanding, ML)
- Can explain how core algorithms work (e.g., Deep Learning, Create Libraries, code patterns)
- Can explain AI toolchain and AI DevOps processes (e.g., WML, Cloud Pak for Data)
Designing for AI Interactions
AI designers are responsible for the iterative and dynamic relationship between humans and AI machines. AI designers make the outputs of an AI model consumable so humans have insight into a machine’s evolving reasoning and results. In parallel, AI designers create interactions that capture human input or preferences, enabling machines to improve over time. At the highest level, an AI designer should be able to produce a variety of consistently high-quality AI interactions, defining standards and best practices for other designers.
Observable Behaviors
- Can translate AI outputs into valuable pieces of information for humans
- Can design for multiple “universal experiences” in AI design from “getting started” to “support”
- Can leverage and reuse existing AI patterns effectively and appropriately
- Can determine the appropriate level of communication about the existence of AI (e.g., frequency, branding)
- Can prototype and test AI solutions early and often
- Can design for common concepts in the AI lifecycle (e.g., setting confidence thresholds, setting up a sustainable model maintenance program, and data collection methods)
To grow your skills (or the skills of your organization) in AI design, here are a couple of things you can do right now.
First, do a self-assessment. Review each of the skills above and give yourself a 0–5 rating, with 0 equating to “No Skill,” and 5 equating to “Expert.” Setting this baseline will help you understand where you are and how you need to grow.
Second, identify one dimension you want to develop. I find that if you choose one dimension for growth, the focus helps you make progress. This is better than trying to get better at everything at once. (A nice side effect of this is that all the skills tend to get better if you focus on just one.)
These skills were co-written with Milena Pribic and Adam Cutler. This work was created with iterative feedback from IBM’s 2021 AI Design guild members including Mats Gothe, Steven Chang, Ciera Raines, Gina Rinalli, Beverly Hrablook, Micheal Zuliani, Diana Tran, Emily Dicesaro, and Chris Noessel. Initial research for this project was done by and with Jenn Aue and Sheri Terwin.
If you work in AI design and you have some thoughts on this subject, I’d love it if you wrote me and told me either how this resonated with you, or what you would change.