They demonstrate this in a simplified design problem: planning a day trip in Paris. ‘An enjoyable day sightseeing’ may be a vague goal for a person, let alone an AI, but it turns out that a system that knows a little about human biases and can learn by observing their choices, combined with a little automation, produces better designs, in this case trip plans, than purely automated systems. ‘People tend to want to stick to the familiar, like a proven trip plan that isn’t too radical. That’s what we call an anchoring bias,’ explains De Peuter. ‘When we build in the capacity for the system to infer people’s biases, it outperforms, for example, inverse reinforcement learning, and improves the quality of advice provided to the user.’
The result allows the trip planner, or designer more broadly, to make high-level decisions, while the AI assistant will tackle the small tasks that make up most of the work—for example, safely assuming that no trip to Paris is complete without a visit to the Eiffel Tower. ‘Currently, the job of designers is to learn how to work with AI tools,’ say De Peuter. ‘With AI-assisted design, designers should be able to stop worrying about what the tool will do, because it will be in the background.’ He estimates that, in this simulated design task of planning a trip, about one-third of the designer’s effort, such as decision-making steps and iterations, is saved thanks to AI assistance.
However,De Peuter says this type of AI assistance will probably not be used to design handbags or cars, because they are one-off cases where the cost of representing and evaluating designs with AI vastly outweighs the benefit. ‘AI assistance is best applied to scientific or engineering fields that have the foundation for deploying AI-assisted design in a worthwhile way. In drug design, where you produce multiple drugs and the chemistry is the same, you can reuse AI systems easily. The investment required in building these systems means they are most useful in repeated use, not for one-off design problems,’ De Peuter clarifies.
Ultimately, AI-assisted design is about improving the productivity of designers. ‘Recommendation systems offer more of the same, exploiting what is already known, but good design needs exploration,’ says De Peuter. ‘An AI assistant can improve existing designs in small steps or explore the space more broadly, which is harder for people. It’s a tradeoff between effort and maximizing quality, with AI pulling up one strap and humans controlling the other.’
‘Helping designers is an important instance of our even broader mission, of building AI that is able to help humans reach their goals, even when they are initially unclear or evolving, as they often are in decision making, design and modeling. This work really shows the value of our collaboration at FCAI, ELLIS Unit Helsinki and my UKRI Turing AI World-Leading Researcher Fellowship program,’ says Samuel Kaski, professor at Aalto University’s Department of Computer Science, director of FCAI and the senior author of this research.
The perspective piece ‘Toward AI Assistants That Let Designers Design’ appears in the spring 2023 issue of AI Magazine.
This work was presented at the AAAI Conference on Artificial Intelligence in February 2023.
Reference: De Peuter, S., & Kaski, S. (2022). Zero-Shot Assistance in Sequential Decision Problems. https://arxiv.org/abs/2202.07364
