Making AI use sustainable: the SUS-AI project

We are still in the very early days of widespread AI usage; the impacts, however, are becoming increasingly clear. In my immediate environment, there are large numbers of hard-working undergraduate and graduate students who face increasingly complex concepts in their courses, heavy workloads, and co-op jobs that sometimes start only a few months into their degree program. AI augmentation of our workload has great potential for benefits (e.g. improved efficiency, shorter working days) and, I would argue, greater potential for drawbacks and even dangers. In the university environment, I am increasingly observing students and employees struggling to benefit from AI and frequently getting themselves into bad situations. While resources, guidance, and training are increasingly being made available to all of us around AI concepts and usage, we are still in the early days and free time to leverage these resources is limited. I have started the SUS-AI project to try to provide a short, simple, and easily understandable “starting-point” resource for those who want to use AI sustainably and safely.

Much of what separates AI tools that help from AI tools that harm are the practices around their use. The research keeps finding the same pattern: unguarded AI use erodes engagement with one’s own work, critical thinking, and unassisted performance, while guardrail-equipped use largely eliminates the harm.¹² The energy costs and environmental impact of AI usage are a significant concern as well, and they are concrete: measured per-query energy varies by one to two orders of magnitude with model size and query shape, from roughly 0.1 to 0.6 watt-hours for a ~120B open-weight model versus 1 to 16 watt-hours for a ~670B model, against 0.3 watt-hours for a typical commercial chat query. Which model you point at which task matters.³⁴

None of this says avoid AI. The evidence points in the other direction: a randomized controlled trial of a pedagogy-informed AI tutor produced learning gains larger than in-class active learning.⁵ Guardrails are supported and not abstinence; those practices are learnable, checkable, and they are what this note is about.

The SUS-AI project

#

The SUS-AI project is a public, evolving, open-source collection of principles, guidelines, checks, and recipes for sustainable AI use. Sustainable AI use can be viewed from at least two perspectives: (i) human health and performance and (ii) tools and infrastructure. For the humans using AI tools, sustainability means building and retaining your skills, judgment, and autonomy. For the tools and infrastructure you use, sustainability means ensuring that they maintain data security, reproducibility, and resource efficiency.

The SUS-AI project has four key parts. Nine principles, from “guardrails, not abstinence” through “practices decay; revisit and revise continuously.” Guidelines in two directions, one for the human using the tools and one written to be handed to your AI agents as standing instructions. Twenty-one “SUS checks,” binary checks you test against your own usage and setup, such as whether any credential-bearing text appears in your AI tool’s history, or whether you can name where your data goes after you press enter. Ten recipes, setup instructions your own AI tool can execute to fix what a check finds. The reading order is short: read the principles, run the checks, fix what fails.

The project is open-source with a CC BY 4.0 license and welcomes external contributions and criticism. SUS-AI is not a certification or training; it requires very little time and is targeted at a non-specialist audience but, necessarily, does involve some complex topics around AI use. The project is meant to quickly and easily provide new or existing AI tool users with principles and guidelines for risk reduction that can be verified (via SUS checks) on their own.

A complementary set of open-source presentations is also being developed along two strands:

• one on responsible AI usage, from sustainable usage patterns and the psychology of AI dependence through graduate research practice and the professional responsibility that arrives with the first engineering job;
• the other dissects a working AI assistant setup component by component: models, tools, memory, external and sensitive access, and orchestration, ending with the assembly of a minimal viable setup.

The presentations are meant for teaching and learning, independent of SUS-AI usage. The initial target audience will be undergraduate and graduate students at the University of Waterloo, but, if reception is positive, I will create high-quality recordings for wider audiences.

SUS-AI Adoption

#

If you currently use AI tools or are planning to, consider reading the SUS-AI principles as a starting point. If you find them useful, run the SUS checks against your own setup, minutes each, producing either reassurance or a specific fix. Fix what fails, using the recipes as instructions your own AI tool can execute. Then revisit occasionally, because every practice in the framework decays as tools change. The framework asks nothing you cannot verify yourself, and it asks it in minutes.

Beyond SUS-AI - Existing Training Resources

#

The SUS-AI project is a starting-point for sustainable AI usage, not an end-to-end training program or practical method. There are many resources available, mainly targeted at university students, that could and should be explored and learned from. The tables that follow provide a summary of some of the larger and more mature resources available. One deserves special mention because of its quality and breadth: the Elements of AI courses from the University of Helsinki.

Table 1: Non-proprietary, tool-agnostic resources.

Table 2: Vendor-tied resources (not recommended, against SUS-AI principles P3+P5).

Footnotes and references compiled with an open-weight AI model.