Jenny Kang, MPH, is a Program Manager at Stanford’s Global Child Health Program in the Department of Pediatrics

When I want to purchase a good or service as a Stanford employee, I’m reminded in the portals and platforms where I can do so of Stanford’s commitment to responsible purchasing. I can easily compare the carbon footprint of one flight with another, and am told that my choices are helping the institution “achieve its goal of net zero greenhouse gas emissions by 2050.”

That commitment, and any reference to responsibility or environmental stewardship appear entirely absent, however, in my engagement with the University-approved AI platforms, such as Stanford’s AI Playground, the GSE’s AI Tinkery, or the School of Medicine’s Secure GPT instance. Their landing pages will warn me about data security, privacy, and offer caveats about AI hallucinations. But nothing about the environmental harms or tradeoffs of use.

Only buried in an infographic in the FAQs of the AI Playground, for example, will you find a broad estimate for the use of these tools: “Generating one image with GenAI with a commercially available large language model (LLM) uses as much energy as fully charging a smartphone.” And, “one text-based query to a GenAI LLM uses less than 1% of the energy required to charge a smartphone or approximately as much electricity as powering one LED lightbulb for about 20 minutes.” 

These are followed by references to articles warning about the “hidden environmental costs of using AI chatbots.” Buried elsewhere in the FAQ — indicating I am not alone in my concerns and questions about the environmental implications of these resources — is the reassurance that the cloud infrastructure powering our instances of these AI tools is hosted in regions that are rated “low carbon.” Low, but not zero carbon. 

And what of these tools’ water consumption? The 2026 AI Index Report, led by Stanford’s own Institute for Human-Centered AI, estimated that annual water use from just one of these models, GPT-4o, may exceed the drinking water needs of 1.2 million people. What of the increased use of PFAS — the “forever chemicals” required to cool the extremely hot computers behind these tools — that are increasing as new, more powerful AI models are developed?

Part of this lack of information is due to the opaque nature of the AI industry, with companies fiercely guarding their proprietary data with the reckless abandon characterized most famously in the early Facebook motto, “move fast and break things.”

As someone who studies the health of children around the world, I know that many human and environmental costs are not accounted for in the costing of new technologies. Instead, they are deferred to future generations who bear the brunt and to communities with fewer resources to identify or remedy these costs. That I don’t pay to use these tools with cash or tokens does not mean there are no costs borne by some other person or community.

It’s not for nothing that these portals to AI use are marketed to us as ‘playgrounds,’ ‘sandboxes’ and ‘tinkeries’; they conjure a world in which we can play like children, freely with no consequences to inhibit us. In reality, it is more likely that unchecked overuse of these tools, much like fossil fuels and plastics, will result in the very opposite: a world in which we see more suffering and less freedom from harm. 

My training and personal values prevent me from treating these powerful, energy-hungry new tools as mere playthings. The very existence of a “human-centered” AI initiative at Stanford suggests to me that others on this campus recognize that we need our human leaders to deliberately center human interests in how we teach, research, and develop policy around these new tools. Recent calls from engineers and leaders involved in developing AI to slow down the speed of development enough to build better guardrails echo this sentiment. 

I believe that is why we need more information to equip us to make responsible, informed decisions about how we engage with these tools. In practical terms, this would require leadership, perhaps in collaboration with peer institutions, to take the lead in demanding more transparency from AI vendors, and negotiations to reduce negative environmental impacts. A carbon footprint or water-use estimator, similar to what is visible automatically when we book flights via Stanford Travel vendors, could be set up. Warnings about what costs are known and unknown should be presented alongside the other warnings in the University-promoted AI pages. If we don’t know something, we should have the courage to say so. And, as a research institution, we can investigate how improving awareness about these costs might influence users’ engagement with these tools.

I look forward to a future in which I am empowered by our institution with at least the same amount of information to make a responsible choice in how to engage with these AI resources as I do in purchasing a box of pencils or an airplane ticket.