Alternative AI Futures

Alternative AI Futures is a hands-on workshop series that invites students to directly experiment with, critique, and modify contemporary AI systems in order to reimage how different design choices, cultural values, and shared goals could produce radically different forms of AI. 

Each workshop will begin with a simple but provocative prompt: What if AI were built differently? More specifically, what if factors like transparency, accountability, and social responsibility were the primary values guiding AI system design? What would those systems look like? How would they behave? And how might they address commonly held concerns about the structural limitations and/or social consequences of AI? These “what if” questions will foreground the assumptions embedded in current AI systems and open space for imagining alternative futures for how AI might be built and used differently.

These 1.5-hour workshops will be split between hands-on experimentation and discussion. Students will work directly with small-scale machine learning systems and language models, using tools like Google Collab Notebooks to modify and reconfigure how these systems behave. Through guided, hands-on exploration, students will experiment with training data, embedding spaces, system instructions, and model parameters, among other elements, to see how changes at the technical level can give rise to alternative forms of AI in the form of small, working prototypes. The discussion portion of each workshop will build on this hands-on work, allowing students to reflect on how technical design choices give rise to broader cultural, social, ethical, and political impacts.

Workshop 1: AI as an Instrument

Token Predictor 

Students generate multi-token sequences and observe how individual predictions compound into seemingly coherent text. Each generated word is color-coded by the model's confidence, making visible the uneven terrain of certainty and uncertainty that underlies every generated sentence. Some words appear with high confidence, typically common function words and predictable continuations, while others appear with notably lower confidence, marking positions where the model could easily have gone a different direction. The sentence reads as fluent prose, but the confidence coloring reveals that it was not produced with uniform certainty. It was assembled one word at a time, with some steps nearly guaranteed and others closer to a coin toss.

Clicking any word opens a panel showing the alternatives the model considered at that step, ranked by probability. Students will often find that at low-confidence positions, several candidates were nearly as likely as the one the model chose. Verbs are particularly instructive: after a noun, students may see six different verbs among the top candidates, and the verb that was selected steers the rest of the sentence in a specific direction. But the model did not pick that verb because it planned the rest of the sentence. It picked a word, and now it is committed to a trajectory. The words that came before paint the model into a corner, constraining what can plausibly follow, and the model continues generating the next most plausible word in front of the last, regardless of where the sentence is headed. This is also the mechanism behind hallucination: when the probability distribution is relatively flat and the model is, in a sense, guessing, it can select a plausible-sounding word that commits it to a factual claim it cannot support. The self-reinforcing nature of autoregressive generation then locks the model into elaborating on that claim with the same fluency and confidence it brings to everything else.

Students can click an alternative word to branch the generation, substituting a different choice and watching the text diverge from that point forward. The model regenerates everything after the branch point, producing a new sentence that may differ subtly or dramatically from the original. Both versions are equally fluent. Both read as confident, coherent prose. But they may say entirely different things, because a single different word at an early step sent the generation down a different path. The branching mechanic makes the contingency of generated text concrete: every sentence the model produces is one path through a tree of possibilities, and the model will complete any path it is placed on with equal fluency, not because it believes what it is saying, but because that is what autoregressive generation does.

By exposing these stages and letting students manipulate them directly, the interface reframes AI-generated text not as the product of understanding or intelligence, but as the output of a specific, modifiable, and contestable technical pipeline.

Go to Google Colab Notebook (next_token_explorer.ipynb).

Go to Token Predictor web interface.

Download instructions for Notebook and interface.

Download sample exercise for Token Predictor.

Workshop 2: AI as Environmental Cost

What if AI systems made legible their use of energy and water in ways that could meaningfully inform users’ interactions?

We tend to treat AI as an infinite resource, as instant, frictionless, and detached from the material reality that makes it possible. This workshop invites students to reimagine alternative AI systems in which environmental costs are visible and legible, by first confronting how thoroughly it has been hidden. A chat window gives no indication that each response draws on a vast physical infrastructure: fossil-fired power plants, electrical grids, data centers, and water for cooling. That absence is an intentional design decision by frontier AI labs. A chat interface could show what a conversation cost, but that would require two disclosures these companies have declined to make. The first is local: tokens are metered precisely enough to bill enterprise customers by the million, yet casual users see no count at all, no sense that a long conversation re-sends everything said so far or that asking for a search pulls in thousands of tokens of web content. The second is global: even a visible token count means nothing without a credible figure for what a token costs in energy, carbon, and water, and no company has published one. Independent researchers reverse-engineer estimates from leaked architecture details, API pricing ratios, and occasional corporate disclosures, which is why two credible published figures can differ by a factor of three. 

Working with a custom-built browser extension for ChatGPT, students in this workshop will examine their own token counts as they accumulate in real time; see the estimated energy, carbon, and water cost of a full conversation; compare competing estimates of what those tokens cost; and conduct their own research into how those figures are produced. 

Environmental Impact Tracker showing token accumulation; estimated energy, CO₂, and water costs; and selectable models for per-token impact rates.

Environmental Impact Tracker (Chrome Browser Extension) 

The Environmental Impact Panel is a Chrome extension that adds a sidebar to ChatGPT, enabling students to observe in real time how different types of tokens (input, output, and search) accumulate as they interact with the model; how those token types grow at different rates relative to one another (input tokens accelerating with each turn as the full conversation history is re-sent, output tokens reflecting the length of each response, search tokens appearing only when the model ingests web content); and how those token counts translate into estimated energy consumption, CO₂ emissions, and water usage.

By offering three different sets of per-token environmental rates, each drawn from a different published source with a different methodology, the extension surfaces the fact that measuring the environmental cost of AI is not a matter of looking up a single authoritative number but rather involves navigating a complex landscape of estimates, assumptions, and undisclosed variables, where the companies that have the actual data have chosen not to publish it and independent researchers are left to reverse-engineer approximations from leaked architecture details, API pricing ratios, and occasional corporate disclosures.

Thus, the extension functions not so much as a measurement tool but as a design provocation: by making token accumulation and environmental impact visible within the same interface where students are already using AI, it invites them to consider what ChatGPT would look like if this information were always present, how their own behavior might change if the cost of each interaction were legible, and what it means that the companies building these systems have deliberately chosen to keep it hidden.

Download .zip of Environmental Impact Tracker Chrome Extension

Download set up instructions and sample lesson for Environmental Impact Tracker Extension.