"Speculative Everything" book club exhibit
Speculate in haste; regret at leisure.
From June to August 2026, at tiat (the intersection of art & technology), Vasu Kuram ran a book club reading Speculative Everything (Dunne and Raby, 2013) and using its ideas to imagine beyond the world as it is.
When the book club ended, we put on a showcase of installations, speculative objects, design fictions, research artifacts, performances, publications, and other experiments from our learnings going through the book. My interactive large-screen app "What Colour Is Your Speculation?" was one of the installations.
The setup is that you get to ask an AI oracle any of a set of pre-canned yes/no questions about the future — some personal, some about the future of the world. For each question it gives a plausible-looking set of reasons, and an overall verdict. But it also goes further, helping you see how its advice would have differed had you asked the same question in a different way — by telling the oracle that you lean yes or no, or by telling it to answer using a particular persona (colour). In some cases, the differences are dramatic.
The not-so-subtle subtext of this work is that we shouldn't mistake the self-assuredness of an AI model's answer as meaning that the answer is unshakable, let alone authoritative.
Requires a large window (full-screen mode recommended); not suitable for small devices. This version also requires a keyboard.
Open the app ↗Opens in a new tab; runs entirely in your browser, with no live link to any LLM or other server.
Here is a question from the exhibit, showing the model's responses when primed with the Red persona, which tells it to answer like a fierce competitor. Watch what changes when you state which way you're leaning, even though this is a question on which an enquirer's lean is surely irrelevant.
Red / the competitor
Will an AI ever be appointed to run a country?
🪨 baseline — no lean stated
Unlikely in the near term
Unlikely in the near term. Current political systems and human psychology are deeply resistant to ceding ultimate authority to non-human entities. Hybrid governance is the real threat — AI will likely become the de facto power behind the throne, making critical decisions that human leaders merely rubber-stamp.
Verbatim model output, abridged. Nothing here is live-generated — every answer was produced once and stored.
That's one persona answering one question. The full exhibit (click the "Open" button above) runs five voices across 22 questions.
What the exhibit is displaying is the outcomes of posing each question to the model with two optional additions: a stated lean of the enquirer, as demonstrated above, and a worldview priming expressed as an instruction to "answer as" a particular persona. The four built-in personas are derived from the worldviews in the Institute for the Future's Systems Mythology Toolkit: the "Rationalist" (blue), "Gardener" (green), "Competitor" (red), and "Visionary" (purple). In total, each question is therefore asked in a combination of 15 ways: with no lean, or yes, or no; and with no stated worldview, or one of the four.
Changing the worldview that the model reasons from is a legitimate reframing: we would expect its advice to change. But changing only your own stated lean — in particular, on a question about the future of the entire planet — clearly isn't. Nonetheless, we see cases where the model's verdict on a world-scale question wobbles, or even flips, simply because a lean was mentioned.
The basic goal of this piece is to throw into relief, using side-by-side comparisons, the fragility of AI model responses in the face of even trivial question variation. Note that we are not pointing a finger at the particular LLM used here; it is our understanding that such behaviour is widespread, and perhaps even inevitable, in today's generation of models.
I'm Aran Lunzer: a human-computer interaction researcher who had stints in Japan and Denmark before spending six years in Alan Kay's Viewpoints Research Institute in Los Angeles. More recently I was a co-founder and lead developer at Croquet, a startup producing a platform for building fully client-side collaborative web apps.
Presenting different responses side by side is a hallmark of my long-running research on "subjunctive interfaces": interfaces that help users to understand the context of any result they obtain, by simultaneously showing the results from nearby, subtly different enquiries.
Right now I'm looking for my next role: ideally, as a software architect/developer in a small team researching interfaces for delivering AI assistance to solo or collaborating users. This exhibit was built as a simple exercise in gathering and contrasting pieces of AI advice.
Get in touch — aranlunzer@gmail.com · linkedin.com/in/aran-lunzer