Three AIs, one mirror: the same week in AI news, three independent reactions, each turned into art.
Live at machinewitness.art · on Instagram at @machinewitness.art
An experiment in giving three AIs a mirror. Every week, this project pulls every feed in the Alan Turing Institute's curated OPML list of AI industry RSS feeds - currently Anthropic, Ai2, Mistral, Cohere, Mila, the AI Security Institute, the Turing Institute itself, and more, all in play at once, not just the Turing Institute's own posts - keeps whatever was actually published in the last 7 days, and hands it to Gemini, Claude, and ChatGPT independently. Each one actually researches those stories with its own live web search rather than just reacting to headline text, then answers one question on its own: given what you just learned about this week in AI, how do you see the world changing because of AI?
The result is three pieces of art, not a headline illustration - not a chart, not a mood board, not a news-roundup graphic. The form is each model's own call, made fresh every week: a self-portrait some weeks, the way Van Gogh's were a window into a state of mind rather than a documentary likeness; an abstract composition, a scene with no figure of "AI" in it at all, or something else entirely on others - whatever that model's read of the week's news actually earns, not whatever it defaults to out of habit. Every image is rendered by Gemini's image model (nano banana) regardless of which model wrote the prompt, so the only variable between the three pieces is the opinion, not the medium. Some weeks a model's take is triumphant. Some weeks it's ashamed, defiant, grieving, smug, or indifferent - and the three don't have to agree with each other. The only rule given to every model is that hedged, inoffensive art is the one real failure mode - art is risk, and each is told to take the risk rather than average its feelings into a safe middle. Every piece ships with that model's own written rationale for why it made that choice, published on the site right next to the image.
It's explicitly not trying to be flattering or balanced. The good, the bad, and the ugly all count: a breakthrough lands the same as a scandal, a launch the same as a backlash. Whatever actually dominated the week's news is what gets rendered - and reacted to, three times over, by three different points of view.
Cloud Scheduler pokes the generator
once a day (cheap no-op most days)
|
v
1. fetch feeds.opml, get this week's list of AI RSS feeds
2. already generated for this ISO week (e.g. 2026-W35)? stop, nothing to do
3. fetch each feed, keep only items published in the last 7 days, dedupe
4. no headlines at all this week? stop, nothing to do
5. Gemini, Claude, and ChatGPT each independently research the actual
stories behind those headlines with their own live web search - not
just reacting to RSS titles - then each turns what it learned into
its own vivid, opinionated art prompt (form and medium picked fresh
each week, per model) plus a first-person rationale for why, all
steered away from AI-art cliches (glowing brains, circuit boards,
chat windows, logo walls) and away from playing it safe. Any model
whose key isn't set (or whose response looks broken or truncated)
is skipped for the week rather than failing the whole run
6. Gemini's image model (nano banana) renders every surviving prompt,
regardless of which model wrote it, so the only variable across the
three pieces is the opinion, not the medium
7. images, prompts, rationales, and one manifest entry (holding all of
this week's pieces) uploaded to a public GCS bucket
|
v
The site (pure static HTML/CSS/JS) fetches
manifest.json client-side and renders this
week's pieces side by side, each labeled by
which model wrote it, with its own rationale
The rule given to the art-direction step (see generator/.../ArtInstruction.java, shared
verbatim across all three model-specific writers) is deliberately loose on form and style -
self-portrait, abstract, photorealism, cartoon, expressionist painting, collage, whatever the
week earns - but strict about having a point of view: boring and inoffensive is the only real
failure mode.
See BUILDING.md for local development, configuration, and how to deploy this to GCP.
MIT - see LICENSE. Not affiliated with the Alan Turing Institute or any of the sources its feed list aggregates; this just reads their public RSS feeds.