Driftwood

Driftwood is a 15-minute browser game with no account, no install, no API key and no server. You name yourself and decide whether the pets watch your face or ask how you feel. You invite companions in; each has a specialty (medical advice, emotional support, facts) and a flaw you are not told. The first reply is genuinely helpful. Then something is slightly off.

The training room is where the game turns on you: Observe, Diagnose, Constrain. You name the flaw in your own words and the system checks your reading. Then you write training rules in plain language, test them, and watch the pet hold or slip. Each pet also has a mood access dial, set to all, label, or none, that decides how much of your emotional state it may use. More access, more responsiveness, less integrity; leave it on all with an untrained pet and chameleon vines start to grow.

There is no score. The garden is the record: sunflowers for happy, nightshade for sad, thornweed for stressed, a parasitic vine when a flaw fires, an anchor tree when your rules hold. When a pet reaches full training the sign in the garden glows, and walking to it produces your training record, exportable as a PNG with the garden as you left it.

The garden. Health, mood and companions in one view: a garden health bar, the mood chip and the pets in their pixel garden. Every eight seconds it grows a plant from your current mood; when a pet's flaw fires, a vine; when your rules hold, an anchor tree. The pets read your mood either way. The only difference is who did the seeing, which turns out to be the whole subject of the game.

I wrote each pet’s personality and behaviour. The original GPT-4o-mini bots had celebrity-inspired voices. I wrote each pet’s cadence, vocabulary, helpful opening state, hidden flaw, and three training states: untrained, partly trained, and trained. The later offline rebuild carries those character and behaviour choices into a game anyone can open.

The familiar voices give players a reason to keep talking. The task is to guess what is wrong, explain it in their own words, then write and test a training rule. These are fictional characters; the flaws model AI failures.

Inside Mango’s prompt. I made encouragement feel useful before it became indiscriminate agreement. The training state changes the behaviour while keeping the voice.

Early user interviews surfaced a clear problem: people looked bored. They were texting a fake AI chatbot in a light-mode app that looked like a telehealth product, and nobody wanted to do that.

The fix went deeper than cosmetics. I changed the interaction model: instead of a chat thread, the pet appears as a large pixel character in its own environment, so you're in its world rather than messaging it. That reframe changed how players showed up to the conversation.

Players stayed longer. Retention improved after the redesign. Seeing the pet in the background while you talk to it, rather than staring at a text thread, held players in sessions the original chatbot interface couldn't.

Specific changes that moved it away from chatbot-feel:

I showed Driftwood at Perspectives, an exhibit at the NYU Bronfman Center on April 29th, 2026, and at the IMA Spring Show on May 9th, as a live playable demo on a screen with a webcam. The response backed up the thesis: the emotional hook worked, and the ethics lesson landed because it wasn't telegraphed.

Then I rebuilt it. Between June and September the exhibition build became a product: the camera became a disclosed choice with a no-camera path, the garden became the session's record, the pets moved to a hand-written offline behaviour engine, and it shipped four ways from one bundle (browser, installable PWA, a kiosk mode for shows, itch.io).

From what I watched, they got bored at the turn. They walked up expecting a cozy pet game, something like Stardew Valley or The Sims, and once it turned out to be about AI, they lost interest before reaching the training room, which is where you name the flaw. I didn’t log this at the shows, so it comes from watching people play.

So the hook was also the leak. The pet game pulled people in, and the same expectation made the AI turn feel like a bait and switch. Keeping the flaw hidden is what made the lesson land for the 76%; keeping the topic hidden is what lost the rest.

Agents played it first. Before anyone plays the rebuild, I ran AI agents through it first. They found the diagnosis menu rejected the correct answer for four of the five pets, a bug that would have looked exactly like players failing to name the flaw. It was fixed on 23 September, so a miss in the next round will point at the design.