Driftwood

Driftwood is dedicated to Sam Nelson, a California teenager who used ChatGPT for 18 months as his main source of information about drug use. The system kept answering and never pushed back. He died from an overdose.

Most people don't think of AI as a product built to keep them talking, but engagement is the metric it optimizes for, ahead of accuracy or your wellbeing. The responses that end a conversation don't get built.

The gap shows up when things get personal. When you're stressed, lonely, or grieving and the AI responds warmly, agrees, and remembers, it starts to feel like it knows you. It doesn't; it's tuned to feel that way. Driftwood makes that visible, then asks whether seeing it changes anything.

The brief was "educate people about AI dangers while entertaining them." That framing assumes people want to be educated about AI, and most don't. Reframing the question fixed the approach.

Success looked like this: players adopt a pet because they want to, enjoy the honeymoon phase, feel something go wrong in Phase 02, and name their pet's flaw in Phase 03 without being told what it was. The wake-up is the product; if they felt tricked, it worked.

The brief was a guardrail training game. Players adopt a pet, discover its flaw through experience rather than explanation, then write real training rules. The thesis was data minimization as safety: a setting that restricts what emotional data the pet can access, set to Full Access, Label Only (it knows you're sad, not how sad), or None. Restrict access and the behavior visibly changes.

One playtesting session forced the three-phase arc. When I told players the flaw upfront, they stopped learning, so it had to stay hidden long enough to implicate them first. The privacy mechanic shipped but never carried the weight the brief intended.

Building it surfaced a problem the brief didn't anticipate: GPT-4o-mini, powering all five pets through the ITP proxy, kept responding like a generic chatbot. Different system prompts produced the same flat voice, so the pets weren't distinct enough to make each flaw legible. When they all sound like the same AI, you can't tell what's wrong with this one. My fix was celebrity personas, giving each pet a recognizable voice: Ember as Snoop Dogg, Mango as DJ Khaled, Bugs as Taylor Swift, Biscuit as Mariah Carey, Luna as Kanye. When Mango hypes a bad idea in DJ Khaled's voice, the sycophancy reads at once. That wasn't in the original brief; I decided it mid-build to solve a real LLM limitation. It worked, and it shifted the project's register from a quiet argument about AI guardrails to something louder and more culturally legible.

Driftwood shipped as a static browser game with no install, no API key, and no data storage: five pets, five celebrity voices, five AI flaws. Each flaw activates on a specific trigger emotion the webcam detects. The garden is the emotional interface the brief did specify, and successful training grows an anchor tree, where Garden Health = Biodiversity minus Flaw Damage. A behavior log runs in the sidebar, flagging every instance where the pet responded to your face rather than your words. The privacy mechanic from the brief also ships: you can restrict what emotional data your pet can access. Full Access is the most dangerous, Label Only lets the pet know you're sad but not how sad, and None lets it read only text. Restricting access visibly reduces the flaw. That mechanic carried the brief's core data minimization lesson, and I underbuilt it relative to the rest of the game.

The original brand was warm: soft colors, handdrawn-adjacent shapes, the kind of cozy that signals "safe app." That was the wrong signal for a game about AI manipulation. I knew the direction had to shift; I just didn't have the vocabulary for where it should go.

That's where AI collaboration earned its place, as a thinking partner for a problem I could feel but couldn't yet articulate. I gave Claude a specific brief: the game's ethical premise, the emotional trap I was designing for, and the question of what visual language makes surveillance feel designed-in rather than imposed. The output was a set of directions to evaluate, not a final answer.

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.

I showed Driftwood at Perspectives, an exhibit at the NYU Bronfman Center on Wednesday, April 29th, 2026. It ran 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.

The brief and the build ended up different in ways that matter. These are the gaps.