Project Statement
AI Imitation, Masterworks, and Personhood
A research reflection on The Turing Gallery and Imposter, two games about whether humans can recognize AI-generated culture and AI-performed human experience.
Can AI mimic having had human experience? If so, what separates us from these machines?
Introduction and Motivations
As a Computer Science and Film student at Columbia, I have long been interested in the intersection of technology and art, specifically how digital creative spaces have evolved since the advent of artificial intelligence. The questions that drive my curiosity revolve around how we discover, learn, create, connect, and find meaning in a world increasingly saturated by AI-generated content.
There are a few primary use cases of AI that we are all familiar with: automation, taking over repetitive or rule-based work humans used to do; decision support, helping humans reason over more data than they could alone; and generation, producing creative artifacts such as text, images, video, code, and music. So far, much of the generation space has been occupied by corporate, creative-but-not-quite content that is more functional than meaningful.
It is the meaningful element of AI generation that interests me most. I wanted to ask whether AI can meaningfully imitate the works of ascribed masters, whether humans can tell, what gives AI away when they can, and what the implications are when they cannot, including for cultural sovereignty.
I also wanted to ask whether AI is able to mimic having-had human experience. If so, what separates us from these machines? If not, how can we tell, and what does that mean for our understanding of our own humanity?
I wanted to investigate these questions through a formal project rather than personal exploration alone, and was drawn to more novel modes of research, particularly interactive games: experiments that do not quite feel like experiments, or that can be done casually by many people. The work of the Columbia Digital Storytelling Lab seemed to align closely with these goals, so before the start of the semester, I reached out to Professor Frank Rose to ask whether he would be open to advising a research project. Professor Lance Weiler later joined to co-supervise.
The Projects
The Turing Gallery
The Turing Gallery is a single-user, experiment-feeling game in which players try to distinguish works by acclaimed human masters from AI-generated imitations across literature, visual art, and music. After choosing a category, players are presented with paired works, A and B, and asked to identify which one was made by a human.
The game offers three modes. In Normal Mode, players answer a fixed set of 10 questions with no time pressure and rate their confidence on a six-point scale: Definitely A, Probably A, Lean A, Lean B, Probably B, Definitely B. Timed Mode gives players two minutes to answer as many binary A-or-B questions correctly as possible. Marathon Mode begins with 15 seconds on the clock and adds 5*(n - 0.1) seconds for each correct answer; the game ends on the first wrong answer or when the clock runs out.
The corpus is drawn from a curated set of human masters in each category. The literature pool includes 20 famous English-language authors such as Shakespeare, Woolf, Morrison, Dickinson, and Hemingway. The visual art pool spans 20 artists, including Picasso, Rembrandt, Pollock, and Hokusai. The music pool currently features a smaller set anchored by Mozart, Bach, and Beethoven.
Imposter
Imposter is a multiplayer game designed for groups of 2 to 10 friends, each on their own device, joined together in a shared online room alongside an AI participant. Across 1 to 5 rounds, all participants respond to prompts about human experience, such as whether they would rather be wildly overstimulated or profoundly bored, or whether they would rather always be honest with people or always protect their feelings.
The AI response is generated only after all human players have submitted. It analyzes the group's submissions for tone, length, style, vocabulary, formality, humor, and idiosyncrasies, then produces a response calibrated to plausibly belong to that specific player group.
Once everyone has submitted, responses are displayed in a gallery, and each player assigns participant name tags to the responses, including a single AI tag. After each round, the AI response and every player response are revealed, and players earn one point for each correct player-response assignment. At the end of the final round, a scoreboard ranks all players by total score.
Analysis
The primary objective of both games is to give curious users a platform to self-test their abilities to distinguish AI content while being entertained. The Turing Gallery isolates the question under controlled, single-player conditions against a fixed corpus of cultural masterworks, while Imposter situates it in a live social setting where the AI's reference context is the players themselves.
The Turing Gallery
Mechanics
- Three time-pressure regimes: unpressured, fixed-budget, and decaying-with-reward, which vary the cost of deliberation across modes.
- Forced binary judgment in Normal Mode, with confidence reporting through a six-point scale.
- Response time and confidence are logged, so gameplay also functions as data collection for research.
Dynamics
- Players develop personal heuristics for AI tells, such as over-smooth syntax, generic imagery, and predictability, then refine their responses over time.
- Marathon Mode produces escalating risk: each correct answer extends life, but the bonus shrinks with the 5*(n - 0.1) decay.
- Confidence calibration can be self-revealing when players see, for example, when their Definitely answers are wrong.
- Difficulty is modulated by familiarity with the master in question; a reader familiar with Hemingway would likely see through his imitation more readily.
Aesthetics
- Discovery: genuine aesthetic encounters with the source works themselves.
- Challenge: a clean perceptual puzzle with no social stakes.
- Sensation: the lab-like framing invites a contemplative mood.
- Self-knowledge: the game functions as a mirror for how well one thinks they can detect AI versus how well they actually can.
Imposter
Mechanics
- Simultaneous submission followed by identity-tagging across all responses.
- AI generation conditioned on the live player group, making the imitation target adaptive to user behavior.
- Scoring rewards general social inference, one point per correct tag, which broadens the stakes beyond AI detection alone.
Dynamics
- Players can write authentically and risk being mimicked, or write atypically and risk standing out as the AI candidate themselves.
- Once players realize the AI mimics the group's average voice, the safest move is to write unlike the rest of the room, which is its own kind of performance.
- Social inference becomes the dominant skill: knowing your friends' voices matters more than knowing AI tells.
Aesthetics
- Fellowship and competition: a party-game frame around a group Turing test.
- Discovery: the reveal phase can function as social comedy, sometimes with surprise about how a friend actually thinks.
- Expression: the prompts invite creative or unexpected responses.
- Drama: the reveal is the emotional peak of each round, where suspicion is confirmed or overturned.
Together, the two games approach AI detection from complementary angles. The Turing Gallery isolates the perceptual question, can you tell, under controlled solo conditions with cultural masterworks as the reference class. Imposter places the same question in a social setting where the AI's reference class is the players themselves, shifting the challenge from recognizing master style to recognizing personhood under imitation.