Large Language Models (LLMs) now perform strongly on Alan Turing's 1950 Imitation Game, so we propose a harder challenge, the Personal Imitation Game, which asks whether a machine can replicate a specific individual in conversation rather than a generic human. Testing with 30 participants, we establish a human baseline and report the identity signals interrogators rely on: stylometric features, personality traits, and biographical facts. To demonstrate the framework, we design an LLM-based system that plays the game for one of the paper's authors, drawing on an autoethnographic dataset spanning ten years of personal digital communication. In a 20-person deception study, interrogators choose our system over the author's own friends in 7 of 10 rounds, and the author over our system in 13 of 20; neither rate is distinguishable from chance at this sample size. We consider design improvements, applications of personal AI agents, and implications for identity and privacy.
Figures & tables
Figure 1. Three variants of the Personal Imitation Game. In each one, a human interrogator chats with two anonymized participants: a target — whose identity is known — and an impersonator who attempts to mimic the target’s identity. In the first variant (left), both the target and impersonator are humans, and all three participants are known to one another. In the second (center), the target is one of the paper’s authors, the interrogator is a (human) friend or colleague, and the impersonator is an LLM-based system powered by the author’s autoethnography, which includes ten years of personal message data. In the third (right), the LLM takes the target author’s place, and the impersonator is again a friend. Three horizontally arranged panels depict variants of the Personal Imitation Game. In each panel, three roles—Interrogator, Target, and Impersonator—are positioned radially around a central chat interface. Thin radial lines form pizza-slice separators between roles, indicating physical separation and text-only communication. Role labels and icons are present: a monocled thinking-face for the human Interrogator; a mustached human for a real human Target; a human with a fake mustache for a human Impersonator; and a robot icon for AI roles (the robot also wears a fake mustache when acting as the Impersonator). Chat bubbles are color-coded with a consistent mapping across panels—blue for Interrogator, green for Target, purple for Impersonator—supplemented by role labels and matching icon halos to avoid reliance on color alone. Left panel: all three roles are humans (all-human baseline). Center panel: human Target (the author), human Interrogator (friend/colleague), and an AI Impersonator grounded in the author's data. Right panel: the Target is an AI model standing in for the author, and the Impersonator is a human. The figure foregrounds text-only, turn-taking interaction and a controlled comparison across variants.
Figure 2. The chat interface powering the Personal Imitation Game. As the game begins, participants are presented with the first-round topic, complete player names, and their respective roles (left). During each game round, the interface automatically assigns random aliases for the impersonator and target, while the interrogator is shown as “I” (center). At the end of each round, the interrogator’s guess and confidence level are recorded (right). Screenshot of a three-panel interface for the Personal Imitation Game. Panel 1: Round Details screen visible to all participants showing the round topic, the viewer's role and chat alias, a goal reminder, and a roster listing player display names with roles. A scenario prompt explains the mission. A Ready to Begin button advances to the chat. Panel 2: live chat during gameplay with a round title, role reminder, and countdown timer. Messages appear as labeled bubbles with sender labels (X, Y, or I). Panel 3: end-of-round survey for the interrogator, asking who is the target (X or Y) with a confidence slider.
Figure 3. Successful interrogators have higher confidence (median = 87, IQR = [75, 100]) than unsuccessful interrogators (median = 76, IQR = [72, 80]). Whiskers are 1.5 × IQR fences and the horizontal axis is broken between 0 and 60. Three unsuccessful-interrogator values fall outside the fences and are not plotted (19, 50, and 100); they are included in the means and standard deviations reported in the text, which is why the reported σ=19.9 is larger than the plotted spread suggests. Boxplot comparing interrogators' confidence (%) by outcome. Two horizontal boxplots are shown side-by-side: the top boxplot represents Successful interrogators (green) and the bottom boxplot represents Unsuccessful interrogators (salmon). Successful: median = 87%, interquartile range (IQR) = 75–100%. Unsuccessful: median = 76%, IQR = 72–80%. Medians are marked by a line inside each box; whiskers extend to the furthest value within 1.5 times the IQR, so the unsuccessful whiskers run from 60 to 89 and three values outside those fences (19, 50, 100) are not drawn. The horizontal axis at the bottom is labeled "Interrogator's Confidence (%)" and is broken: it shows 0, then jumps to the 60–100 range in which the boxes lie. Category labels ("Successful," "Unsuccessful") appear on the vertical axis so the figure does not rely on color alone.
Figure 4. A heatmap showing the relationship between group tie strength TG (top-blue) and interrogator success rate (bottom-orange). Groups are sorted left to right from lowest to highest TG . Heatmap with two rows comparing group tie strength and interrogator success rate across ten groups, ordered left-to-right by increasing tie strength. Columns (left to right): G7, G9, G3, G8, G2, G10, G4, G6, G5, G1. Top row (blue) encodes group tie strength TG with values: G7=2.0, G9=2.7, G3=3.2, G8=3.3, G2=3.3, G10=3.3, G4=3.7, G6=4.0, G5=4.0, G1=4.5. Bottom row (orange) encodes interrogator success rate (%) with values: G7=33.3, G9=33.3, G3=66.7, G8=33.3, G2=66.7, G10=66.7, G4=66.7, G6=100, G5=100, G1=0. The x-axis labels groups by ID; the two-row layout places TG directly above the corresponding success rate for each group. The figure shows that higher tie strength at or above 4.0 is associated with extreme outcomes—either 100% success (G6, G5) or 0% (G1).
Figure 5. The distribution of signals used by participants across groups, sorted left to right from lowest to highest TG . Each cell gives that signal’s share of the group’s coded signal mentions, so columns sum to 100%; darker shades indicate higher shares. Because per-group totals differ (7 to 25 mentions), shares are not comparable across columns as raw counts.
Platform
Conversations
Messages
Facebook Messenger
30.58%
64.51%
Apple iMessage
28.25%
17.82%
Tinder
26.57%
1.20%
Instagram
13.05%
11.21%
Hinge
1.10%
0.06%
WhatsApp
0.45%
5.19%
Table 1. Platform composition of the historical chat record. Shares are of the 4,637 conversations and of the 434,085 messages. Table of platform composition for the historical chat record, with columns: share of conversations, share of messages. Facebook Messenger 30.58 percent and 64.51 percent; Apple iMessage 28.25 and 17.82; Tinder 26.57 and 1.20; Instagram 13.05 and 11.21; Hinge 1.10 and 0.06; WhatsApp 0.45 and 5.19.
Metric
Mean
SD
Q1
Median
Q3
Max
Duration ( days )
227.7
502.7
0.4
5.1
159.1
3,454
Messages ( count )
93.6
504.4
2.0
5.0
20.0
10,004
Table 2. Summary statistics per conversation for the historical chat record. We compute duration over the N=3,485 conversations containing at least two messages and a nonzero span, and message counts over all 4,637. Two-row table of summary statistics from the historical chat record, with columns: Mean, Standard Deviation, Q1, Median, Q3, Maximum. Row 1 — Duration (days): mean 227.7; SD 502.7; Q1 0.4; median 5.1; Q3 159.1; max 3,454. Row 2 — Messages (count): mean 93.6; SD 504.4; Q1 2.0; median 5.0; Q3 20.0; max 10,004. The two rows use different denominators: duration over the 3,485 conversations with at least two messages and a nonzero span, message counts over all 4,637.
Figure 6. The system architecture of the AI Impersonator. Historical messages are preprocessed and indexed for retrieval in a vector database (left). A dynamically constructed prompt combines static identity signals with relevant messages from the history and the current PIG chat context (right-top). The LLM updates the chat in real time (right-bottom). System architecture diagram in three panels (left, center, right) illustrating the AI Impersonator pipeline from data ingestion to live chat response. Left panel (Data ingestion and normalization): personal messages are pulled from multiple platforms (WhatsApp, Hinge, Tinder, Instagram, iMessage, Facebook). An arrow leads into a Pre-processing node. The output is a stack of Normalized Messages, each record containing a timestamp, sender name, and message text. Third-party senders are anonymized (e.g., Person_1). An arrow leads from Normalized Messages to the Vertex AI platform, where messages are embedded by an embedding model and stored in a vector database for retrieval. Center panel (Prompt assembly for the AI Impersonator): the prompt includes (1) instructions to imitate a specific person in conversation; (2) three identity signals for that person: Personality scores, Stylometric profile, and a short written autobiography; (3) the current PIG chat transcript; and (4) relevant historical messages retrieved from Vertex AI via vector search. This assembled prompt is sent to the Gemini LLM. Right panel (Game live chat interface and response loop): the chat UI displays the round title, a role reminder, and a countdown timer. Messages appear as labeled bubbles. When a new user message is sent, the AI Impersonator is triggered; Gemini generates a reply that is returned to the chat, completing the loop.
Rounds ( N )
Target
Impersonator
ISR
AI win
Conf.
Human Baseline
All (30)
Human
Human
56.7%
—
79.8%
AI Study
1 (10)
AI
Human
70.0%
70.0%
75.0%
2 (10)
Human
Human
50.0%
—
50.0% †
3 (20)
Author
AI
65.0%
35.0%
76.7%
Table 3. Interrogator performance across the four PIG rounds. In the baseline and Round 3 a high ISR means the impersonator failed, whereas in Round 1 the AI holds the target seat, so a high ISR means the AI out-impersonated the human friend. The AI’s win rate is therefore given separately. Conf. is interrogators’ average confidence. No rate in this table is distinguishable from chance (exact binomial p≥0.26 ). † The target author was the interrogator in all ten Round 2 games. Six-column table reporting interrogator performance across four PIG rounds. Columns: Rounds; Target; Impersonator; ISR, the interrogator success rate (%); AI win rate (%); Interrogator confidence (%). Rows are grouped under two sub-headings, Human Baseline and AI Study. Row 1 — Human baseline study, all 30 rounds; Target: Human; Impersonator: Human; ISR: 56.7%; AI win rate: not applicable; Interrogator confidence: 79.8%. Row 2 — AI Impersonator study, Round 1; Target: AI; Impersonator: Human; ISR: 70%; AI win rate: 70%; Interrogator confidence: 75%. Row 3 — AI Impersonator study, Round 2; Target: Human; Impersonator: Human; ISR: 50%; AI win rate: not applicable; Interrogator confidence: 50.0%, marked with a dagger because that interrogator was the target author rather than a participant. Row 4 — AI Impersonator study, Round 3; Target: Author (human); Impersonator: AI; ISR: 65%; AI win rate: 35%; Interrogator confidence: 76.7%. No value is bolded, because the columns measure different constructs and are not comparable as a single ranking.
Large language model-based role-play simulation creates interactive agents that portray people, characters, or users for training, testing, and research.
1AILab, MIGe, University of Trieste, 34127 Trieste, IT · 2Computational Statistics and Machine Learning, Istituto Italiano di Tecnologia, Genova, IT · 3DIA, University of Trieste, 34127 Trieste, IT