hoshen/Beyond-the-Turing-Test

Social Turing Test: 50 study games

Data for Beyond the Turing Test: Evaluating AI Agents in Group Conversation, Hoshen Lugasi, Uri Berger, Niv Eckhaus, Itay Itzhak and Gabriel Stanovsky (2026).

Code | Project website

Humans chat and vote to find one AI player. Humans win by eliminating the AI; the AI wins by reaching the final two. This is data and frozen calculation inputs, not the game server.

Browse and load

Each game has seven files: metadata, transcript, ballots, surveys, timeline JSONL, full text rendering and coverage manifest. JSONL is event authority; text is its deterministic full rendering. No HTML viewer or credentials are needed.

Five configs use only split data: games (default, 50 rows), messages (2,954), votes (618), surveys (229), events (53,360). No train/test split is invented. The size category describes the default game-level view. Events load nested timelines directly, not a root event table; the other four views are under tables. See dictionary and dataset manifest.

After downloading, these optional datasets examples use this package's root as the working directory, without assuming a hosted URL:

from datasets import load_dataset
games = load_dataset(".", name="games", split="data")
messages = load_dataset(".", name="messages", split="data")
events = load_dataset(".", name="events", split="data")
social = events.filter(lambda row: row["component"] == "Social State Tracker")
players = messages.filter(lambda row: row["speaker_role"] != "system")

dataset_infos.json declares explicit nullable nested types, so fields that are all null in the first event shard do not become null-only columns. The same configuration names apply to the published Hub ID hoshen/Beyond-the-Turing-Test:

games = load_dataset("hoshen/Beyond-the-Turing-Test", name="games", split="data")
events = load_dataset(
    "hoshen/Beyond-the-Turing-Test", name="events", split="data", streaming=True
)

Standard-library reading needs no installation:

import json
from pathlib import Path
with Path("games/game_001/events/timeline.jsonl").open(encoding="utf-8") as handle:
    events = [json.loads(line) for line in handle if line.strip()]

Optional pandas loads JSONL with pd.read_json(..., lines=True) after importing pandas. Neither pandas nor datasets is required for companion offline calculations. Local examples do not imply remote Hub viewer certification.

Study and timeline scope

There are 50 games, 135 rounds, 2,414 player messages (1,895 human / 519 AI), 540 system notices, 618 cast ballots (483 human / 135 AI) and 229 surveys. Enjoyment has 224 ratings; five terminal-format surveys did not ask it. Missing values are not zeros. AI ballots affect elimination but are excluded from human-only identification. Demographics are aggregate-only for 226 records.

Sixteen games started with four humans and 34 with five. Inclusion was accepted before outcomes were inspected; nobody was removed after START in those sessions. Use roster-adjusted references and four-/five-human sensitivity results. The 234 starting human seats do not establish distinct people across games. Alternative model or test games do not belong to this historical Flash condition.

Scheduler, Generator and Reviewer used Gemini 2.5 Flash (gemini-2.5-flash), four candidates and verbatim reviewer selection, not rewriting. Recorded study diagnostic evaluation used Gemini 2.5 Pro (gemini-2.5-pro) with feedback disabled; its 1,558 records are excluded from released timelines. Of 54,918 source events, 53,360 are retained.

Timelines include Internet Context, Round Summarizer, Persona Monitor, Social State Tracker, Behavior Adapter and Repetition Monitor, alongside foreground and System records. These are reviewed structured outputs, not full prompts, snapshots or raw provider responses. Selection is not delivery; available context is not proof of use/influence. Returned-item counts are not independent cited-source counts. Persona/social observations are model interpretations, not verified traits.

Historical timing is one-second precision since the first source event, not START. No chat-clock merge, timezone correction, inferred date or causal concurrency order is asserted. See dataset guide.

Numerical reproduction

The companion code supplies four standard-library scripts and the same data. Use it as-is or substitute these games and inputs in its corresponding folders. Scripts print JSON without API calls or figures; four core files alone cannot recover every calculation input. Preserve cohort order, audit boundaries, original numeric word counts and frozen judgments/ annotations. Redacted text cannot recreate original model decisions. The original midnight-window behavior is retained; event timing is not a correction. Later-stage detection conditions on AI survival. See input provenance.

Privacy, rights and validation

public_redactions.json records four additional spans across three files relative to the submitted core (197 of 200 files unchanged), with no changed ratings, ballots or outcomes. A first pass reviewed 7,408 unique text values; 52 were withheld at 135 payload occurrences. A further 22 withheld values occur at 111 payload locations. Repeated rendering is not counted again.

Contextual checks were selective, not exhaustive; review is not anonymity proof. Do not re-identify or externally link participants. Author permission is conditional on not identifying individuals, not independent consent/IRB certification. MIT within the distributor's rights does not override privacy or third-party rights. Cite the title, authors and year above; machine-readable citation metadata is available with the code. No public arXiv identifier or paper DOI is available yet.

Release notes describe validation boundaries. The final outside-ZIP validation report identifies the exact artifacts checked, including 12 offline scientific reproductions, tests in a new external Windows virtual environment, server checks and local dataset loading. Its evidence does not automatically certify edited files or rebuilt packages. No new model-backed game, scientific-quality validation or remote Hub viewer validation is claimed here.