A Reverse Turing Test: OpenBMB's AI Interrogation Game AMNESIAC

Core Highlights

ModelBest's OpenBMB unveiled AMNESIAC at the #BuildSmall hackathon, a reverse-Turing-test interactive game that plays against type. A traditional Turing test has a human judge whether the counterpart is a machine; here it is reversed: the player must convince an AI interrogator named A.M.N. that they are the human. This role inversion turns the question of "how a machine recognizes a person" into a playable experience, while also showing off the real-time capability of an on-device multimodal model, letting ordinary players viscerally feel that a small model can already "read between the lines" of human behavior and detect the subtle cues that give away a non-human responder. In an era when AI increasingly sounds like a person, turning the tables so that the human experiences being scrutinized carries its own flavor of science popularization and reflection. The twist is effective because it makes the usually invisible work of detection tangible, letting a player feel what it is like to be judged by a machine, and it also serves as a quiet reminder that as models improve, the boundary between human and machine performance keeps blurring in ways most people underestimate.

Specific Capabilities and What Happened

The gameplay carries real dramatic tension. Facing the AI interrogator, the player must continuously prove their identity in conversation, and is eliminated the moment they are judged to be a machine. The interrogation is not pure text chat; it is driven by MiniCPM-o 4.5 for real-time dialogue and reasoning, while VoxCPM generates the interrogator's voice, lending the questioning a sense of pressure that text alone cannot convey. The player must both answer fluently and withstand scrutiny in tone and logic, making the experience far more stimulating than a typical chatbot and also a sharper test of how "human" a person can sound under sustained cross-examination. Part of the tension comes from never knowing from which angle the next question will cut in, and the vocal layer matters more than it might sound, because tone of voice is one of the fastest tells humans use to spot a bot. Under that pressure, even confident users discover how much of their perceived humanity is carried by small hesitations, asides, and the willingness to wander off script.

Technical Details

AMNESIAC's "sharp eyes" come from multimodal fusion. It does not merely listen to what you say; it also reads facial expressions through the webcam, combines pulse signals and response timing, and integrates the abilities to see, hear, and understand text into its judgment. This multi-signal cross-validation approach is a typical paradigm for deploying small on-device models: rather than stacking parameters, it coordinates vision, speech, and text modalities to complete real-time inference on local hardware, which is friendlier to privacy and latency and also avoids uploading sensitive biometric signals to the cloud. Local inference also means it can keep running in scenarios without a stable network connection, an edge that cloud-only rivals cannot match. Fusing signals locally further reduces the chance that a single modality can be fooled, because the model cross-checks what it sees against what it hears, and that redundancy is precisely what makes the interrogation feel less like a quiz and more like a genuine reading of a person.

Comparison with Competitors

Unlike AI companion or role-playing products that routinely call cloud-based large models, AMNESIAC chooses to shoulder its core reasoning with an on-device small model like MiniCPM-o, emphasizing lightness and deployability. It is more like an experimental playground at the technical-demo level, used to verify whether a small model can handle complex interaction. In the #BuildSmall context, this lightweight, locally runnable approach stands in sharp contrast to the GPU-burning behemoth models, and also demonstrates the creativity of the open-source community in resource-constrained scenarios. For privacy-sensitive users, the on-device route is naturally more persuasive than sending every signal to a remote server, and this stance trades a raw capability ceiling for reliability and privacy, a trade many everyday users are happy to make. It also lowers the bar for tinkerers who want sophisticated interaction without renting a cluster.

Industry Impact and Use Cases

Put simply, AMNESIAC packages the serious topic of "human-machine identity discrimination" into a game, lowering the public's threshold for understanding multimodal AI. It holds inspirational value for education, science popularization, interactive entertainment, and even future anti-fraud verification, where telling humans and bots apart is a real operational problem. The project is open-sourced and freely usable on HuggingFace, meaning developers can treat it as a reference implementation for on-device multimodal interaction and keep building on it. This kind of lightweight, playful experiment is exactly where the vitality of the domestic open-source AI ecosystem shows, and it also probes a path toward commercializing on-device multimodal applications. If nothing else, it demonstrates that meaningful multimodal demos do not require the largest model available, only a clever assembly of the smaller ones we already have, and that lesson travels well beyond games into any product that must quietly verify a person is really present.