Transparency
AI methodology
A technical, non-marketing account of the identification pipeline, including where it fails.
The model
Identification uses a general-purpose multimodal large language model with vision capability, accessed through a hosted API. It was not trained by us and is not fine-tuned on a proprietary film corpus; its film knowledge comes from broad pre-training on public data.
The same model family handles text queries (quotes, plot descriptions, character and actor descriptions) and speech-to-text transcription for video links. We change the underlying model when a better one becomes available, and this page is updated when we do.
What the model actually reasons about
Faces and apparent casting; wardrobe and period detail; set design, architecture and location; lens choice, aspect ratio, film grain and colour grade; subtitle and caption typography; visible interface elements from streaming apps; and any legible on-screen text.
It does not perform biometric face matching, and it does not compare your image against an index of known frames. Any actor name it returns is an inference from appearance and context, not an identity verification.
How confidence is expressed
The model reports a self-assessed confidence alongside its answer, and the interface bands it into high, medium and low. Self-reported confidence is imperfect: it correlates with correctness but does not guarantee it.
Below the high band the result presents ranked alternatives rather than a single answer. We would rather show three plausible titles than one wrong one that looks authoritative.
Known failure modes
Very dark, heavily motion-blurred, extremely low-resolution or tightly cropped frames often carry too little signal. Generic modern interiors and stock-looking landscapes are frequently unsolvable from a single frame.
Films with strong visual similarity — same director, same period, same franchise, remakes and their originals — are the most common source of confident errors. Very recent releases may post-date the model's knowledge, and obscure regional titles are under-represented in its training data.
Enrichment data (ratings, cast, streaming availability) is generated alongside the identification and can contain errors even when the title is right. Verify anything that matters before relying on it.
Human oversight and corrections
No answer is manually reviewed before you see it — the pipeline is fully automated. When users report a systematic error, we adjust the prompt or the model and note material changes here. Report errors to hello@whatmovieisthis.com.