how do you spot a Sora video?

Sora is OpenAI's text-to-video model. announced in February 2024 with a research demo that briefly broke the consumer-AI news cycle, publicly released as “Sora Turbo” in December 2024, and upgraded with Sora 2 in September 2025. the architecture is a diffusion transformer that treats video as patchified spatiotemporal tokens, analogous to the way large language models treat text.

Sora 1 outputs were striking and obvious. Sora 2 improved physics fidelity more than any other video AI model shipped in 2025, so the tells are subtler.

what a Sora video looks like

the canonical Sora 1 failure modes, most of which still occur in Sora 2 at lower frequency:

morphing background objects. items in the background of a shot subtly change shape, multiply, or disappear across frames (a coffee cup on a desk in second one becomes a bottle in second four), making this the most reliable Sora tell on long-enough clips.

physics violations in liquids and hair. water, smoke, fire, and hair behave inconsistently with momentum and gravity. Sora 2 explicitly targeted this and OpenAI's positioning was “prior video models morph objects and deform reality to satisfy the prompt.” the residual cases in Sora 2 are subtler and harder to catch by eye.

clothing motion decoupled from body motion. a jacket sways at a different beat than the body underneath it. a scarf flutters in a direction inconsistent with the wind blowing the subject's hair. this remains a reliable Sora-family signal.

the “Sora cat”. extra limbs, doubled joints, hands with too many fingers appearing mid-frame in otherwise-stable shots. the meme name comes from a particularly gnarly Sora 1 cat clip. Sora 2 reduced these dramatically but not to zero.

warped text and reflections. signage and small printed text gets unreadable; reflections in mirrors and water don't track the source.

how amige. detects Sora

amige. doesn't run one test. a trained model reads the marks a generator leaves and routes the scan to the detectors strongest for video, then a panel of independent detectors built by different teams weighs in and the reads get fused and calibrated. for Sora it leans on three layered signals:

C2PA Content Credentials. every Sora output ships with a cryptographically-signed provenance manifest and a visible watermark in the lower corner. when this metadata is intact, the call is trivial: amige. reads the C2PA chain and reports Sora origin with very high confidence. the problem is that almost every social media platform strips C2PA metadata during re-encoding, so a Sora clip arriving from TikTok or Twitter usually has no provenance left to read. absence of C2PA indicates re-upload, not an authentic source.

temporal coherence and optical flow. diffusion- transformer video leaves measurable artifacts: subtle flicker between frames, warping in regions of high motion, AI-typical motion blur where the blur direction doesn't match the subject's actual motion vector. optical-flow residuals (second-order motion approximations) detect Sora-family output even on detectors that have never seen Sora training data (per arxiv 2406.19568, “What Matters in Detecting AI-Generated Videos like Sora?”).

frame-level AI scoring. the video classifiers in amige.'s panel score AI probability per frame rather than averaging across the clip, so a fully-synthesized Sora scene scores consistently high across every frame in a way that a real video with a single deepfaked element does not. this distinction shows up in the per-detector breakdown.

the detection difficulty curve

Sora 1 outputs are reliably caught by temporal-consistency methods. Sora 2 made detection harder by improving physics fidelity, but added a new signal in synchronized audio: AI- synthesized audio and video that don't quite agree in timing or environmental acoustics expose their own forensic surface. public peer-reviewed Sora-2-specific benchmarks are thin as of mid-2026, and the wider detector ecosystem is still adapting to the synchronized-audio surface.

for an amige. user: a Sora flag with intact C2PA metadata is near-certain. a Sora flag without C2PA, on a clip that shows morphing background objects or hair-physics anomalies, is strong evidence. a confident Sora flag on Sora 2 output without the obvious tells is meaningful but read the per-detector agreement signal before treating it as a final answer.

the likeness controversy

Sora 2's launch in late 2025 was followed within weeks by a nonconsensual-likeness scandal. Bryan Cranston discovered Sora- generated videos of himself including a “selfie with Michael Jackson” clip; SAG-AFTRA filed; OpenAI shifted likeness use to opt-in only on October 20, 2025. Zelda Williams publicly asked users to stop sending her AI-generated Robin Williams videos. the Martin Luther King Jr. estate forced a Sora block on October 17, 2025 after viral clips depicted him shoplifting and being chased by police, exposing the post-mortem-likeness gap that reporters called the “dead celebrity loophole.” in November 2025 the consumer advocacy group Public Citizen formally called on OpenAI to pull Sora 2 from public release.

these controversies are why the detection question matters. for more on the methodology, see the machine or read about what counts as a deepfake.