Video AI models plateau at 50% of human reasoning despite perfect pixels

The largest reasoning study to date finds leading video generators fail basic physics tests while producing photorealistic scenes

Watch Sora 2 generate a basketball bouncing upward after being dropped. Or Veo 3.1 rendering water flowing uphill in a kitchen sink. These systematic patterns emerged from testing 32 institutions' worth of video AI models against the new Very Big Video Reasoning benchmark, released today.

The VBVR suite represents what happens when researchers stop measuring visual fidelity and start measuring comprehension. Despite the industry pouring billions into scaling compute and training data, flagship models achieve only half of human performance on basic physical logic tasks, according to The Decoder. The finding suggests video AI has hit what researchers are calling a reasoning ceiling, a point where adding more data improves familiar scenarios but fails to generalize to new physical situations.

The study found that models excel at reproducing training-adjacent content but struggle with novel spatial relationships, producing perfect pixels with broken physics. The consortium tested everything from object permanence (does the ball still exist when occluded?) to causal chains (if you push domino A, what happens to domino C?).

The timing matters. OpenAI just launched Codex Security for enterprise vulnerability detection. Anthropic opened a marketplace for third-party tools. Yet their video models can't reliably predict which way a door swings open.

According to AI Tech Suite, the research identifies controllability as the core bottleneck. Current architectures lack state tracking and self-correction mechanisms that would allow them to maintain physical consistency across frames. The models regenerate each moment without understanding how the previous moment constrains what comes next.

This echoes the moment when image generators could produce stunning portraits but gave everyone six fingers. The stakes are different now. Vietnam mandated AI content labeling starting March 1. India requires platforms to remove flagged manipulated videos within three hours. The UK's House of Lords just urged adopting a licensing-first regime with C2PA watermarking standards, according to Computer Weekly. The US Supreme Court declined to hear an appeal on AI-generated art copyright, leaving in place rulings that require human authorship for protection, MarketingProfs reports.

The industry's response has been telling. Neither OpenAI nor Google provided comment on the VBVR findings. Their recent product launches suggest a pivot away from pure generation toward hybrid systems that combine AI with traditional simulation engines.

The benchmark itself deserves scrutiny. Testing reasoning assumes these systems reason at all, rather than performing sophisticated pattern matching. The 50% performance figure comes from averaging across different tasks. Models excel at some, like object tracking, while failing completely at others, like multi-step causality.

The findings puncture the narrative that video AI just needs one more breakthrough to achieve human-level comprehension. The researchers tested scaling strategies explicitly: doubling parameters, tripling training data, extending context windows. Performance increased logarithmically at best, plateauing well below human baselines.

Video creators should expect continued improvements in aesthetic quality while physical accuracy remains limited. Platforms implementing AI detection will need to look beyond visual artifacts to physics violations. Copyright frameworks assuming human-like creativity from AI systems require fundamental rethinking. Training on synthetic data from flawed models risks compounding reasoning errors. Hybrid approaches combining generation with physics engines may become the industry standard.

The consortium plans quarterly VBVR updates to track whether architectural shifts can break through the ceiling. They're particularly interested in whether models trained on interactive settings, where physics violations have consequences, perform differently than those trained on passive video.

If perfect photorealism without physical understanding becomes the norm, how do we handle a media environment where seeing doesn't guarantee believing, and reasoning about what we see becomes equally unreliable?