A video AI can recognise a person, a chair and a cup in every frame and still fail to tell whether the person moved faster, farther or in the opposite direction. Object recognition answers what is visible. Motion understanding requires comparing change across time—and a new benchmark called MotionBlind suggests that many video-language models remain surprisingly weak at that basic task.

A fluent description can hide temporal blindness

When a model watches a clip, it usually does not process continuous time as people experience it. It samples frames, converts them into visual tokens and reasons over that compressed sequence. This is efficient, but it can preserve appearance better than movement. If the first and last frames look similar, the system may rely on familiar scene descriptions instead of measuring what happened between them.

That produces a convincing illusion. A model says, “A person lifts a cup,” so it appears to understand the action. Ask whether one lift was faster than another, whether the hand travelled farther or whether the direction reversed, and the answer may collapse.

What MotionBlind changes

Released on 8 September 2026, MotionBlind uses pairs of near-identical self-recorded videos. The room, person and objects stay the same; speed, magnitude or direction changes. Each clip receives two complementary yes-or-no questions, so each pair produces four answers. A model receives credit only if all four are consistent. Random guessing therefore has a 6.25% chance of passing an instance.

The researchers tested six open and two frontier video-language models while varying frame count, temporal order and frame-selection strategy. According to the preprint, open models remained near the chance floor. Removing video reduced all models to zero instance accuracy, and shuffling frames reduced performance to chance, confirming that the questions genuinely require ordered video. More frames and smarter sampling did not solve the problem. Only one tested frontier system cleared the benchmark overall, and it still failed on speed.

These results are author-reported and the benchmark is new. They do not prove that every video model is motion-blind in every setting. They do show why scene fluency should not be used as evidence of physically grounded motion understanding.

A concrete example

Imagine two clips of the same reach. In both, a hand begins near the body and ends at a cup. In one, the hand accelerates quickly and brakes sharply. In the other, it travels slowly and continuously. A model focused on endpoints can label both clips correctly while missing the difference a movement teacher considers central.

Adding more sampled frames may help only if the system uses them relationally. Twenty-four still images are not automatically a velocity measurement. Speed requires change in position divided by change in time, while magnitude requires a stable reference and direction requires ordered comparison. The model needs a representation and training objective that preserve those relations.

Why this matters beyond benchmarks

Video-language models are increasingly proposed as evaluators, reward signals and perceptual front ends for robots and world models. If such a model cannot distinguish two speeds of the same action, it may reward a generated movement for looking plausible while missing the physical property the instruction requested.

The problem is especially serious for somatic and creative practice. “Move more slowly” is not a change in object identity. “Reach with less amplitude” cannot be verified by naming the action. A system must compare temporal organisation, not simply recognise the category.

The Lab’s earlier article on what gets lost between sampled frames explains the input bottleneck. Its discussion of Motion-as-Prompt describes one attempt to draw trajectories directly onto visual inputs. MotionBlind adds the missing audit: test whether the model’s answer actually changes when only the movement changes.

What a better test looks like

A credible motion evaluator should pass contrastive tests before judging complex performances. Keep the person, setting and action constant; vary one physical property; reverse or shuffle time; remove the video; and check whether answers remain logically consistent. These controls reveal whether the model is measuring motion or guessing from language and appearance.

Three shortcuts that can look like understanding

The first shortcut is appearance prior. If most training videos labelled “running” contain sports clothing or outdoor tracks, a model can predict the action from the setting before reading the legs. A controlled pair removes that shortcut by keeping clothing and background constant.

The second is language prior. Questions often contain clues about the likely answer. If “Did the person speed up?” is usually answered yes in a benchmark, a language model can score well without using the video. No-video controls reveal that behaviour.

The third is endpoint matching. A model may compare the beginning and end while ignoring the path between them. That can distinguish a large reach from a small one, but it cannot reliably separate a smooth deceleration from an abrupt stop or identify a reversal that returns to the same endpoint. Reversing and shuffling frames tests whether temporal order matters.

These shortcuts are not evidence of deception. Machine-learning systems optimise the easiest reliable pattern in their data. The benchmark designer’s job is to make the intended physical relation the only dependable route to a correct answer.

For practitioners, the lesson is simple: never accept a fluent caption as proof that an AI understood the movement. Ask it to discriminate carefully controlled differences and compare its judgement with measured trajectories and human observation. Motion understanding begins where object naming stops.

References

Bhatia, D., Galoaa, B., Fritsche, O., Kamal, S., Salam, M. O. A., Saleem, U., Rastogi, O., Chettiar, F. F., Erdogmus, N., & Ostadabbas, S. (2026). MotionBlind: Probing the illusion of motion understanding in Video-LLMs. arXiv:2609.09528. https://arxiv.org/abs/2609.09528

Sun, X., Liu, K., Wang, H., Liu, L., Wang, X., & Liu, Y. (2026). Motion-as-Prompt: Enhancing motion reasoning in multimodal large language models via motion-guided cross-frame visual prompting. arXiv:2608.11655. https://arxiv.org/abs/2608.11655