Why Motion, Physics, and Hands Go Wrong

A glass falls off a table and bounces like rubber. A person's legs swap places mid step. A hand grows an extra finger while it picks up a cup. AI video can look real for a second, then do something no real camera would ever record.

These mistakes are not random. They come from how video models learn. This lesson explains the main reasons, so you know what to expect and where to look.

What You'll Learn

• Why a video model knows how things look, not why they move
• Which kinds of physics are hardest for AI video
• Why hands and bodies are such a common problem
• Why cause and effect often breaks
• Why some motion looks fine and other motion falls apart

Patterns, not physics

A game engine or a physics simulator follows rules. It knows gravity pulls things down, solid objects do not pass through each other, and water flows downhill. It calculates each frame from those rules.

A video model has no rules like that. As you saw in lesson 1, it learns from a huge number of clips. It learns what motion usually looks like. It does not learn why things move.

This works well when the motion is common and simple. Waves rolling in, clouds drifting, a car driving down a road. The model has seen thousands of examples, and they all look alike.

It works badly when the motion is rare, complex, or depends on exact cause and effect.

The hardest kinds of physics

Some kinds of motion trip up AI video again and again:

• Collisions. When two things hit each other, the result depends on weight, speed, and material. The model often blends the objects together or lets one pass through the other.
• Liquids. Pouring, splashing, and stirring involve many tiny parts moving at once. Water may pour upward, or a glass may fill without water reaching it.
• Breaking and cutting. A cut cake should stay cut. A broken plate should stay broken. The model may undo the change a few frames later.
• Gravity and weight. Objects may fall too slowly, float, or land without any bounce or impact.
• Counting. Juggling balls, stairs, or fingers on a hand. The number often changes, because the model tracks the look, not the count.

Why hands and bodies fail

Hands were a well known problem in AI images. In video they are harder still. A few reasons:

• Many small parts. A hand has many joints that bend in many ways. Small errors are easy to make and easy to notice.
• Often hidden. In real video, fingers hide behind each other, behind objects, and behind the body. The model sees many partial hands, so it has fewer clear examples of how a full hand moves.
• Fast and blurry. Hands move quickly. In real footage they are often blurred, which gives the model unclear examples.
• Easy to spot. People look at hands and faces closely, so even small mistakes stand out.

Full bodies have similar issues. Walking looks simple, but the legs and arms have to swap in the right order. Models sometimes let a leg slide through the other, make feet slide across the floor, or turn a person's body in a way that is not possible. Dancing, sports, and fighting are even harder, because the movement is fast and the limbs cross over each other.

When cause and effect breaks

Real events have a clear order. You bite a sandwich, and the bite mark stays. You open a door, and then you walk through it. You knock over a cup, and the drink spills.

A video model has to fake this order from patterns. Common failures:

• The bite mark appears, then disappears.
• A person walks through a door that has not opened.
• A cup tips over, but the drink stays inside.
• A ball hits a wall, and the wall moves instead of the ball.

These mistakes happen because the model cleans up the whole clip at once and aims for frames that each look believable. It does not check that frame 50 follows from frame 20 for a real reason. Each frame can look fine on its own while the story they tell together makes no sense.

Why some motion looks fine

You might wonder why some AI clips look almost perfect. The answer is that they play to the model's strengths:

• Common scenes. Landscapes, city streets, slow camera moves.
• Slow, smooth motion. Less change between frames means less to get wrong.
• Few interacting objects. One person walking is easier than two people shaking hands.
• Short clips. Fewer frames means fewer chances for errors to show.

Put together, scenes fall on a rough scale:

• Easier: slow, common, one subject
• Harder: fast, many moving parts
• Hardest: contact, liquids, hands, cause and effect

When you see a clip that looks flawless, it often uses a scene from the easy end of this scale. Newer models move the line further each year, but the order of difficulty stays about the same.

Does more data fix this?

More and better training data does help. Newer models show fewer of these errors than older ones, and some companies talk about their models learning more about how the world works. But the core method has not changed. The model learns from what videos look like, so it is still weakest where good examples are rare or the motion is complex.

That is why you should not treat any of these mistakes as a fixed rule. A tell that works today may fade in a year. What stays useful is knowing where to look: contact, hands, liquids, counts, and cause and effect.

Key Takeaways

• Video models learn what motion looks like, not the physics behind it.
• They do well with common, slow, simple motion and badly with rare, fast, complex motion.
• Collisions, liquids, breaking, gravity, and counting are frequent failure points.
• Hands and bodies fail because they have many parts, are often hidden, move fast, and are easy for people to check.
• Cause and effect breaks because the model makes each frame look right, not the story that links them.
• Errors are shrinking with better models, but the places to look stay the same.