What AI Video Generation Can and Cannot Do Right Now

Demonstrations of generated video are selected. They show the shot that worked, not the nineteen before it, and they are chosen from the categories where the technology is strongest. That is not dishonest — it is what a demonstration is — but it makes the technology look uniformly capable when it is sharply uneven.

The question worth answering is not whether generated video can look good. It can, and sometimes remarkably so. It is which shot types come out usable often enough to plan around, because a production schedule cannot be built on a tool that succeeds unpredictably. This guide maps that unevenness, names the failure modes that recur, and sets out where generation currently earns its place.

Reliability, Not Peak Quality

The single most useful reframing is to stop asking how good the output can be and start asking how often it is usable.

A shot type that works nine times in ten is a production tool. You can put it in a schedule, quote for it, and rely on it. A shot type that works once in twenty is a lottery ticket — it may produce something extraordinary, and you cannot promise anyone it will arrive by Thursday.

This distinction explains most of the disappointment people report. They saw an impressive example, assumed the category was solved, and discovered that the example was the successful outcome of a process they did not see.

The pattern in that table is consistent. Generation is strong where nothing has to stay the same and weak where something does.

Why the Failures Cluster

Four failure modes account for most of what goes wrong, and they share a cause.

Hands and fingers. A hand has a fixed structure with strict constraints. The model produces something that looks like hands generally look, and "generally" includes a distribution of finger counts. It is not making an error so much as sampling from a space where the rule is not enforced.

Text and logos. Letterforms have to be exactly right to read as writing. Approximately-correct letters are immediately wrong to a viewer, in a way approximately-correct foliage is not. This is why generated signage and packaging so often look convincing until you read them.

Physical contact. One object touching, holding, or supporting another requires consistent physics across frames. Fingers pass through handles, objects rest on nothing, and the moment of contact is where it usually breaks.

Continuity. Nothing in the process holds a character's face constant between two separate generations. Each shot is produced independently, so consistency has to be imposed from outside — what can and cannot be carried between shots is set out in the continuity techniques and their limits.

Everything on that list is a case where being statistically plausible is not enough, because a viewer is checking against a rule they know.

Judge It Against the Alternative, Not Against Perfection

A comparison people get wrong: generated footage is usually assessed against filmed footage, when the real alternative is often something else entirely.

If the choice is between a generated establishing shot and a filmed one, filmed wins on almost every axis. But that is frequently not the choice. The actual alternative is generic stock, or a static graphic, or simply not having the shot — and against those, a serviceable generated clip is a clear improvement.

This is why the same output can be a success or a failure depending on the project. A generated cityscape replacing a location shoot is a downgrade. The same clip replacing a stock library search that was going to take an hour and cost money is a straightforward win.

The practical form of this is to decide what a shot is competing with before evaluating it. Shots on your load-bearing list are competing with filmed footage and should be held to that standard. Shots on the decorative list are competing with nothing, and the bar is whether they work, not whether they are indistinguishable from a camera.

There is a second-order effect worth naming. Because generated shots are cheap, they tend to accumulate — a video ends up with fourteen atmospheric inserts because each one cost almost nothing. Cheapness is not a reason to include a shot, and the same test applies as to any other cutaway: it earns its place by doing something the surrounding footage was not doing.

Where It Genuinely Earns Its Place

Set against those limits, several uses are straightforwardly good.

Illustration for abstract points. Narration about growth, complexity, connection or time has always needed imagery that stands for an idea rather than depicting a fact. Generated footage is well suited to this, and the alternative was generic stock.

Texture and atmosphere. Establishing mood, filling transitions, providing motion behind titles. High reliability, low stakes, and no continuity requirement.

Impossible or expensive shots. Aerial views, historical settings, scale changes, anything a shoot could not reach. Here the comparison is not against filmed footage but against not having the shot at all.

Concept and previsualisation. Showing a client or a team what a sequence might look like before committing to a shoot. Failures cost nothing at this stage, which changes the reliability calculation entirely.

What these have in common is that no viewer is being asked to believe a specific thing happened. Where the content does make that claim, disclosure becomes the relevant question rather than quality — the test for it is in disclosing AI-generated media to your audience.

The Capability Moves; the Shape Does Not

Any specific claim about what these tools can do has a short life. Models improve, and shots that were unreachable last year are routine now.

What has been stable is the shape of the difficulty. The hard cases have consistently been the ones requiring a rule to be satisfied exactly — correct anatomy, legible text, physical contact, the same face twice — and the easy cases have consistently been the ones where plausible is sufficient. Improvements have moved the line without changing which side things sit on.

That gives you a way to evaluate a new tool that does not depend on anyone's claims. Test it on the hard categories rather than the easy ones. Everything generates convincing clouds; the question is whether this one can produce a hand holding an object, twice, recognisably.

It also means the workflow advice below outlasts the capability figures above. Planning which shots are load-bearing, starting from reference images, keeping clips short and generating batches rather than iterating — none of that becomes unnecessary as models improve. It just applies to a larger set of shots.

Working Within the Limits

The last point is worth emphasising because it changes what you can use. A clip that falls apart after two seconds is still a usable two-second insert, and short inserts are most of what b-roll consists of anyway.

Next Steps

Take a video you are planning and mark every shot as load-bearing or not. Then look only at the second list — that is the honest scope for generation on this project, and it is usually smaller than people expect and larger than nothing.

Generate three shots from that list and count the attempts each took. That number, on your own material with your own tools, is more useful than any general claim about capability. Before you start, it is worth having the sequence planned rather than discovered, which is the case made in storyboarding before you generate anything. Everything else in this area is collected in the generation cluster.