Taste is the last creative advantage in AI

What changed when making got cheap

For most of the history of this industry, the hard part was execution. You had an idea, and then you needed a crew, a camera, a grade, and six weeks. The idea was cheap. The making was expensive.

That ratio has flipped. Generation is now the cheap part. On a typical AI25 project we throw away far more material than we keep - the selects folder is a fraction of what the pipeline produced. Nobody pays us for the volume. They pay for the folder of keepers and for the confidence that someone with a trained eye decided what went in it.

Which means the scarce skill moved. It is no longer "can you make this". It is "should this exist, and is it better than the other forty versions". That skill has an old-fashioned name: taste.

Machine taste already runs your feed

There is a tempting objection here - recommendation algorithms curate billions of pieces of content daily, so machines clearly have taste. Look closer at what they optimize for. A feed algorithm is trained on what held attention yesterday. It is pattern-matching against the past, at scale, with no opinion about what should come next.

That kind of curation flattens things. It rewards the average of what worked, which is why so much feed content converges on the same look, the same pacing, the same thumbnail grammar. A studio's taste has to do the opposite: commit to a specific look before the market has voted on it. Machine taste reflects what was. Human taste places a bet on what should be.

We feel this on every pitch. The references a client brings from their feed are usually the safest possible version of the idea. Our job is to move them off that average.

How a studio trains taste on purpose

Taste sounds like a talent. Inside a production company it behaves more like a process, and it can be built:

1. Reference before generation - No AI25 project starts with a prompt. It starts with a reference deck - film stills, photography, grade references - argued over until the look is locked. The model only enters after the eye has decided.

2. A named owner of quality. Every project has one person whose call is final on selects. Committees average; a single accountable eye commits.

3. A kill list. We keep notes on what we never ship: the plastic skin, the drifting hands, the lighting that looks like nowhere. Studying failure is faster than studying success.

4. Looking outside the feed. The team watches films, not timelines. If your inputs are algorithmic, your outputs will be too.

None of this is romantic. It is checklists and arguments in a review room. But it compounds, project after project, into something a prompt cannot replicate.

Taste is slow, and that is the point

Everything else in this field accelerates. Models improve monthly, render times shrink, costs fall. Taste refuses to speed up. It is built by years of watching, making, and being wrong in front of colleagues.

That slowness used to look like a weakness. Now it is the moat. Anything fast gets copied fast. The one asset your competitors cannot download is a decade of decisions about what good looks like.

AI is the tool. Craft is the foundation. The studios that survive the next few years will not be the ones with the best model access - everyone has the same model access. They will be the ones where somebody in the room can say "no, not this one" and be right.