A year ago I put a fully AI generated thumbnail live on one of our own channels. It looked incredible in the folder. It looked expensive. It got a click rate roughly a third below the hand built version we swapped back in two days later. That was the first honest data point I had, and it set the tone for everything since.
Twelve months on, AI image tools are in almost every part of how we work. They are just not where most people assume. The generator is not the designer. It is the intern who is very fast, very confident, and completely unable to tell you why a thumbnail works.
So here is the year written down properly. What we adopted, what we quietly dropped, and where the money actually went.
The first six months were mostly wasted
I want to be blunt about this because a lot of studios will not be.
For the first half of the year we tried to use AI at the top of the funnel. Prompt the concept, generate the scene, composite the face, ship. The output was fast. It was also strangely lifeless in a feed, and I could never explain why to a client without sounding like I was making excuses.
Eventually I stopped guessing and started looking at what the images had in common. They were all beautifully lit from nowhere in particular. Every surface had the same soft plastic sheen. Backgrounds had detail everywhere instead of detail where it mattered. Faces were symmetrical in a way real faces are not.
None of that is a rendering flaw. It is a taste flaw, and it is the exact same problem we have been writing about with libraries of pre shot imagery for years. If you have read our piece on why stock imagery flattens a channel, the failure mode here is identical. The tool optimises toward the average of everything it has seen. The average is invisible.
A thumbnail does not compete with other thumbnails on quality. It competes on difference. AI is trained to produce the least different image possible.
That sentence took me about eight months to arrive at and it now governs how we brief every generation.
Where AI actually earned its place
Once I stopped asking it to have ideas, it got genuinely useful. Three jobs, all of them unglamorous.
Cleanup and rescue. Creators send us stills pulled from footage. Compressed, soft, sometimes 720p from a phone. A year ago a bad still meant a reshoot request or a heavy crop. Now it usually means fifteen minutes of repair. We wrote up the actual workflow in our notes on rescuing soft source frames, and that piece is now the most used internal document we have.
Background separation. Cutting a subject out of a messy frame used to be the slowest part of a build. Hair against a busy wall could eat forty minutes. It is now near instant and near perfect, and it fails in predictable ways we know how to catch.
Filling space that was never shot. Extending a frame sideways so a face can sit on the third instead of dead centre. Nobody sees it. That is the point.
Notice the pattern. Every one of those is a labour problem, not a design problem. AI is very good at removing tedium and very bad at removing ambiguity.
What we tested, and what the numbers actually said
We test styles on our own channels before they go anywhere near client work, which is a policy I would recommend to anyone. It means the expensive mistakes are mine.
Across roughly a year of that, on a tech commentary channel we run in the low tens of thousands of subscribers, fully generated scenes consistently underperformed hand composited ones. On a finance style channel we test with, sitting around 80k subs, the gap was smaller but still there. On a couple of gaming style tests, generated backgrounds behind a real photographed subject actually did fine, sometimes marginally better.
I am not going to hand you a percentage and call it a rule. The honest read is this. When the audience expects a real person to be present, generated faces cost you. When the background is understood to be decoration, generation is free money.
That is a much less exciting conclusion than "AI thumbnails get X percent more clicks", and it is the only one I would defend.
Two more things I learned about testing itself. First, most of the comparisons people run are noise. A swap that moves click rate by half a point on four thousand impressions has told you nothing, which is the whole argument in our piece on tests that end too early. Second, judging a thumbnail on a big monitor is close to useless. We check every build the way a viewer meets it, surrounded by competing images at small size, and the method is written out in how we preview thumbnails in context.
The colour problem nobody warned me about
This one caught me off guard.
Generated images tend to arrive with gorgeous, subtle, wide colour. Lots of midtones. Gentle transitions. Photographically lovely. In a feed it turns to grey mush next to a thumbnail built on three committed colours.
We now flatten almost everything that comes out of a generator. Crush the palette, force separation between subject and background, push one accent colour that the rest of the image is not allowed to touch. It looks aggressive in isolation and correct in context, which is the whole game. If you want the underlying logic rather than the fix, it is laid out in our thumbnail color theory reference, which covers why a palette that looks restrained on its own dies in the feed.
The tools have improved a lot in twelve months. This particular tendency has not improved at all, and I suspect it will not, because tasteful and legible are different goals.
The uncanny tax is real and it is getting worse
Early on, obviously synthetic images were just ignored. Now audiences have learned to spot them, and ignoring has turned into mild resentment.
I see it in comments on tests. Not "this looks fake" so much as "why does this channel look like everything else". Viewers cannot always name what is wrong, but the trust hit is measurable in the second click, the one where they decide whether to watch a second video.
There is a psychology layer here about faces, eye lines, and implied effort that goes deeper than I can fit in one post. We keep it in the thumbnail psychology guide because it applies whether the image was generated or shot.
My practical rule after a year: the human in the thumbnail must be the real human from the video, unedited in ways that would embarrass them. Everything behind that person is negotiable.
What this changed about how we price and work
Almost nothing, and I think that is the interesting part.
Our pack is still $250 for 8 custom thumbnails, one per video, which works out around $31 each. Still unlimited revisions, still a dedicated designer, still title consultation included, still 48 to 72 hours per thumbnail. Add ons are unchanged too, an A/B variant at $15, a title pack at $25, rush at $20.
People expect AI to have collapsed that number. It has not, because the generator was never the cost. The cost is the judgement about what the thumbnail should say, the argument about the title, the four revisions where a creator slowly works out what they actually meant. That work did not get cheaper. It got slightly more pleasant, because the person doing it spends less time masking hair.
What did change is throughput on the boring end. Rescue work that used to trigger a "can you reshoot this" email now just gets done. Fewer blocked jobs. Faster turnarounds inside the same window.
We also never use templates, and AI made me more committed to that rather than less. A template and a generator are the same trap wearing different clothes. Both produce the average. You can see the range that comes out of custom work in our work, and the range is the whole argument.
What I would tell a creator doing this themselves
Use it for repair, not for invention.
Do not generate a face. Photograph one, badly if you must, then fix the photograph.
Generate backgrounds and environments freely, then flatten them until they are almost rude.
Decide your concept before you open the tool. If you prompt your way to an idea, you will land on the idea the model finds most likely, and so will everyone else in your niche that week.
Test on real impressions, over enough of them to mean something, and be suspicious of any result you were hoping for.
And if your bottleneck is not thumbnails at all but the sheer time video takes, that is a different conversation. Our sister studio handles video podcast editing from $199 with clips at $25, and quite often the honest answer to "how do I grow" is to publish more, not to redesign the same thumbnail a sixth time.
One year in, the short version
AI did not replace the designer. It replaced the tedious hour before the designer starts.
The images got better all year and the click rates did not follow, which tells you the constraint was never image quality. It was clarity, difference, and knowing what the video is actually about. No tool has offered me help with any of those yet.
Ask me again in a year. I will still be flattening the colours.