Last month I ran a small experiment I am not especially proud of. I took a face from one of my own thumbnails, ran it through a face swap model, and pasted it onto a body shot I had generated. Total time, eleven minutes. The result looked good on my monitor. It looked fine at 320 pixels wide. And when I put it in a click test against the honest version, it lost. Not by a little. The honest version won by roughly a third more clicks in the sample I ran on my own channel.
That result surprised me enough that I kept poking at it for three weeks. This post is what came out of that.
The thing nobody says about face swaps
Face swap tools are now good enough that the seams are gone. Skin tone matches. Lighting direction mostly matches. Hair edges are clean. If you are looking for the classic tells, the smeared ear, the wrong jaw angle, the eyes that point in slightly different directions, you will often not find them anymore. The technology cleared that bar sometime in the last eighteen months.
So the argument against using them has quietly moved. It is no longer "it looks fake." It is "the viewer knows something is off and cannot tell you what."
That is a much harder problem, and it is the one that actually costs you clicks.
I think of a thumbnail as a promise. You are standing in a feed next to twenty other promises, and the viewer is deciding, in something under a second, which one is worth their next eight minutes. Everything in the frame is evidence. Your face is the strongest single piece of evidence you have, because a real face carries a hundred small signals of a real moment. A swapped face carries the geometry of a face without the moment behind it.
What I actually tested
I ran this on my own channels, not on client work, because I am not willing to use anyone else's audience as a lab. Small samples, so treat the numbers as directional and nothing more.
Three setups:
- Real photo, real reaction. A frame pulled from the actual video, color graded, cut out, composited.
- Swapped face on a generated body. My face, a generated pose, blended.
- Fully generated person. No real reference at all.
Across nine videos on a mid sized channel, the real photo version held the best click through rate every time. The swapped version sat in the middle. The fully generated person did the worst, and it also had the worst average view duration, which is the number that made me stop the test early. Getting a click and then losing the viewer at fifteen seconds is a worse outcome than not getting the click.
The gap between the real photo and the swap was smaller than I expected, somewhere in the range of ten to twenty percent relative on those particular videos. The gap in retention was not small at all.
A face swap does not fail because it looks fake. It fails because it removes the one thing a thumbnail is actually selling, which is that a specific person had a specific reaction to something.
Why the uncanny part moved from the face to the eyes
When I zoomed in on the losing versions, the face itself was clean. What was wrong sat behind the eyes.
Real reaction photos have a chain of cause and effect. The eyebrow moves because the eye widened. The mouth is asymmetric because one side of the face reacted first. The neck tenses. A tendon shows. The head tilts a few degrees off center because nobody holds their head perfectly straight while being surprised.
Swaps break that chain. The face is doing one emotion and the body is doing another, or the face is doing an emotion with no physical cause underneath it. Your viewer does not analyze this. They just feel a small friction and scroll past. This is the same failure mode I wrote about when I looked at where generative image tools still break down, and I keep running into it from different directions.
The practical version of this rule: if you are going to fake a face, you have to fake the whole body language too, and at that point you are doing more work than shooting a real reaction with a phone in your kitchen.
The trust cost is delayed, which is why people miss it
Here is the part that makes this genuinely hard to argue about. The damage from a swapped thumbnail does not show up in the first twenty four hours. Click through rate might even go up briefly, especially if the generated version is prettier than reality. Prettier does get clicked.
The cost lands later, in three places.
First, retention. The person who clicked expected the version of you in the thumbnail. The person in the video is someone slightly different. That mismatch reads as bait even when nothing dishonest was intended.
Second, repeat viewers. A channel builds a face memory over time. If your thumbnails drift toward an idealized version of you, the returning audience notices the drift before they can name it, and returning viewers are the only audience that compounds.
Third, comments. This one is brutal and public. Once one person types "why does this look AI" under a video, that comment sits at the top and reframes the whole thumbnail for everyone who reads it.
I have watched channels absorb all three at once and conclude their content got worse. Their content did not get worse.
Where AI in thumbnails is genuinely fine
I am not an AI purist. That would be a strange position for someone who uses these tools most days. The line I draw is about the subject, not the pixels.
Fine, in my opinion:
- Background generation and extension. Nobody has a parasocial relationship with your background.
- Object generation. A phone, a stack of cash, a car, a chart. Props are props.
- Cleanup and relighting on a real photo of a real person.
- Cutout and masking. I compared the current crop of tools in my notes on background removal and the good ones now beat manual pen tool work on hair.
- Upscaling a good frame from bad footage.
Not fine, in my opinion:
- Swapping your face onto a body that was never yours.
- Generating a fake reaction you never had.
- Putting a real person who is not you into your thumbnail wearing an expression they never made. This one is not a taste question. It is a legal and ethical one, and I will not do it.
The middle ground is the interesting part. Generating a crowd in the background of a shot is fine. Generating a specific recognizable person in that crowd is not. The test I use is simple: if the viewer knew exactly how this image was made, would they feel tricked about something that matters? Backgrounds, no. Your face, yes.
The sameness trap sitting underneath all of this
There is a second problem with leaning on generated faces, and it has nothing to do with trust. It is that everyone is pulling from the same models, and those models have opinions.
Generated people converge. Same skin quality, same symmetry, same slightly waxy highlight on the cheekbone, same eyes. Put four generated thumbnails side by side from four unrelated channels and they look like siblings. This is the exact failure I described in the piece on stock photography flattening everything, and generative tools have industrialized it.
Your face is the only asset in your thumbnail that no competitor can produce. Trading it for a smoother version of a generic face is a bad trade at any click through rate. It also leaks into your whole channel presence, because the first three seconds on your channel page are mostly a grid of your own thumbnails arguing with each other.
What I do instead, practically
None of this is useful without a replacement workflow, so here is the one I actually use.
Shoot ten seconds of reaction after every video. Not posed. Phone on a stack of books, window light, react to the thing you just talked about. Ten seconds gives you roughly three hundred frames and two or three of them will be usable. This is the single highest leverage habit for anyone making thumbnails, and it costs nothing.
Grade and cut out, do not regenerate. Fix the light, fix the color, remove the background, sharpen the eyes. Everything that was true about the photo stays true.
Use AI for everything behind the person. Backgrounds, props, atmosphere, glow. This is where these tools earn their keep, and it is where the visible craft in most of our work actually lives.
Test with variants, not with vibes. If you genuinely do not know whether the honest photo or the retouched one wins, run both. That is what the A/B variant add on exists for, and fifteen dollars is cheaper than a month of guessing.
Write the title before the image. Half the thumbnails that get replaced with a generated face were never a face problem. They were a title problem. If that is where you are stuck, a thumbnail consultation will usually find it faster than another round of image generation.
If you are handing this to someone else, be explicit about the AI line in the brief. I laid out the format I like in how to brief a designer in five lines, and "no face generation, real photo only" is one line that saves an entire revision cycle.
The version of this problem coming next
Face swaps are the current argument. Full video generation is the next one, and it is going to be worse, because a fake thumbnail can at least be checked against the real video. When the video itself is generated there is nothing to check against.
That pushes the value of provably real footage up, not down. The channels that keep showing up as themselves, on camera, in a room, with a real face doing real things, are going to have an asset that cannot be cloned by anyone. The same logic is why I still think unedited long form conversation holds value, which is most of the reason we run podcast editing as a separate service rather than folding it into everything else.
Closing thought
I do not think face swaps are evil. I think they are a shortcut that trades a permanent asset for a temporary lift, and most people making that trade have not looked closely at what they are giving up.
Use the tools. Generate the background, generate the props, fix the light, remove the halo around the hair. Then put your actual face in the middle of it, badly lit hair and asymmetric eyebrow and all. That is the part that no model can produce, and it is the part your audience is actually clicking on.
If you want a second opinion on whether a thumbnail is reading as honest or as generated, send it over. I will tell you what I see.