Case study
AI presenter looks fake: the test that catches it
We wanted a daily video without filming every day, so an image model drew the presenter in new places and clothes. The pictures were tidy and still not quite him. Here is the test that caught it, and how to run it on your own pictures.
Key takeaways
- A picture only has to be slightly off for someone who knows your face to distrust the whole video.
- Turn each of your standards into a yes-or-no question, so no score can wave a failure through.
- Compare generated pictures with proper portraits, because a selfie flatters the model.
- Agree before you publish what happens when nothing passes, and let the answer be no video that day.
- Keep cutaways free of lookalikes, and label every post as made with AI.
An AI presenter looks fake long before a viewer can say why, and the pictures can pass every score you set. What caught it for us was a different test: plain yes-or-no questions that rule a picture out, and likeness judged against proper portraits instead of selfies. Every daily video we publish now starts from a real photograph, and that decision came out of the test.
The test that beats a score when an AI presenter looks fake
Turn your standards into questions with a yes or no answer, and let any yes rule the picture out before it is used.
A score from a model tells you how tidy a picture is, not whether the person in it is you. Ours was generous. A face with a wide smile and teeth showing came back at 9 out of 10 while the model's own instructions said teeth should rule it out. A stronger model, shown the same picture, saw the teeth and gave it 6.
Questions leave less room to argue: teeth showing? another person in the frame? writing anywhere? eyes closed? Any yes and the picture is out, whatever the score said.
The second half of the test is what you compare against. Old selfies, taken at arm's length and close to the lens, make a drawn face look closer to you than it is. A few good portraits, facing the camera in ordinary light, are a harder and fairer standard for judging likeness.
Why a daily AI presenter goes wrong at all
A viewer checks a face without deciding to, so a presenter who is almost you reads as false even when the clothes, the light and the background are right.
We make a short daily video for YouTube Shorts, Instagram Reels, TikTok and Facebook on our own channels. The person who presents our work tells the day's article in under a minute, using a voice and face cloned with his consent, and every post is marked as made with AI on each network. To keep the videos from looking identical, each was meant to show a new outfit and a new place, drawn by an image model.
That is where a channel built on a person's face can lose what it was built for. For a business, the face on camera is the trust. A stranger at the next table, or an emblem on a shirt nobody chose, tells viewers the whole thing was manufactured, and they stop watching for the reason they came.
What our first run turned up
The drawn pictures were tidy and wrong in small ways nobody had asked for.
In a busy cafe they put strangers at the next table. They added an emblem to a plain polo shirt. Asking for empty places and plain clothes helped; it did not answer the question that mattered, which was whether the person in the picture was him.
The fallback failed too. On a run where all sixteen pictures were ruled out, the video platform's own presenter came out with his face badly distorted.
The turning point was the comparison, not another setting. Judged against real portraits rather than old wide-angle selfies, the drawn pictures were only partly like him.
What we changed
Two things: harder questions, and real photographs to judge against.
The rules became plain yes-or-no questions, and any yes ruled the picture out before the score was read. Judging went in two steps: a quick model rules out plain failures across four pictures, and a stronger one sees only the best two and alone may pass one. No video is made from a face that failed.
Then came the decision the evidence pointed to. Every video now starts from a real photo, as it is, one of 42 used once each in turn, with the varied place as the background. The drawn looks are switched off and kept.
We also asked for footage between his lines again, because a talking head on its own is dull, on one condition: no stock clip may show a man who could pass for him. A lookalike in a cutaway leaves viewers wondering which one is real.
Where our daily videos stand now
Live since 16 October: every daily video since has started from a real photo and gone out on all four networks.
The judging work is not wasted. It is what told us, in numbers, that the drawn pictures were not good enough, something we could argue about but not measure until the questions and the scores existed.
What you can check yourself
- Ask someone who knows your face to pick the AI picture out of a handful of real ones, without saying how many there are.
- Play the video with the sound off and watch only the mouth and eyes, where a face that animates badly shows first.
- Look at the hands if they appear at all, and count the fingers.
- Look through the background of every generated picture for a logo, lettering or a passer-by who was never in the room.
- Go through your scheduled posts and confirm each one that uses a cloned voice or face says so.
What to do next
Decide now what happens on a day when nothing passes, before that day arrives. Making no video is better than publishing a weaker picture, and an unchecked fallback is how a distorted face reaches an audience.
Have a few proper portraits taken, in good light, facing the camera. That is the standard anything generated will be measured against, and it is what you fall back on.
Settling that test, and building it into a publishing routine, is the kind of work our engineers take on for clients who want to be on camera without filming every day.
Questions people ask
Can an AI presenter be made to look exactly like me?
An image model draws someone who resembles you rather than you, so the pictures come out close and not exact. In our experience the gaps show in the mouth, the hands and the background, and people who know your face find them quickly. You can raise the standard a picture has to meet, but raising it does not close the gap.
Why did the AI pictures include strangers and logos nobody asked for?
Image models fill in whatever makes a scene look complete, and a room with people in it looks more complete than an empty one. Asking for plain clothes and empty places helps, but it does not remove every addition. Asking a model to leave something out can even put it in, so check each picture rather than trusting the instruction.
How do I check whether an AI picture of me is good enough to publish?
Put it beside a real portrait and ask someone who knows your face which is which, without telling them the answer first. Then read it on its own: the mouth, the eyes, the hands, the clothes and the background are where the problems show. If it does not pass, publish nothing that day rather than something weaker.
Should I stop using AI video if the face is not perfect?
Not necessarily. What we settled on for our own videos was to start each one from a real photograph of the presenter and vary the place around it, which keeps a daily video going without asking viewers to accept a face that is slightly off. If you have a set of real photos and a routine for using them, you may not need a drawn face at all.
Do I have to say a video was made with AI?
We label ours on every network, and labelling is the safer habit whatever you decide. The rules differ between platforms and between countries and they change, so check the current guidance for where you post, and take advice where the law applies to you.