The Good Kind of Fake Video

Open any feed today and you'll find a face saying something it never said. A politician, a celebrity, a stranger — dropped into a scene that never happened, mouthing words they never spoke. In the space of a couple of years, face-swapped video went from a lab curiosity to background noise, and most of the time it's pointed at you: to sell, to smear, to scam, to make you doubt what you're even looking at.

At Annoying-Is-Caring, we build the exact same technology. We use it to make Grandpa the hero of an action movie. On purpose. And, quietly, to teach a machine something profoundly human.

Same tool, opposite intent

Here's the uncomfortable truth about face-swap: the technology itself is neutral. The pixels don't know whether they're faking a confession or putting your dad's face on a slow-motion, walk-away-from-the-explosion hero shot. What separates the two isn't the software — it's intent, consent, and craft.

Ours is deliberate on all three:

Intent

To make one specific person you love smile — not to fool a crowd.

Consent

The videos are made by you, for your person, from photos you own, shared privately with someone who's expecting it. Nobody's face is broadcast to deceive a stranger.

Craft

And this is the part worth lingering on, because it's where creepy and delightful actually part ways.

Teaching a machine to be funny

People assume a joke company is a thin wrapper around a big AI model. It isn't. Comedy is one of the genuinely hard, unsolved problems in artificial intelligence — and it turns out to be a wonderful teacher.

A model can write a grammatically perfect sentence all day. Getting it to write one that makes a particular 78-year-old in a particular town chuckle at breakfast is a completely different sport. It takes timing, surprise, brevity, and a small personal detail that says this one was written for you. So we don't take the first draft: the system auditions several versions of every joke and lets an AI judge pick the one that lands hardest. On video, the narrator even learns the oldest trick in stand-up — the pause, right before the punchline.

Face-swap comedy is where all of that gets grounded. A joke is abstract until it's on a face you recognize, in a scene that's absurd precisely because it's them. Building that well forces the models to understand not just “is this funny?” but “is this funny about this person?” — a much richer, much harder question, and a much better lesson.

Getting the right face on the right person

Which brings us to the unglamorous engineering that actually eats our time — and the cleanest line between a malicious deepfake and a joyful one.

A face-swap can look flawless — crisp, smooth, believable — and still be completely wrong, because it put the right face on the wrong person. In a clip of one person that never happens. But point it at a real scene — two people dancing, a dinner table, a movie moment with a cut in the middle — and suddenly identity is the whole ballgame.

Our clearest teacher was a two-dancer clip. Frame by frame it looked great. And right after a scene cut, it swapped both faces onto a single dancer's head — one person wearing two identities, fighting for the same skull. A beautiful frame that was completely wrong about who was who. A number on a chart shrugs at that; a human recoils instantly, because we are all exquisitely tuned to recognize the people we love.

Fixing it is the job: we cut the video at each scene change and decide who's who within each shot; we keep only the single best detection per head, so a close-up doesn't earn two identities; we run cheap, near-instant checks that veto obviously wrong swaps a pure “quality” score would happily wave through; and we model the face with tens of thousands of points so an identity survives a head-turn instead of melting into someone else.

We obsess over this for a soft reason with a hard consequence: the gift only works if the person on screen is unmistakably the person they love. Get identity wrong and you don't have a funny video — you have an uncanny one, and the warmth is gone. Getting it right is exactly the discipline that separates a tool built to delight from one built to deceive.

The future we actually want

Fake video isn't going back in the box. The interesting question was never “can we stop it?” — it's “what do we point it at?” You can aim this technology at eroding trust, one convincing lie at a time. Or you can aim the very same pixels at a thirty-second surprise that makes someone laugh out loud at their phone and text back “SEND ME THAT AGAIN.” One of those futures we'd like to live in. And the wonderful part is that building the joyful version well — with consent, with care, with the right face on the right person — is also how we're teaching machines to do something no benchmark fully captures yet: to be funny, about someone, on purpose.

Same tool. Opposite intent. That's the whole company.