Show Notes
Sixty percent of the faces training AI were never asked
Katrien Grobler woke up on the 29th of December and saw a word written above her bed. Not metaphorically. She describes seeing the name in her mind's eye, getting up, and telling her team they were pivoting that day.
The word was Twinnin.
She had spent the previous year running AI KAT, making AI-generated video ads for clients like Capital.com and Reckitt's Dettol brand. On one project she needed a trader character. Rather than generate a synthetic face, she pulled photos of an investor she knew, ran them through Nano Banana, and built the character. The client's first reaction was the whole business model in one sentence: is this a real person?
It was. He hadn't been asked. She called him afterward, apologised, and made him sign a licence. That awkward phone call is now a company.
Her argument runs on arithmetic she calls the math of human likeness. There are 8.3 billion people on the planet. Frontier labs have already consumed images of humans at a scale in the tens of billions apiece, and by 2030 she projects the training requirement climbing into the trillions. Models cannot reuse the same face indefinitely. They need novelty, and novelty means new humans. Meanwhile the legal ground is moving under everyone: the EU AI Act's obligations landed, Denmark is rewriting its copyright law to treat likeness as property, and the NO FAKES Act cleared the Senate Judiciary Committee unanimously in June 2026. Demand is accelerating. Supply is becoming illegal. There is no clearing house in between.
That gap is the company. Twinnin verifies a human with government ID and a liveness check, lets them set granular terms for how their face can be used, hashes the consent record so it is auditable, and routes payment when a brand or lab licenses it. Seventy-five percent goes to the human, always, agent or no agent.
Then there is the half of the market that wants nothing to do with any of it. When Deadline clipped her panel and posted it, the comments were not friendly. One person told her she looked like an Incredibles villain, which she thought was fantastic. Most of the rest were worse. She did not dismiss them. She built for them. The Defend product exists because she concluded that people who want to opt out have nowhere to formally do so, and that a documented, timestamped "no" is the only evidence that holds up when a deepfake surfaces.
The other half of this episode is about how she operates. She has been to a contrast therapy club in Canary Wharf more than three hundred times since April 2025. Sauna, then a three minute plunge at one degree, then sauna again, sometimes three cycles before a call. She does not describe it as recovery. She describes it as a forced thinking environment. You cannot leave, you cannot distract yourself, so the problem you brought in gets solved. Most of her best product decisions happened underwater at one degree.
Twinnin: https://www.twinnin.ai
Katrien Grobler on LinkedIn: https://www.linkedin.com/in/katriengrobler
Ryan Estes on LinkedIn: https://www.linkedin.com/in/estesryan/
Build your audience for life! https://inboxalchemy.co/
Free Competitor Analysis from CEOfriend.ai!
If you're not AI native; you're not getting the job! https://ainativestudent.com/
Get 35% off any supplement subscription with Momentous! https://crrnt.app/MOME/8RDrnXDd
Use code RYAN30 to save $30 on your AI men's fashion stylist! https://taelor.style/
Your podcast's autonomous AI sponsorship agent! https://gethowdi.com/
