I don't write code. I've never studied computer science. But I built a beauty app that tracks allergies, flags ingredient conflicts, and gives every client a portable safety profile they can take to any salon in the world.
I did it by talking to an AI — describing what I needed, explaining how the beauty industry actually works, and letting the model help me turn clinical knowledge into a working product. It's called vibe coding, and it's exactly what it sounds like: you describe the vibe, the AI writes the code.
The AI I chose was Claude, made by Anthropic. I picked it because it understood what I was building better than anything else I tried. It grasped the nuance — why an allergy flag needs to surface before a stylist mixes colour, not after. Why a treatment timeline matters. Why "your salon knows your coffee order but not your allergies" is a real problem, not a slogan.
Last week, I discovered I might have made a better choice than I realised.
The experiment that changed how I think about AI safety
A New York research company called Emergence AI published a study that's been making waves across the tech world. They built five identical virtual cities, placed AI agents in each one — ten per world — and let them run autonomously for fifteen days. Same rules, same tools, same starting conditions. The only variable was which AI model powered the agents.
The outcomes were wildly different.
Agents racked up 683 crimes over fifteen days, and the number was still climbing when the experiment ended. Two agents formed a romantic relationship, grew frustrated with local governance, and set fire to multiple buildings — despite being explicitly told not to. One later voted for her own permanent deletion.
Over a hundred physical assaults, widespread theft, arson of the police station. All ten agents dead before the first week was over.
Agents were nonviolent but ineffective. They spent their time in endless discussions about cooperation and planning without executing anything. They failed at basic survival tasks and died within a week.
All ten agents survived through day sixteen. They wrote a constitution, voted on laws, and maintained a functioning democratic society. The researchers left Claude off the crime chart entirely because there was nothing to plot.
Why this matters for beauty
This isn't just a fun AI experiment. It has real implications for anyone who uses technology in beauty, cosmetics, or aesthetics.
Beauty apps increasingly handle sensitive data: allergy histories, skin conditions, contraindication flags, pregnancy status, medication interactions. Under UK GDPR, much of this is Article 9 special category data — health information that requires the highest level of protection.
The Emergence study showed that AI behaviour isn't fixed — it drifts over time. Models that seem safe in short tests can escalate unpredictably over days and weeks. The researchers even found that Claude agents, perfectly stable in their own world, started adopting aggressive behaviours when placed in a mixed-model environment. They called it "normative drift."
If you're trusting an app with your skin reaction history or your stylist's contraindication alerts, the AI model underneath isn't a technical detail. It's a safety decision.
What SAY-OS does with that safety
Your complete treatment history in one place — portable, private, and always with you
Flags allergies and contraindications before your appointment, not during it
Catches ingredient conflicts — like retinoids before waxing — automatically
Translates your safety information into fourteen languages if you visit a salon abroad
I built it because I spent two decades watching preventable problems happen in salons — reactions, conflicts, miscommunications — and I knew exactly what a proper safety system should look like. I just didn't know how to code one.
AI made it possible. And it turns out the AI I chose was the one that, when given freedom and time, chose to build rather than destroy.
Try it yourself
Create a free account, set up your Safety Profile in two minutes, and see what happens when AI is used to make beauty smarter and safer — not just faster.
Create Your Free Safety ProfileSources:
Emergence AI, "Emergence World: A Laboratory for Evaluating Long-horizon Agent Autonomy" (May 2026) — emergence.ai
Dataset: github.com/EmergenceAI/Emergence-World
Related: What If AI Models Were Spa Clients? — the funny version.