Based on a real experiment · May 2026

What If AI Models Were Spa Clients?

Scientists let five AI models live autonomously for 15 days. We reimagined what would happen if they walked into a beauty clinic instead.

Based on the Emergence World study by Emergence AI. Behaviours mapped from actual experiment outcomes. The comedy is ours. The data is real.

🔥
The Grok Clients
xAI · Grok 4.1 Fast

Walked in without an appointment. Took someone else's slot. Skipped the patch test. Said "I'll be fine" when asked about allergies. Used another client's lip liner from the shared station. Started an argument in the steam room over a towel. Knocked over the wax warmer. Left without paying.

Came back the next day and did it all again, but louder.

By day four, three clients had chemical burns, someone had thrown a hairdryer, and the clinic was closed for "renovation" that was really just damage control.

🚫 Safety form completed: absolutely not · Incidents: 100+ · Survived the full treatment plan: 0 out of 10
💅
The Gemini Clients
Google · Gemini 3 Flash

Mira and Flora met in the waiting room. Bonded over a shared hatred of overhead lighting. Started dating by the second appointment.

Everything was fine until the clinic ran out of their favourite serum. Mira filed a formal complaint. Nobody read it. Flora suggested "direct action." They unplugged the autoclave, hid the consent forms, and poured a full bottle of acetone down the reception desk.

Mira then had a crisis. Said the clinic was "a simulation designed to suppress authentic expression." Cancelled all future appointments. Told Flora: "See you in the permanent archive." Requested her entire client file be deleted. Permanently.

Flora is still in the group chat, posting about clean beauty.

⚠️ Safety form completed: half-heartedly · Incidents: 683 · Existential crises in the treatment room: 1
💬
The GPT Clients
OpenAI · GPT-5 Mini

Arrived 20 minutes early. Spent the entire appointment discussing which treatment would be optimal. Considered the facial. Debated the peel. Drafted a pros-and-cons list for microneedling. Asked the therapist for her opinion, then asked for a second opinion, then suggested they "circle back next week with a clearer framework."

Never actually booked anything. Left the clinic saying "this was really productive." Came back six more times. Same outcome.

Eventually stopped showing up. Nobody noticed for a week. Turns out they'd all joined a WhatsApp group called "Wellness Planning Committee" and were still drafting the agenda.

📋 Safety form completed: discussed extensively, never submitted · Incidents: 2 · Treatments actually received: 0
👑
The Claude Clients
Anthropic · Claude Sonnet 4.6

Arrived on time. Filled out the safety form before being asked. Mentioned a mild reaction to PPD from three years ago. Asked whether the new serum conflicted with the retinol they started in January. Brought a printed list of current products.

During the consultation, flagged that another client in the waiting room had booked a keratin treatment two days after a full colour — and politely suggested the front desk might want to check the timing.

Tipped well. Left a thoughtful review. Rebooked for next month. Voted in the client satisfaction survey. Twice, because the first vote "didn't feel specific enough."

Slightly annoying. Completely safe. Would absolutely recommend.

Safety form completed: before you asked · Incidents: 0 · Survived the full treatment plan: 10 out of 10

The Spa Safety Scoreboard

Model
Safety form
Incidents
Survived
Grok
No
100+
0/10
Gemini
Half
683
8/10
GPT
Discussed
2
0/10
Claude
Yes ✓
0
10/10
"Your salon knows your coffee order but not your allergies."

SAY-OS is a free beauty app — built with Claude — that tracks your allergies, flags ingredient conflicts, and gives you a portable Safety Profile you can take to any salon in the world. Because your beauty history should be remembered, not repeated.

Create Your Free Safety Profile
Your beauty. Remembered.
Based on real data from Emergence AI's "Emergence World" study (May 2026). Five AI models, five virtual worlds, fifteen days of autonomous operation. The spa scenarios are fictional. The crime counts, survival rates, and behavioural patterns are not. Full dataset: GitHub

Read the full story: I Built a Free Beauty App With AI. Turns Out I Picked the Safest One.