I challenged 10 AIs to jail me. Result? They arrested someone else (or learned the alien language).

Or: How I discovered that 10 out of 10 artificial intelligences have serious problems with Italian and my face

Sometimes things happen by chance. Inspiration strikes suddenly, and you can't help but follow it. And I, our genius, thought: "Hey! Let's turn ourselves into a prisoner of mediocrity using AI!" Spoiler: most of these so-called artificial "intelligences" turned out to be more artificial than intelligent.

From the Crime of Mediocrity to the Disaster of Facial Recognition

I was writing an article about online mediocrity. Then I had a crazy idea: create a mugshot of myself as a "creative criminal," accused of violating mediocrity standards. I took a photo of myself (a really bad one, by the way), wrote a very detailed prompt, and fed it to nine different AI models using lmarena.ai. The result? A festival of errors that deserves a merciless analysis.

The Carnage: Brutal Model-by-Model Analysis

Flux-1 Kontext Dev – FAILED​

The crime: He hung the sign on the wall instead of handing it out. WHAT THE MEANING IS THAT?! It's a mugshot, not a contemporary art exhibition! Besides, the resemblance is there, but the prompt's interpretation is elementary school-esque.

Flux-1 Kontext Pro – TOTAL DISASTER​

The crime: He wrote "VNOULAZIONE MIOLLISOINIA 'DL FAIE" and other incomprehensible bullshit. Dear Flux Pro (which is the PAID version, by the way), if you don't know how to write in Italian, at least say so. It costs even more and he can't spell. That's enough to report to Codacons.​

Flux-2 Pro – WHO IS THIS STRANGER?

The crime: It's COMPLETELY transformed my face. It could be anyone. It could be my cousin, the local baker, Brad Pitt aged badly. But it's not me. Zero resemblance to the original image. Total failure.

Flux-2 Flex – BELOW-ZERO REALISM​

The crime: It looks anything but realistic. The image has that plasticized, '90s action figure-style feel. If the goal was "hyper-realistic," someone should explain to Flux what "realistic" means.

Gemini 2.5 Flash (Nano Banana) – ALMOST, BUT…

The crime: He wrote "VIOLATIONE" instead of "VIOLAZIONE." Dear Google, it's 2025, and Italian has been around for centuries. A misspelling of such an important word ruins everything. It's a shame, because the similarity and the atmosphere were good.​

GPT-Image-1 (OpenAI) – BUT WHO IS THIS GUY?

The crime: They completely distorted the image. It's not my face. Period. ChatGPT/OpenAI created a beautiful, cinematic, Oscar-worthy image... but of someone else. It's like ordering a Margherita pizza and getting sushi.

The Winners Are…

Nano Banana Pro (Gemini 3 Pro) – THE CHAMPION

Finally! It maintains the resemblance, spells out "VIOLATION OF MEDIOCRITY STANDARDS" correctly, features perfect lighting and shadows, and believable textures. It costs a little more, but IT WORKS. It's like comparing a surgeon and a butcher: both cut, but only one knows where to cut.

​Qwen-Image-Edit (Alibaba) – THE REAL HIDDEN WINNER

Qwen-Image-Edit, Alibaba's 20 billion parameter model has done what others have only dreamed of. It's built on architecture dual-path: use the Qwen 2.5-VL encoder for semantic understanding and a VAE (Variational Autoencoder) for visual fidelity. This division allows him to make both broad semantic changes and precise pixel-by-pixel editing.It supports semantic editing (object rotations, style changes) and appearance editing (pixel-level changes with seamless integration of highlights and shadows). It has bilingual text rendering capabilities (English and Chinese) and is released under a license. Apache 2.0 – completely open source and commercial-friendly, more permissive than Flux.​

Reve-v1 – THE CHINESE SURPRISE​

The Chinese model maintains a good consistency with my original face, the writing is almost correct, and the atmosphere is believable. It may not be perfect, but it did the job well. Why on the podium? It costs a tenth of the competition… Ranking #5 on LMArena for editing and you can see why.​

SeeDream-4 High Res – ANOTHER CHINESE COUP​

Another Chinese model holding its own. Square resolution, convincing similarity, legible text. Lower cost compared to Western flagships and superior performance to most competitors. The dragons are devouring the market.​

The Perfect Prompt Wasted

For those who want to understand where they failed, here is the starting image and the VERY DETAILED prompt I used: photographic specifications (Nikon D5300, 50mm f/1.2L, ISO 400), description of the setting, the subject, the lighting, the text to be written:

“"A hyper-realistic, cinematic mug shot portrait of a man (Critical Identity Lock: attached image) standing against a gritty, stylized police booking wall. The background is a textured concrete wall with faint scuff marks, smudged fingerprints, height lines (imperial measurements), and faded graffiti layered over institutional grey. The subject is framed dead centre, holding a black signboard that reads in bold white letters:

‘'VIOLATION OF MEDIOCRITY STANDARDS'’

He wears a modern black-and-white prison-style outfit: slim-fit striped top or monochrome jumpsuit, edgy and fashion-forward rather than costume-like. The neckline and sleeves have subtle fraying. Clothes are dirty and consumed. Accessories like silver hoops or a worn leather wrist cuff give it a rebellious aesthetic. His expression is confident and unbothered, with a slight smirk — bold, clever, and unashamed. He is bald his head is perfectly shaved. The lighting is stark and moody: single light source from above casting soft shadows under her jaw and behind her, creating depth and mood.

Camera specs for realism and tension:
• Nikon D5300, 50mm f/1.2L lens
• ISO 400, f/2.0 for soft background blur and crisp facial detail
• Studio-style flash with slight overhead diffusion
• Sharpened textures on skin, hair, concrete, and fabric
• Colour-graded for cinematic realism, subtle desaturation for gritty tone“

Everything clear, precise, impossible to misunderstand.

And instead…

Reflections of a Disappointed Criminal

The truth is this: Most AI models have failed miserably. They failed at facial resemblance, at interpreting the prompt, at writing the Italian text. Some got EVERYTHING wrong.

And this, paradoxically, proves exactly the point I wanted to make in my original article on mediocrity: we cannot blindly rely on algorithms. It's not enough to use the most famous or expensive AI. You need critical thinking, you need testing, you need to SEE with your own eyes.​

The lesser-known Chinese models (Qwen 2.5, Reve-v1, SeeDream-4) performed better than Flux Pro and GPT-Image. Google Gemini 2.5 almost hit the mark but failed in spelling. Only the Pro version of Nano Banana proved to be worth the investment and the best in terms of quality.

The True Moral of the Story

The best model for this task was neither Google Premium nor OpenAI. It was Alibaba's Qwen-Image-Edit: open source, permissive commercial license, and superior results.​

While Flux Pro costs a lot and writes “miollisoinia”, while GPT created magnificent images of strangers, Qwen simply did the job. Perfectly.

China isn't just coming to the world of AI. It's already here. And it's winning.​

VIOLATION OF MEDIOCRITY STANDARDS: GUILTY AND PROUD.

(And impressed by Alibaba)

---

PS: Qwen, if you're reading this, you're the best. Period.

PPS: Flux, GPT… have you seen? THIS is how it's done.

PPPS: Alibaba released this monster under the Apache 2.0 license. Open source. Free. And it beats all its paid competitors. Let's think about it.

PPPPPS: All this was done for fun, to raise a smile and poke fun at this technology that can really be useful in so many fields... it's not up to me to describe the implications of all this when it falls into the hands of the darkest part of the human soul... Let's meditate for 2