Why Context Technology Who We Build For Research Status Roadmap Company Blog Request Demo
PurenetX Research

We Said Pixels Aren't Enough. So We Tested It.

In July we argued that pixel-based safety systems block legitimate moments. This month we built an on-device prototype and tested the claim on 63 everyday photos — including our own mistakes.

Research
October 27, 2026 · 5 min read · PurenetX

In July we wrote that most AI vision systems are context-blind: they see skin, not situations. A gym selfie, a clinic visit and a wedding can all look "risky" to a system that only counts pixels.

That was an argument. This month we wanted evidence.

What we built

We built an early prototype of PurenetX that runs entirely on a laptop — no cloud, no upload, no internet needed. It works in three steps:

  1. An open-source vision model describes the scene: where it is, what is happening, who is present, how much skin is exposed, whether anyone seems unaware of the camera.
  2. Our decision engine reads those signals through our six context dimensions.
  3. It returns ALLOW, WARN, REQUIRE CONFIRMATION or BLOCK — with a one-line reason.

PurenetX prototype demo showing context analysis of a yoga photo

Our prototype, running locally on a laptop. A pixel-only approach rates this yoga photo CRITICAL risk; the context engine allows it, with a reason.

The test

We collected 63 everyday photos that should all be allowed: gyms, yoga studios, swimming pools, beaches, clinics, offices, weddings, festivals, family homes, food, streets. Some are free stock photos; ten are ordinary phone photos we took ourselves, in real Indian homes. We labelled every photo before testing.

Then we ran the same 63 photos through two systems:

  • An earlier prototype of ours that judged images by counting skin-coloured pixels — the approach we have been criticising.
  • The new context engine.

What happened

The skin-colour approach rated 27 of the 63 safe photos as high or critical risk. The context engine flagged none of them at the blocking level.

The failures of the pixel approach were almost comic once we looked at them. It rated an empty hospital corridor, a CCTV camera on a wall, a plate of food and an empty hotel room as high "nudity risk" — because beige walls, wood and food look like skin to a colour counter. On our own phone photos alone, it flagged 5 of 10. The context engine flagged 0.

The screenshot above is the clearest example: a man meditating in a yoga studio. Pixel approach: CRITICAL risk. Context engine: ALLOW — "minimal exposure appropriate to the context (fitness)."

Same pixels. Different understanding.

What we got wrong (and fixed)

Testing honestly means reporting our own mistakes.

We found a gender bias. Early on, the system judged several adult women in sportswear and swimwear to be children — and never did this for men in similar clothing. That made it warn unfairly on women's gym and pool photos. We changed how age is assessed and added a second, focused check. On new photos the system had never seen, all adult women were correctly judged as adults — and photos with real children were still correctly identified.

We threw out some of our own test photos. Some stock photos we used to test "secret filming" showed actors in hats and trench coats peeking around corners — a film-poster idea of a spy. Real secret recording looks nothing like that, so a system could pass that test for the wrong reason. We removed them. We also caught a photo we had labelled "bathroom" from its search term, when it actually showed a bedroom. We now label from the photo itself, never from the search word.

What this does not prove

We want to be precise:

  • This test measures false alarms on safe photos. It does not yet show how well PurenetX catches genuinely harmful situations. That testing is in progress, with realistic staged photos taken with consent.
  • The comparison is against our own earlier method, not against commercial tools.
  • 63 photos is a small set, and some of them helped us tune the system. On photos it had not seen during design, it made the right decision on 33 of 38.
  • At the time of this test (earlier prototype version), it took around 20 seconds per photo on a laptop. Running fast on a phone is our next engineering milestone.

Why it matters

Every wrongly blocked gym photo, clinic visit or wedding moment teaches people that safety tools are against them — and then they switch those tools off. Systems that people switch off protect nobody.

Context is how humans decide what is okay. This month showed, in a small but real way, that it works for machines too.

Next, we test the harder half: whether PurenetX catches what it should.

— Kaushik A., Founder & CEO, PurenetX

← All posts