Unseen test photos
Every model is measured on photos it never saw in training, including photos from different sources, cameras and types of home, category by category.
Our AI
How our AI works, and where it stops
AI Scan uses computer vision to look for visible signs of eight common property problems in ordinary photos. This page explains the process, what the results mean, how we test it and what it can't do.
The pipeline
From photo to verifiable report
Every photo goes through the same four stages, so results are consistent across homes, people and devices.
1
Quality gate
Each photo is checked for focus, exposure, glare and resolution, and that it actually shows part of a property. Poor photos are flagged before analysis, with tips to retake them.
2
Multi-model analysis
Several independently trained computer-vision models score all eight categories at once, because one wall can show damp, mould and peeling paint together. Their scores are combined against calibrated, category-specific thresholds, and the report shows how strongly they agree.
3
Visual explanation
Areas of the photo that most influenced a finding are marked, so a person can see what the AI was looking at and judge it for themselves.
4
Tamper-evident record
The report records when it was made and which model version was used, and gets a unique reference and verification code. Any later change to a shared record shows up.
Reading the results
What the scores mean
Signal strength (0–100) shows how strongly the photo matches the visual pattern of a category, relative to its calibrated threshold. A higher number means a clearer visual signal.
Agreement shows how many of the models flag the same category. High agreement means the finding is consistent; low agreement means a person should look more carefully.
What a score is not
It is not the probability that a defect exists, not a measure of severity, and not a health risk rating. It is a visual indicator to help people decide what to look at next.
Quality
How we test and improve it
Versioned releases
New models are compared against the current release on the same tests before they go live. Every report records its model version, and any release can be rolled back.
People in the loop
Uncertain and disputed results are reviewed by people, and feedback from users and housing professionals feeds into the next round of improvements.
Honest limits
What our AI can't do
- See hidden moisture inside walls, floors or ceilings
- Measure moisture, temperature or air quality
- Find the root cause of a problem, or judge structural safety
- Assess health risk or legal compliance
Results can be wrong. Shadows, low light, wood grain, tile grout, patterned wallpaper and stains that look like defects can all mislead a visual check. Always look at the photos yourself, and ask a qualified professional about health, safety, structural or compliance concerns.
Your photos and our AI
- RentProof photos and records are never used to train our AI
- You can object to AI Scan photos being used for training at any time
- Contributed images are used only with permission
See our privacy notice for details.
What we keep confidential
We don't publish our model designs, training data or internal thresholds. This protects our technology and makes the system harder to game.
Try it with your own photos
Free during beta. Your report is saved to your account.
Start with AI ScanQuestions about our AI?
Housing professionals, researchers and surveyors are welcome to get in touch.
Contact us