Responsible AI
What VisionObject is for, what it must not be used for, how uncertain its outputs are, and who is responsible for what. Controls the software enforces are kept apart from the ones that rest on an agreement or on you as the operator.
- 1Monitored zone · Yard and entrance only
- 2Out of view · Neighbouring home not covered
- 3Stated purpose · Site safety and access
- 4Retention limit · Deleted on schedule
Four principles in practice
A purpose for every deployment
What the cameras are for, which events matter and who may search are written down before a pilot starts.
Candidates, then review
A match, a group or a reading is a candidate. We agree with you in writing that a person compares it with its source before anyone acts.
Uncertainty, stated
Every capability lists its known limits, and nothing is described that is not in the release.
Measured on your footage
We publish no accuracy figures. Read rates and false-alarm rates are measured on your cameras in a pilot.
Four different things
A detection is not a match, and a match is not an identity.
Each page keeps these four apart. A result never moves from one to the next because a score crossed a threshold.
Detection
An object was observed in a frame or a sequence of frames: a person, a vehicle, a bag.
Appearance similarity
Two observations resemble one another: by clothing and colour, or by a plate read. Grouping uses appearance only. A plate is searched on its own, and a near-match plate is flagged beside the plate as read. Comparing faces is not offered in the current release.
Reviewed association
A reviewer accepts that two observations belong together, within one investigation and for its purpose. That decision is recorded as a note on the alert.
- Outside the product
Identity verification
Establishing who a person is. That is a separate, specifically authorised process with corroborating evidence, carried out by your organisation. VisionObject does not do it.
Who is responsible for what
Software, agreement, operator.
Three kinds of safeguard, kept apart. Only the first holds whatever anyone agrees.
Enforced by the software
These hold whatever anyone agrees.
- An alert stays New until a named person marks it
- Uploaded footage never raises an alert
- Every read by a person is limited by organisation, role and camera scope
- Changes, and searches by plate, scene description and route, are written to the audit log
- A near-match plate is flagged as a near match
- Sightings that overlap in time are not grouped as one
- A module that is switched off neither runs nor shows what it produced
Agreed in writing with you
Decided before a pilot starts.
- The purpose of the deployment and of each sensitive capability
- That face search stays switched off: it is off in every deployment we deliver, its recognition model is not licensed for production use, an organisation's administrator cannot switch it on, and the change to the deployment's configuration that could is one our agreement does not allow
- Who holds each role, and who may search
- Retention periods for records, stills and uploaded footage
- That a person checks the source before any consequential action
- How sample footage for a pilot is shared, used and deleted
Your responsibility as operator
The software cannot do these for you.
- The lawful basis for the cameras, and the notices people see
- Reviewing alerts, so the queue and the false-alarm rate stay meaningful
- Checking your recorder before acting on a still
- Keeping watchlists accurate, with a reason and an expiry for each entry
- Answering people's requests about their own data
- Where cameras point, and what they should not see
What it is not for
These are not in the product, or are deliberately not offered.
- Emotion or mood
- Intent, dangerousness or likelihood of offending
- Gender, ethnicity or other demographic classification
- Who a person is, from appearance alone
- Automatic enforcement or sanction without a person reviewing the source
- Attendance records, or face enrolment of students
- Replacing a certified safety, fire or medical system
Known limits
We would rather you hear these from us.
- Detection quality drops with poor light, low resolution, steep camera angles and distance.
- Grouping uses overall appearance, such as clothing and colour, so similar-looking people and vehicles can be grouped together.
- Not every visible plate is readable. Small, distant, angled or poorly lit plates are skipped. In our own testing, on night and dawn street footage, only a small share of vehicles produced a usable read, and daytime rates have not been measured.
- An AI scene reading can be wrong or missing and arrives after the sighting. Its instructions and its preferred tag list are written for city street scenes in India.
- Line-crossing, zone-exit and dwell rules, plate watchlist alerts and routes with more than one stop are verified by automated tests only; none has yet been confirmed on a live feed in our own testing.
- A count rule counts everything in a camera's view. It cannot be limited to a drawn zone.
- Colours shift with lighting, and a descriptive tag is not a fact about a person.
- Live sightings are timed by when the frame reached the server, not by the camera's clock.
- A human review does not make a model accurate or fair. It makes a person accountable for the decision.
Questions we are asked first
Does VisionObject identify people?
No. It describes and groups what it sees, such as a person in a red top or a white van, without knowing who anyone is. Face search, which would compare images and return candidates, is not offered in the current release.
Does it make decisions on its own?
No. It raises alerts and groups candidates, and no alert acts on its own. What happens next is decided by a person.
How do you check it works fairly in our environment?
We publish no general accuracy claim. In a pilot, your reviewers label a sample of your own footage and we measure the results against it, including across lighting conditions and camera positions.
Can we restrict a capability to certain people or sites?
Partly. Each module is switched per organisation; a person has one of five fixed roles and can be limited to named cameras. A capability cannot be limited to named people or to one site.
Agree the safeguards before the first camera goes live.
We will work through purpose, sensitive capabilities, access and retention with your policy and legal teams as part of pilot planning.