Until recently, one thing quietly constrained what organisations could do with video: it was difficult to search. Finding a single moment across hours of CCTV, body-worn or dashcam footage usually meant somebody watching a lot of it, or running analytics built to find something relatively specific.
That friction is disappearing. Video-language models increasingly let users search with ordinary descriptions, an example image, a vehicle, a location or a pattern of behaviour.
The biggest change is not speed but scope. The same footage can yield far more information without a single new camera being installed.
From “who is this?” to “what can be found?”
The privacy debate around cameras has largely focused on identification – can this system tell me who somebody is?
Facial recognition is controversial, but its basic task is to compare a face against a defined reference or watchlist. The boundaries can be explicit – a particular list, location, period and retention policy.
Visual search raises a different question: what can somebody discover about people without knowing who they are to begin with?
WIRED’s recent investigation into Flock Safety’s OS Investigate shows how that capability is being designed for police use. It found 69 pre-built prompts available to officers. Of those, 19 were designed to find patterns rather than retrieve a known record, and 14 reportedly needed no number plate, name or description of a suspect at all. A location, timeframe and behaviour could be enough to start searching.
The version WIRED examined was still in testing and was reconstructed from code served by Flock’s own login pages. The example nonetheless shows the shift from retrieving a known record to open-ended discovery.
Open-ended visual search is much less tidy than a watchlist. A user – whether in policing, retail, transport, automotive or another sector – might look for a person in a red jacket near a particular entrance, every white van appearing at several locations, or people following a pattern of movement over several days. Depending on the context, each could be a perfectly legitimate search.
The difference is that the technology no longer defines the question in advance. The user does.
Why access controls are no longer enough
Two authorised users can have access to exactly the same footage and use it in fundamentally different ways.
One searches a specific location and time because an incident has been reported. Another searches weeks of footage to build a picture of one person’s movements.
Even where footage was collected for a legitimate purpose and the user is authorised to access it, a search can take its use far beyond the original reason it was captured.
Access alone cannot tell those cases apart. Organisations also need to govern what people are using the system to find.
A natural-language interface makes that power feel routine. Typing a sentence into a box feels more like using Google than operating specialist visual-analysis technology. Behind that box may sit months or years of information about people’s locations, behaviour and interactions.
The search history may matter as much as the footage
Once search becomes open-ended, the query needs scrutiny too.
A query can contain personal or sensitive information. Its wording can reveal assumptions about a person or behaviour. A sequence of searches can show how an investigation developed and why particular people or events came into focus.
In recent discussions I have had with criminal lawyers about visual search, we have discussed whether search terms themselves should be retained – and whether that history could ultimately need to be disclosed in a case, showing how investigators arrived at the evidence they selected.
If AI filters thousands of hours of footage down to the handful of clips an investigator sees, the record of what they asked the system to find becomes part of the chain of accountability.
Once the search history can explain how evidence was found, it stops being a technical log and becomes part of the accountability record. An audit trail needs to show more than whether somebody logged in. It should make it possible to reconstruct what was searched, when, across which data and what happened next.
The same principle applies outside criminal justice. If a business uses visual search to investigate theft, a safety incident, an insurance claim or an operational problem, it should be able to show how that search was conducted and why.
Permission, purpose and proof
For organisations deploying visual search, the controls fall into three areas: permission, purpose and proof.
Permission needs to be granular. Access should reflect the datasets, locations and time periods somebody genuinely needs, rather than a single decision about whether they are an authorised user.
Purpose needs to be explicit. A search should have a defensible connection to the reason the footage is being used, rather than becoming an invitation to explore whatever the system can surface.
Proof needs to be retained. Search creates a new layer of activity that should be auditable: who searched, what they asked for, which results they opened and what happened next. Retention policies also need to account for queries, results and any collections created through search, not just the original footage.
The words in the query are part of that record. A search can encode an assumption about who or what is suspicious, while the results can pull in people who were never relevant to the original event.
More insight without more cameras
Privacy debates often focus on expansion – more cameras, longer retention, more sensors.
AI breaks the assumption that greater surveillance capability requires more collection. The camera network can stay exactly the same while what an organisation can learn from it expands.
A network that once required hours of manual review can become searchable across locations, time periods and behaviours. The cameras have not multiplied, but what can be extracted from their footage has.
Visual data makes that particularly acute because cameras do not capture only the subject of interest. Footage of one incident may also contain colleagues, customers, children, homes, screens, number plates, private conversations or people simply passing through.
They may be irrelevant to the incident. Their information is not.
Making footage more searchable can make that surrounding information easier to discover and combine too.
Visual search has obvious value: finding critical evidence quickly, tracing an incident across several cameras or making huge archives operationally usable. The challenge is ensuring that the ability to find one useful thing does not quietly become permission to discover everything else.
