Artificial intelligence is becoming more deeply embedded in video surveillance, improving visibility in difficult conditions and helping security teams identify events faster. But as cameras and analytics become more capable, manufacturers face a growing challenge: ensuring that AI does not undermine trust in the footage it is supposed to help interpret.
Two concerns are emerging as particularly important for security buyers. The first is whether AI image enhancement can alter footage in ways that affect its evidentiary value. The second is whether video collected by customers can later be used to train AI systems without their knowledge or approval.
For surveillance technology providers, addressing both issues increasingly depends on separating what AI can do from what organizations should allow it to do.
Preserving the original footage
Callum Wilson, co-founder and CEO of Eluviant, said AI-enhanced video can have a legitimate role in law enforcement and private surveillance, provided organizations preserve a clear distinction between the original recording and any version modified for easier interpretation.
“There is absolutely a role for AI-enhanced footage in both law enforcement and private surveillance, as long as the fundamentals are respected,” Wilson said.
The most important safeguard, he said, is maintaining the original footage without modification.
“The original footage should always be preserved unchanged and kept separate from any enhanced version, and enhanced footage should be clearly labeled as such,” Wilson said.
That distinction becomes increasingly important as AI systems move beyond traditional image processing. Surveillance cameras have long adjusted exposure, noise, contrast and other image characteristics. AI-based enhancement can potentially go further, making difficult scenes easier for operators to interpret.
For investigators, however, improved visibility and evidentiary integrity are not necessarily the same thing.
Wilson said enhancement should function as an aid to interpretation rather than replace the underlying recording.
“Enhancement is there to help a human, increasingly assisted by AI, interpret a scene, while the unaltered recording remains the evidence,” he said.
Why provenance matters
This approach places greater emphasis on provenance. Security teams may increasingly need to know not only when and where footage was captured, but also whether what they are viewing is the original recording, a conventionally processed image or an AI-enhanced derivative.
For systems integrators and enterprise security teams, the issue could become especially relevant when surveillance footage moves between operational monitoring and investigation.
Video enhanced to help an operator understand an incident may later become part of a formal inquiry. Clear separation between original and processed footage therefore becomes important.
Eluviant's approach is focused on understanding events in live video rather than reconstructing imagery, Wilson said. Its AI works with video produced by an existing camera estate, including footage that may already have been processed by cameras designed for low-visibility environments.
“Our own AI works with whatever video the camera estate produces, and that already includes processed images from cameras optimized for low-visibility conditions,” he said.
Wilson said the objective remains interpreting activity occurring in the scene rather than changing the underlying record of the event.
Customer footage and AI training
The integrity of the video itself is only one part of the emerging AI governance problem.
Organizations are also becoming more concerned about what happens to their surveillance footage after it enters an AI-enabled security platform.
Video surveillance systems can collect unusually sensitive information. Cameras may record employees, customers, visitors, building layouts and patterns of activity over extended periods.
As AI vendors seek larger datasets to improve their models, customers may question whether footage collected for security purposes could later become training material.
Laurent Villeneuve, senior manager for product and industry marketing at Genetec, said customers retain control over their data and that the company does not automatically use customer surveillance video for AI development.
“Genetec does not use customer video to train AI models without the customer’s explicit authorization,” Villeneuve said.
According to Villeneuve, customer data is processed to provide the functions that organizations choose to deploy, including security, investigation and automation capabilities.
That processing takes place within the use case and deployment model configured by the customer, he said.
A broader data governance question
The distinction matters because AI-enabled surveillance does not necessarily require that customer footage be absorbed into a vendor's general-purpose training datasets.
Analytics can process video to perform a particular function without giving the provider unrestricted rights to reuse the footage for model development.
For enterprise customers, that creates a procurement question that extends beyond whether a product includes AI.
Buyers may also need to understand where video is processed, how long it is retained, who can access it and whether the same data can be reused for purposes beyond the security operation for which it was originally collected.
Villeneuve said Genetec applies guidelines covering the creation, improvement and maintenance of its AI models, including controls around privacy, data governance, transparency and safety.
Those controls include limiting access to datasets and testing models to reduce bias, he said.
Keeping humans in control
Human oversight also remains part of Genetec's approach.
“Human oversight is also maintained so that people retain control and make critical decisions,” Villeneuve said.
That principle is becoming more significant as surveillance systems move from passive recording toward automated interpretation.
AI can help identify objects, behaviors or events across large volumes of video that security staff could not feasibly review manually. Yet the more responsibility delegated to software, the more important it becomes to understand what the system is doing with both the images and the conclusions drawn from them.
For customers, the challenge is therefore not simply deciding whether AI analytics are accurate enough to deploy.
They must also determine whether the technology preserves a reliable record of events and whether data collected for security remains governed according to the customer's expectations.
Trust becomes part of the AI equation
The comments from Eluviant and Genetec point to safeguards that may become increasingly important as AI adoption in surveillance expands.
Preserving original footage, identifying enhanced material, restricting the secondary use of customer data and keeping people responsible for consequential decisions can all contribute to maintaining confidence in AI-enabled surveillance.
AI may make surveillance video easier to interpret, particularly when large camera estates or difficult visual conditions overwhelm human operators.
But its value will depend in part on whether organizations can still establish what the
camera originally recorded, how that footage was processed, and who retained control of the data after capture.