AI orca detection sounds like a simple task until a whale surfaces in poor light and disappears almost immediately. A partnership announced in October 2026 aims to improve the ability of maritime cameras to recognise marine mammals by collecting footage in the real conditions that make them difficult to find. Technology company SEA.AI is working with conservation organisation WeWhale, whose teams spend time in the animals’ habitats.
The SEA.AI partnership briefing, published on 5 October, explains that its existing systems use cameras and computer vision to detect objects near vessels. The new collaboration focuses on gathering the images and thermal observations needed to train better marine-mammal classification. It is an active data-collection effort, not evidence that a finished autonomous warning system can already prevent collisions or every dangerous encounter.
Why AI orca detection needs real sightings
A machine-learning system needs examples of what it is supposed to recognise. Cars, packages and manufactured components can often be photographed repeatedly under controlled conditions. Wild marine mammals cannot be told where and when to surface. They may appear briefly against waves, spray, sun glare or darkness, and their visible shape can change dramatically with distance and movement.
SEA.AI explains that it can stage some maritime detection scenarios. Staff can enter the water safely in controlled exercises or deploy objects that resemble floating hazards, producing images in different sea states. It cannot reproduce a genuinely free-swimming whale on demand. That makes encounters observed by conservation teams unusually valuable training material.
The partnership brings together complementary activities. WeWhale has field experience monitoring cetaceans and working with boat users. SEA.AI supplies maritime sensing technology and the ability to turn recorded examples into inputs for its visual classification system. Each sighting can add evidence about how animals appear in the conditions crews actually face.
The difficult technical problem is not simply identifying an obvious photograph of an orca. It is distinguishing an animal from wakes, birds, floating equipment and other confusing shapes early enough for a person on board to make a sensible decision.
A small and vulnerable population raises the stakes
The conservation focus includes Iberian orcas in waters around the Strait of Gibraltar. The Save the Iberian Orca campaign says fewer than 30 individuals remain in this threatened population. That figure is the campaign’s published estimate and should be treated as such, rather than an independently measured count taken by LiveAIWire.
These animals use waters shared with fishing boats, leisure craft and commercial traffic. Encounters can produce anxiety for boat users and risks for wildlife. A warning system that improves awareness could be useful, but its value would depend on whether it gives crews enough time and sufficiently reliable information to respond appropriately.
The conservation group describes field monitoring, collecting information about orca distribution and providing navigation guidance. This is an important complement to the technology. A camera can help bring an animal to someone’s attention, but it cannot on its own settle which course of action is safe, lawful or appropriate in a particular situation.
Avoiding a harmful encounter also requires clear instructions for people operating vessels. The broader objective is coexistence rather than simply placing more cameras at sea.
A detection and a species identification are not the same
Machine vision can flag that something unusual appears on the water, but recognising its precise identity is another challenge. A floating object might be a log, a person, an animal or an artefact of light and waves. Incorrect alerts are frustrating; missed detections may be more serious. A useful onboard system has to balance both errors.
The SEA.AI collaboration is specifically intended to improve the examples used to classify marine mammals. Its briefing describes cameras collecting normal and thermal imagery during real field activity. Thermal information can highlight heat differences, while visible-light footage may reveal shape and movement. Neither input guarantees an accurate classification in every condition.
That distinction matters when headlines suggest that AI has already solved wildlife monitoring. AI animal identification in other settings may work well on carefully selected photographs, but performance can change when the location, camera angle or species differs from training material. Robust field testing therefore needs to include ordinary difficult cases, not just impressive demonstrations.
For boat crews, the ideal interface would be clear about uncertainty. An alert could help them look towards an area of concern and use established guidance. It should not invite the assumption that every absence of an alert means the water ahead is clear.
Field data may be more valuable than a bigger model
Much attention in AI development goes to the size or apparent cleverness of a model. For a maritime warning system, one of the important constraints is the quality of examples. If the system has never seen an animal under a particular combination of sea state, lighting and distance, it may struggle to make sense of the image even if its general vision abilities are strong.
The collaboration illustrates why specialist knowledge remains useful. People who understand the animals can tell engineers whether a recorded shape really represents a whale, what behaviour preceded the sighting and which conditions made the event unusual. Accurate labels are essential when the same model must distinguish several kinds of floating object.
Lessons from robot learning from video offer a useful comparison. Watching a real activity provides information that artificial examples may miss, but it also introduces noise and variation. Researchers need to decide what the model should learn from that variation and how to assess it on footage not used during training.
The resulting dataset may become more useful as it covers different seasons, areas and conditions. Its value cannot be measured simply by the number of videos collected. A carefully described rare encounter might reveal a weakness that thousands of nearly identical daytime images would not expose.
Who should respond when the camera spots something?
There is an operational question beyond recognition. If cameras detect a possible marine mammal, what does the crew actually do? The correct response may depend on water conditions, speed, traffic, regulations and the latest guidance for the area. An AI alert can inform that decision without assuming authority over the helm.
Autonomous technology and safety is an important comparison because the cost of a false assumption rises when software moves from sensing to acting. On a vessel, a poor automatic manoeuvre could create a separate danger. It is therefore sensible to distinguish early-warning assistance from systems that physically control movement.
SEA.AI’s announcement supports claims about a partnership and data collection. It does not establish a published independent benchmark for orca-detection accuracy, field sensitivity or collision reduction. Those would be the measurements needed before declaring the approach proven for this purpose.
Even without them, the research direction is understandable. The harder task is gathering enough trustworthy examples from the actual environment. Better recognition can only be achieved if researchers first have something genuine to learn from.
The next milestone is reliable warning at sea
For this project, success should eventually mean more than a model correctly labelling attractive wildlife footage. Crews would need useful warnings in rough conditions, at realistic distances, with false alarms low enough that people do not begin ignoring the system. Conservation groups would also want evidence that better awareness translates into safer behaviour.
Those outcomes require field evaluation and transparent reporting. It would be useful to know how the system performs by species, lighting conditions, weather and vessel type, and whether its accuracy changes as new footage is incorporated. Independent testing would make it easier to separate a promising partnership from verified conservation benefits.
There is a wider opportunity here for technology that helps human observers without displacing their judgement. The people best equipped to recognise changes in animal behaviour may be the same people who can help improve an algorithm. When engineers and conservation teams cooperate, the technology can become more grounded in the world it is supposed to serve.
For now, the central development is the creation of a better real-world evidence base. The ocean has not suddenly become predictable. The researchers are trying to teach their system how to work with that uncertainty.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.