The name tells you something about where military AI development is heading. Ukraine calls its combat artificial intelligence platform “Avengers Labs,” and the branding is apt enough for a country that has spent three plus years finding asymmetric technological answers to a larger and supposedly better-resourced enemy. When Britain signed an agreement on 24 August to become the first foreign nation to access that platform, it marked a genuine shift in how Western nations are approaching defence AI development, moving from theoretical capability to systems built and tested in live combat.
The deal is genuinely significant. Prime Minister Andy Burnham and President Volodymyr Zelensky, meeting in Kyiv on Ukraine’s Independence Day, formalised a defence AI partnership that extends the 100-year agreement signed between the two countries in 2025. Under its terms, British universities, researchers and technology companies gain access to a dataset of five million battlefield frames collected by thousands of daylight cameras and infrared sensors deployed across Ukrainian combat zones. The imagery covers tanks, artillery, air defence systems, infantry positions and aerial targets. Ukraine’s military has already used it to build automatic target detection systems capable of identifying 70 per cent of enemy targets in real time. The UK is now first in the queue to learn from what they built.
On paper, this is an extraordinary asset. No exercise, no simulation, no synthetic data generation programme has ever produced anything like it. Three years of high-intensity warfare, fought across a vast front with sophisticated sensors on both sides, has generated a repository of real-world combat imagery that NATO members have been unable to accumulate in decades of lower-intensity operations. The Avengers Labs platform is the closest thing that exists to a live laboratory for military AI development, and Britain has just been handed the keys.
The official framing is accordingly enthusiastic. Burnham described the partnership as bringing together “Ukraine’s unrivalled operational experience with Britain’s world-class AI ecosystem.” The two pilot projects announced alongside the deal, converting fibre-optic cables into AI-enabled sensor networks, and developing low-power AI chips for drones and weapons systems, suggest an ambition that goes beyond simply absorbing Ukrainian data and into genuinely co-developing the next generation of battlefield technology.
The question now is how to build on it.
Where the data has limits
Nobody involved in this deal is likely claiming that the Avengers Labs dataset is a complete picture of modern warfare. It is, by definition, a snapshot of one conflict fought in one country over three years Understanding where those limits lie is the necessary first step to addressing them.
Ukraine’s war has been fought almost entirely across flat or gently rolling steppe, agricultural plains and the industrial lowlands of the eastern Donbas. The terrain is open, the sightlines are long, and the combat has developed tactics suited to that landscape. AI systems trained on this data are consequently very good at finding targets in that environment, under those light conditions, in that climate.
British defence planning, however, spans a considerably wider range of scenarios. Norway’s northern flank, where Article 5 obligations could pull UK forces into high-altitude arctic terrain. Mountain environments in the Caucasus or the Balkans, where elevation, shadow and broken sightlines make long-range optical detection a fundamentally different problem. Potential deployments where the threat profile and equipment encountered bear little resemblance to the Ukrainian steppe. The Avengers Labs data is built for one theatre, and AI models carry the assumptions of the data they were trained on.
The technical specifics matter here. Optical and infrared sensors behave differently across different ground. In mountain terrain, targets are masked by ridgelines and vegetation at angles that flat-terrain algorithms have not encountered. A tank hull-down behind a rocky outcrop presents a completely different visual and thermal signature to the same vehicle advancing across open farmland. Arctic conditions alter infrared signatures dramatically: snow and ice change how vehicles present thermally, extreme cold affects sensor performance, and the extended darkness of polar winters produces lighting conditions that even Ukrainian winter data does not represent.
The specific equipment in the dataset creates a further consideration. The Avengers Labs imagery is heavily weighted towards Soviet-designed armour in Russian service, alongside the mixture of NATO and Soviet kit fielded by Ukraine. That is directly useful for any scenario involving Russia, and Russian equipment remains relevant across a broad arc of potential conflict from the Baltic to the Barents Sea. For other theatres and other adversaries, the training set would need to be widened.
But let’s face it, Russia is the main issue we face in the short term.
Filling the gaps
The more interesting question is what addressing those gaps actually looks like, and it is here that the partnership’s longer-term potential becomes clearer.
The most straightforward route is transfer learning: using the Avengers Labs dataset as a foundation and fine-tuning AI models on smaller supplementary datasets gathered in other environments. The underlying architecture and learned features from Ukrainian combat data are not wasted in a mountain or arctic scenario; they provide a starting point that synthetic data alone could not. The task becomes generating targeted supplementary data rather than building from scratch.
Exercise programmes offer one avenue. NATO’s northern European exercises, conducted annually in Norway and Finland, already involve the kind of terrain and equipment that would fill some of the gap. Instrumenting those exercises with the same density of sensor coverage that Ukraine has deployed in combat would generate genuinely useful complementary data, though at nothing like the volume or variety that three years of live warfare produces.
Allied data sharing is another. Nordic nations have extensive experience of arctic operations and the sensor behaviour that goes with them. A broader multilateral approach to AI training data, perhaps under a NATO framework, could assemble a far more comprehensive dataset than any single conflict or exercise programme can provide.
The pilot project on fibre-optic sensor networks points in a useful direction too. If that technology can generate dense sensor coverage of fixed areas, it could be deployed in training environments to build datasets for terrain types not represented in the Ukrainian data. The infrastructure developed for one purpose may prove useful for the other.
None of this is a criticism of the current deal. It is the logic that follows from it. The Avengers Labs partnership is the beginning of a data strategy, not the end of one.
The ‘gamification’ debate
Some critics have seized on the cultural dimension of Ukraine’s approach to drone warfare, particularly the use of points systems and gaming aesthetics, as evidence of a troubling trivialisation of combat. The concern is that presenting warfare through the vocabulary of video games creates psychological distance from the reality of killing, and that Britain risks importing that mindset alongside the data.
It is a critique worth noting, but it sits awkwardly with the evidence. Ukraine’s drone operators have developed some of the most effective battlefield AI integration seen in modern warfare, under conditions of genuine existential pressure. The gamified elements of their approach, the points systems, the game controllers, the social media footage, are the cultural tools of a generation that grew up with them, applied to a problem that is lethally serious. That they look unfamiliar to some more ‘traditional’ observers does not make them less effective. By most measures, they are working.
The legitimate version of the concern is about keeping the human-in-the-loop standards as AI moves further into targeting decisions. That is a real question for British doctrine, and it applies to any AI-enabled system regardless of its cultural packaging. It is a governance and legal question, and it deserves to be treated as one, separately from the aesthetics of how Ukrainian operators have chosen to motivate themselves through a grinding and costly war.
What Britain has bought
What Britain has secured is a foundation of real-world data and engineering experience that would take years to replicate through any other means. The Avengers Labs dataset is imperfect in the ways that any single-theatre dataset is imperfect, and the work of broadening it is substantial. But the alternative, building AI targeting capability purely from simulation and exercise data, is demonstrably less effective.
The pilot projects in low-power chips and fibre-optic sensor networks suggest that the partnership is oriented towards capability development rather than a simple data transfer, which is the rihgt framing. What Britain does next, in terms of building supplementary datasets, establishing the governance frameworks for AI-enabled targeting, and investing in the research infrastructure to make use of what Ukraine has provided, will determine whether this agreement delivers lasting capability or remains a headline.
Ukraine has offered something earned at extraordinary cost. The obligation now is to use it well.
