I explore in this post how the use of AI-based Decision Support Systems (AI-DSS) could disrupt the three criteria developed by Alexander Wentker for identifying co-parties to an armed conflict. I first set out Wentker’s criteria and then define AI-DSS and how it may map (or not) onto the criteria.
Three criteria for co-party status
Wentker’s criteria are, in my view, a significant improvement on pre-existing approaches because they are fairly general, simple and based on an objective assessment on the facts.
First, the relevant conduct of the individual must be attributable to the collective entity, applying the law as reflects in the ILC Articles.
Second, the acts of the purported co-party are directly connected to the hostilities. He observes that sharing military intelligence in real time and in relation to a concrete military operation or air to air refuelling of fighter jets carrying out strikes are examples of a direct connection. By contrast provision of weapons or other supplies, provision of finance or political support as part the general war effort would not satisfy the criterion of direct connection.
Third, there is some degree of cooperation or coordination between the purported co-party and the other party that they support. Wentker suggests that it suffices for State A to coordinate with State B and then for State B to coordinate with States C and D in order for States A, B, C, and D to be co-parties. The litmus test is whether the purported co-party is involved in the decision-making processes as to whether and how the coordinated military operations are taking place.
These criteria presuppose that a co-party acts with knowledge of the facts that establish the direct connection to hostilities and the element of cooperation or coordination. According to Wentker, knowledge means awareness of the essential factual patterns constituting the coordinated military operations, that they are directly related to the harm to the adversary, and the circumstances enabling involvement in decision-making. This is objectively inferred from factual patterns. He explains that Five Eyes intelligence alliance would not, without more, be sufficient to meet the standard of knowledge whereas the provision of intelligence for concrete military operations would suffice, such as nominations of persons for a targeting list.
AI-DSS in military operations
Militaries are incorporating increasingly complex forms of AI-DSS into their procedures. By displaying, synthesising or analysing information, AI can provide complex assessment and nuanced outputs to aid humans in deciding who or what to attack and where, when and how. Even though AI-DSS do not “make” decisions, they directly influence the decisions of humans (see further). Examples of potential uses of AI-DSS are set out in the figure below (source).
After some hesitation about cooperating in the military applications of AI, some tech companies are now keen to work with certain national militaries.
Meta previously prohibited the use of its open-source large language model (Llama) for “military, warfare, nuclear industries or applications, [and] espionage”. However, in November 2024, it announced it would allow the use of Llama by US national security agencies and defence contractors as well as national security agencies in the UK, Canada, Australia and New Zealand.
Anthropic’s Claude 3 and Claude 3.5 models will be used by Palantir, a defence contractor, to sift through secret government data. OpenAI, recently hired former Palantir’s Chief Information Security Officer and appointed a retired US Army General to its board of directors.
We are likely to see an exponential increase in AI-DSS development and deployment in the coming period.
Mapping onto the three criteria
AI-DSS poses a challenge to the first criterion of being able to attribute the relevant conduct to the collective entity because it obscures the control that humans exercise over, for example, targeting decisions. This in turn complicates the Article 8 ARSIWA analysis as to whether there have been “instructions” issued, or “effective control” exercised, by the collective entity.
These concerns may be illustrated by the use of AI-DSS by Israel. “Lavender” is an AI-based programme designed to mark all suspected operatives in the military wings of Hamas and Palestinian Islamic Jihad (PIJ), including low-ranking ones, as potential bombing targets. Lavender analyses information collected on most of the 2.3 million residents of the Gaza Strip through a system of mass surveillance, then assesses and ranks the likelihood that each particular person is active in the military wing of Hamas or PIJ. In the first weeks of the war, the system marked up to 37,000 Palestinians as suspected militants and identified their homes as targets for possible air strikes. Outputs of Lavender are reportedly treated “as if [they] were human decisions” (but also see the view that humans still make the critical decisions). Reports are that Lavender errs in 10% of cases of target identification.
Another AI-based programme, “Hasbora” (or “The Gospel”) compiles and cross-references information from different datasets in order to “generate” targets at a rapid rate It has been said to facilitate a “mass assassination factory” in which the “emphasis is on quantity and not quality”. The IDF emphasised that Gospel does not pick targets and its suggestions are an insufficient basis for concluding that an objective is a lawful target. The system apparently does not assess potential collateral damage for a proportionality analysis nor identify viable precautions—compliance with these rules is implemented separately during the stages that follow target identification.
Even if AI-DSS like Lavender and The Gospel are meant to be used as ‘human in the loop’ systems (meaning that the recommendation about who or what to target is sent to a human decision-maker for review and final decision), the speed and scale of generation or nomination of targets and the complexity of data processing “may make human judgment impossible or, de facto, meaningless”. Moreover, research indicates that in situations involving high cognitive complexity, pressure, stress, and time constraints, humans are more likely to defer to the judgment of AI. Bringing this back to Wentker’s first criterion: how can ARSIWA apply to attribute the conduct of AI-DSS to a collective entity?
Wentker’s second criterion is that the acts of the purported co-party must be directly connected to the hostilities. The supply of real-time targeting intelligence would constitute a direct connection whereas mere supply of weapons would not. Would the supply of AI-DSS by State A to State B constitute a direct connection to the hostilities?
The Gospel and Lavender can immediately make available information on the status of a potential target. AI DSS may assist with monitoring the battlefield, including predicting the behaviour and reactions of other actors. Existing military research projects are developing AI DSS that allow users to examine the interconnectedness of objects, to model the interiors of buildings, and to assess the capabilities of friendly or adversary forces. On Wentker’s analysis, it would appear that supply of such AI-DSS by State A to State B would make State A co-party, given the powerful insights that AI-DSS can provide on the battlefield. This blurs the line between the supply of weapons (an insufficient connection, according to Wentker) and the supply of real-time targeting intelligence (a sufficient connection).
The third criterion of a degree of coordination between the purported co-party and the other party that they support is also complicated by AI-DSS. There will likely be less need for coordination between States, at least on targeting decisions, if decisions are being largely managed by AI-DSS. The potential co-party is also less likely to be aware of circumstances of decision-making if AI-DSS is being deployed. Indicia of coordination that Wentker points to in the book, such as coalition command/joint operations centres may become less common or meaningful in terms of how a military campaign is being conducted on the ground.
Wentker’s three criteria for co-party status is understandably based on a vision of humans from different states or collective entities sharing real-time intelligence and meeting in a coalition command HQ. My concern is that this vision may soon be replaced by the increasing delegation of military decision-making to AI. That has implications for the conduct of hostilities far beyond the identification of co-party status…
As regards co-party status, the use of AI-DSS may make attribution of conduct more difficult and the existence of coordinated action on targeting less common. On the other hand, the supply of AI-DSS by one State to another may well constitute a direct connection to hostilities. Overall, AI-DSS may call for a reconsideration of the criteria for co-party status and I look forward to Alexander Wentker’s thoughts in this regard.
