The Israeli Defence-AI Deployment Index records, for each of 50 companies, the strongest published evidence that its system has been trialled, bought or used. It stops there. No entry connects a product to a specific named incident, operation or casualty event, and no entry will, even where press reporting has drawn such a connection. This is a standing methodological rule, and in a field where defence AI accountability reporting is contested it needs a stated rationale rather than a silent boundary.
What the rule permits and what it forbids
The rule is narrow and mechanical. An entry may record that a named outlet reported a military using a system in operations. That is the L3 threshold and thirteen entries meet it. Globes reported the IDF using Axon Vision's AI against drones. Janes reported the IDF using UVision's mini loitering munition. Breaking Defense reported the US Army expanding its use of SMARTSHOOTER's fire-control system.
Each entry records the fact of reported use and the URL. It does not extend to where, when, against what, or with what result. The grading definitions contain no field for any of those things, so there is nowhere in the data model for an attribution to be stored even if one were accepted.
The same rule applies in the other direction. The index does not infer which military unit a founder served in, and does not treat a company's own account of where its product has been used as establishing anything beyond the fact that the company said it. Sightec's index source is a release published through an accelerator describing over 3,000 "battle-proven deployments" of its navigation technology. The entry is graded L0 because no independent public source corroborates a defence deployment. That phrasing, treated as evidence, would function as an attribution claim without any of the sourcing an attribution requires; the broader pattern is examined in what "battle-tested" means when nobody publishes the test.
First reason: the verification standard is not met
Linking a specific product to a specific incident is an evidentiary exercise with an established methodology. The Office of the UN High Commissioner for Human Rights and the Human Rights Center at UC Berkeley publish the Berkeley Protocol on Digital Open Source Investigations, which sets international standards for using digital open-source information to investigate alleged violations of international criminal, human rights and humanitarian law. It specifies procedures for collection, preservation, authentication and verification.
This index does none of that. It reads published articles and records what they say. It does not authenticate imagery, establish chains of custody, corroborate across independent collections, or assess the reliability of underlying witnesses. An index built on secondary reading cannot produce a finding of the kind the Protocol governs, and asserting one anyway would borrow authority it has not earned.
Second reason: presence is not causation
Even a well-sourced report that a system was present, fielded or in service at a given time does not establish that it was the system used in a given event, still less that its output determined an outcome.
Modern targeting chains involve multiple sensors, several software layers and human decision points whose sequencing is not public. The International Committee of the Red Cross has argued that AI decision-support systems in military decision-making challenge the human's ability to exercise judgement, and that these systems should support rather than replace human decisions. That argument depends on facts about where a human sat in a particular loop. Those facts are not in any public source this index reads. The vocabulary problem alone is significant enough to warrant its own treatment in in the loop, on the loop, out of the loop.
A supplier index that asserted causation would be asserting a chain it has not traced.
Third reason: an index should not become an accountability document by accident
The index is a procurement-facing research artefact. It is read by buyers, analysts and journalists trying to establish whether a supplier's product has left the demonstration tent, which is the question set out in the deployment evidence gap.
If entries carried incident attributions, the document would function as a register of alleged responsibility while remaining sourced as a market survey. Companies would face reputational consequences generated by a research method never designed to carry them, and would have no proportionate route of reply. The index's correction policy — send a published source that contradicts an entry and the entry is amended or removed — works for a claim about a contract announcement. It does not work for a claim about an incident.
The case for attribution, stated fairly
There is a serious argument on the other side, and it is not made by fringe voices.
In 2025 the UN General Assembly adopted a resolution on lethal autonomous weapons systems by 164 votes to six with seven abstentions, stressing the role of humans in the use of force to ensure responsibility and accountability. Human Rights Watch has called for the start of negotiations on a treaty addressing these systems. Campaign groups have recorded that a majority of state submissions to the UN Secretary-General's report expressed support for a legally binding instrument.
The position underlying that work is straightforward: accountability for the use of force requires knowing which systems were involved, who supplied them and what they did. On that view, supplier-level indices that stop at "reported in use" are precisely the evasion the field cannot afford, because they document the commercial layer while leaving the consequential layer unexamined. Researchers and human-rights reporters who pursue attribution argue that if nobody links products to outcomes, no supplier is ever answerable for one.
This index does not adjudicate between that position and its own. The two are doing different jobs with different evidentiary equipment. Attribution work requires investigative methods, source protection and legal review that a published-sources index does not have and does not claim. What the index can offer that work is a clean, checkable, dated record of which suppliers have publicly reported military relationships at all, including across the autonomous air systems category where the question bites hardest.
Limits of this reading
Applying the rule has a cost, and it should be named. Readers looking for consequences will not find them here, and the index's silence on those questions is not neutrality about whether they matter. It is a statement that this document cannot answer them.
The rule also means the index may understate what is publicly known. Where investigative reporting has established a connection to an operational context, the index records only the underlying fact of use and drops the rest. A reader who wants the fuller picture should go to that reporting directly. Related boundaries are set out in what a deployment index cannot tell you, alongside the full index.
Frequently asked questions
Does the index say which systems were used in specific military operations?
No. Entries record that a named outlet reported a system in operational use by a military. They do not record where, when, against what, or with what result, and no entry links a product to a specific incident, operation or casualty event.
Why not, if journalists have already reported those links?
Because this index reads published articles rather than conducting verification to open-source investigation standards, and because a report that a system was in service does not establish that it was used in a given event or that it caused an outcome.
Does that make the index a defence of the companies listed?
No. The rule applies symmetrically. It also prevents the index accepting a company's own claim about where its product has been used, which is why several entries with far-reaching self-descriptions are graded L0.
Sources
- idf.ai, The Israeli Defense-AI Deployment Index v1.0, 24 August 2026
- OHCHR, Berkeley Protocol on Digital Open Source Investigations
- United Nations Meetings Coverage, General Assembly adopts more than 60 resolutions and decisions of its First Committee, 2025
- Human Rights Watch, UN: Start talks on treaty to ban 'killer robots', 21 May 2025
- ICRC Humanitarian Law & Policy blog, AI in military decision-making: supporting humans, not replacing them, 29 August 2024
- Janes, IDF using UVision mini loitering munition
- Starburst, Sightec announces over 3,000 deployments of its GPS-free visual navigation technology
Independent publication of idf.ai. Not affiliated with, endorsed by, or connected to the Israel Defense Forces, the Israeli Ministry of Defense, or any government body. Compiled entirely from publicly published sources. No classified, restricted or non-public information. Listed companies may dispute any entry: send the published source that contradicts it and the entry will be amended or removed.