UNIDIR Plans Military-AI Forum as Human Control and Industry Responsibility Take Centre Stage

UNIDIR is preparing to launch a dedicated forum on artificial intelligence, peace and security on October 22. The initiative brings governments, industry and researchers into a continuing process as concerns grow over human judgement, accountability and the military use of AI.

A forum launch is scheduled for October 22

The United Nations Institute for Disarmament Research is preparing a new forum intended to connect the organisations shaping artificial intelligence in peace and security. Its launch is scheduled for Thursday, October 22, in New York, with Switzerland and Pakistan partnering on the event. The initiative follows an October 20 briefing by UNIDIR director Robin Geiss and arrives amid wider concern about how military decision-making is changing as automated systems become more capable.

UNIDIR’s event announcement describes an opening for expressions of interest in an organisational-level forum. It places the event alongside the General Assembly First Committee and identifies it as part of the institute’s Centre of Excellence on AI, Peace and Security. The centre itself was launched in June. The new forum should therefore be distinguished from the establishment of the broader centre.

The immediate development is a planned launch and an invitation to participate. It is not the adoption of an international treaty, the creation of a new enforcement authority or confirmation that governments have agreed on a single approach to military AI. The significance lies in whether a sustained process can connect technical knowledge with decisions made by states and the organisations supplying their systems.

Military AI reaches beyond autonomous weapons

In its report on the October 20 briefing, the Associated Press describes Geiss raising concerns about the narrowing space for human judgement, including in nuclear-related settings. He pointed to AI entering functions such as early warning, information collection and reconnaissance. AP also reports work on industry conduct and the risks of misuse by armed groups. These are concerns and initiatives described by the institute, not evidence of a particular imminent loss-of-control incident.

The breadth matters. A system does not need to choose and strike a target to influence a consequential decision. Software that sorts reports, prioritises alerts or summarises intelligence can change what a person sees first and what information receives less attention. A human may formally retain the final decision while relying on an automated picture that is incomplete or difficult to interrogate.

As an illustrative example, an analyst receiving a ranked list of possible threats needs to understand what generated the ranking and where the uncertainty lies. More information on a screen is not necessarily more usable understanding. The governance question includes the quality of the evidence presented to people, the time available to assess it and the practical ability to challenge the system’s recommendation.

What meaningful control would need to accomplish

The International Committee of the Red Cross’s position on autonomous weapons provides an independent humanitarian perspective. The ICRC recommends binding prohibitions on unpredictable autonomous weapons and on systems designed or used to apply force against people, together with restrictions on others. Its recommendations include limits on targets and operating conditions, effective supervision and the ability to intervene or deactivate systems. These are the organisation’s recommendations to states; they are not new rules adopted by the October forum.

That approach illustrates why the phrase “human control” needs an operational meaning. A person who can only watch an irreversible sequence unfold has a different role from one who can understand, constrain and stop it. A formal approval step can also be inadequate if the person has neither enough information nor sufficient time to make a meaningful choice.

The question is consequently about the entire arrangement: what the system is permitted to do, how its output is presented, what the operator knows and when intervention is possible. Different applications raise different risks. A general label such as AI-enabled cannot establish that all uses are equally dangerous or that all should be governed by identical technical measures.

Industry decisions are part of the governance problem

UNIDIR is also developing a Framework of Responsible Industry Behaviour for AI in the Military Domain, in partnership with the UN human rights office. The proposed framework consists of voluntary principles and practical guidelines. It focuses on decisions across design, data practices, testing and support after deployment. Its development process includes governments, industry, researchers and civil society.

The institute’s description explicitly says that contributing to development does not imply endorsement or create binding commitments. That caveat is useful when company names appear beside an international initiative. Participation can help shape a framework without demonstrating that a company has implemented every future recommendation, passed an independent assessment or received a UN certification.

The practical importance of involving suppliers is straightforward. A customer can set requirements, but the supplier may know most about how the system was built and tested. Responsibility can become blurred when several firms provide models, data, integration and maintenance. A useful framework would help clarify the information each participant must supply and the responsibilities that remain after delivery.

A separate national-strategy guide is already available

On October 19, UNIDIR published revised guidelines for national AI strategies in security and defence. The publication distinguishes the institutional steps of a strategy cycle from the technical, legal and political issues a strategy must address. It describes work that began in 2024 and was revised through feedback and research. This is a related policy resource, separate from the forthcoming forum launch.

The connection between those efforts is important. International discussion can identify shared problems, while national institutions still need to decide who has authority, how procurement works and how implementation will be reviewed. A strategy with broad aspirations but no allocation of responsibility may leave operational decisions to develop without a clear public framework.

A review cycle is particularly relevant to software that can change after its first assessment. An institution needs a way to recognise when a change is significant enough to require renewed scrutiny. That does not mean every update is equally consequential. It means the process should explain who makes that judgement and what evidence they need, rather than assuming that an initial approval settles the matter indefinitely.

Testing needs to follow the real conditions of use

UNIDIR’s overview of its AI work identifies risks including malfunction, misperception, bias and inadvertent escalation. Those are different failure modes. A system can operate consistently while reflecting unsuitable assumptions, or it can perform well in one setting and respond poorly when conditions change. Assessing only a headline accuracy score would miss those distinctions.

Consider a hypothetical tool that summarises surveillance reports. Its average performance may look strong while rare omissions concern the most consequential cases. A realistic evaluation would examine the kinds of errors, who is exposed to them and whether users can recognise uncertainty. It would also ask what happens when information is incomplete, contradictory or deliberately misleading.

Testing should therefore support a decision about a defined use, rather than serve as a general declaration that a model is safe. The more consequential the decision, the more important it becomes to connect the evidence with the actual conditions, constraints and human roles. These are analytical implications of the governance problem, not findings about a named deployed system.

Transparency must make accountability possible

An effective process needs to explain its boundaries. Public reporting can distinguish a research partnership, a voluntary commitment, a national policy and a binding obligation. These instruments can reinforce one another, but they have different consequences when someone fails to follow them. Treating every announcement as an enforceable agreement would overstate the progress made.

The same discipline applies to outcomes. A successful meeting might produce common terminology, a consultation plan or a draft framework. Those can be valuable steps without establishing that risks have already been reduced in deployed systems. Later evidence would need to show changes in procurement, testing, supervision or other practices, and explain how those changes were evaluated.

Inclusive participation will also need more than a list of prominent organisations. Institutions with limited technical resources must be able to understand proposals well enough to scrutinise them. Otherwise, the people most familiar with a technology can dominate the definition of acceptable practice, even when other participants bear important consequences from its use.

What to watch after the launch

The October 22 event is scheduled from 13:00 to 14:45 in New York, equivalent to 19:00–20:45 in Europe/Belgrade. The published programme includes introductory remarks, an invitation to engage and a moderated exchange. As of this October 21 report, the event remains ahead. There is no basis to describe future remarks as having been delivered or future proposals as agreed.

The next substantive questions will concern who joins, what work is prioritised and how the process reports its progress. Readers following our coverage of the importance of early-warning information will recognise the stakes of keeping information support distinct from an automatic decision to use force. UNIDIR’s initiative is an attempt to make those distinctions governable before the technology and its institutional uses move further ahead.

A practical measure of progress will be whether participants make their work understandable beyond the meeting room. Publishing the scope of consultations, the questions left unresolved and the evidence required for future recommendations would allow outsiders to distinguish agreement from aspiration. It would also let researchers identify gaps without assuming that the absence of a public dispute means the difficult questions have been settled. For a forum dealing with rapidly changing systems, that record can be as useful as a broad statement of shared concern.