IMRF Members Briefing Paper

IMO Maritime Safety Committee (MSC) Paper 111-21-2

Considerations on the need for performance standard(s) for AI systems embedded in marine equipment

1. Background

At IMO MSC meeting 111, a paper was submitted (MSC 111/21/2) which is a position and scoping paper arguing that IMO needs to set formal performance standards for AI systems embedded in marine equipment, not just high‑level principles.

In summary, the paper states:

• AI is already being embedded in navigational, radio, and marine decision‑support systems.
• Existing IMO instruments (SOLAS, COLREGs, GMDSS, performance standards for radar/ECDIS, etc.) assume deterministic behaviour, not probabilistic, data‑driven AI.
• That without specific performance standards, AI may:
• behave unpredictably in safety‑critical situations,
• be impossible to verify or type‑approve properly,
• undermine existing legal responsibilities (of the vessel’s master, company, manufacturer).
• IMO should therefore:
• develop performance standards for AI systems embedded in marine equipment,
• ensure human‑centred design, transparency, and controllability,
• require testing, validation, and documentation appropriate to safety‑of‑life functions.

It is not an “AI for SAR” paper but its logic lands squarely on SAR craft once applied to their use of GMDSS, navigation, and electronic search systems.

2. Core themes that may affect SAR units

Key themes in the paper:

• Embedded AI vs. stand‑alone AI: The focus in the paper is on AI inside equipment (e.g. radar, ECDIS, radio, electronic sensors, decision support).
• Safety‑critical functions: Where AI influences collision avoidance, navigation and routeing, watchkeeping, or distress/safety communications, higher assurance is needed.
• Performance standards: AI should be subject to clear, testable requirements - accuracy, reliability, latency, failure modes, environmental limits.
• Human in control: AI must support, not replace, the master/bridge team; humans must be able to understand, override, and disable AI functions.
• Transparency and explainability: For incident investigation and accountability, AI decisions must be traceable and reconstruct-able.
• Lifecycle and updates: AI models may change over time (re‑training, updates, etc.); this demands configuration control, re‑approval, and documentation.

For SAR vessels, these themes touch search planning, target/object detection, radio watchkeeping, and decision‑making.

3. Likely implications for AI in radio systems on SAR vessels

Affected systems may include: GMDSS consoles, DSC, VHF, MF/HF, satellite terminals, digital watchkeeping, automated triage of calls/messages by command-and-control systems, etc.

a) Performance and reliability requirements

• AI that filters, prioritises, or classifies distress, urgency, and safety traffic will likely need:
• minimum detection/false‑negative rates for distress calls,
• strict limits on false positives that could desensitise operators,
• defined latency (how fast a distress alert or voice call is notified to the relevant operator).

b) Human override and visibility

• Any AI that:
• auto‑routes calls,
• suppresses ‘noise’,
• or clusters messages
• will need:
• clear indication when AI is active,
• full visibility of all raw traffic on demand,
• a simple way to bypass AI (e.g. “show everything”, “AI off” mode).

c) Compliance with GMDSS and SAR obligations

• AI must never block or delay:
• distress alerts,
• distress relays,
• SAR coordination messages.
• Performance standards will likely require:
• demonstrable proof that AI cannot prevent a station in distress from using any means to seek assistance,
• test scenarios where AI is exposed to ambiguous or degraded signals and must still surface distress alerts or voice messages correctly. This may require additional ‘teaching’ of AI as human operators often become attuned to voice calls that have an implication of an emerging emergency despite there being no explicit distress pro-words used e.g. Mayday, Pan-Pan.

d) Logging and investigation

• AI decisions in radio systems (e.g. “classified as routine”, “filtered as noise”) will need:
• event logs,
• model/version identification,
• input–output traces for post‑incident analysis.

For an RCC or SAR operator, that means: if AI mis‑classifies a distress call, you must be able to prove what it did and why.

4. Likely implications for AI in navigational and detection systems on SAR vessels and aircraft

Affected systems may include Radar, ARPA, ECDIS, EO/IR, AIS, target and object recognition, search‑pattern optimisation, collision avoidance.

a) AI‑assisted detection and tracking

• AI used to:
• detect small targets (persons in water, life‑rafts),
• classify contacts,
• fuse radar/EO/IR/AIS information,
• will likely need performance standards specifying:
• probability of detection vs. sea state, visibility, clutter,
• false alarm rates,
• update rates and latency,
• degradation behaviour (what happens in heavy rain, high clutter, sensor failure).

For SAR craft, this is directly tied to probability of detection (POD) and search effectiveness.

b) COLREG Compliance and SAR‑specific manoeuvring

• If AI influences routeing or collision avoidance, standards will need to ensure:
• full compliance with COLREGs, even when SAR units are operating at high speed or in congested areas,
• clear rules for priority of SAR tasks vs. collision risk avoidance (AI must not “optimise” in a way that creates new hazards).

c) Search planning and decision support

• AI that proposes:
• search patterns,
• drift predictions,
• resource allocation,
• will need:
• validated models (e.g. drift, leeway, environmental data),
• clear presentation of uncertainty (confidence intervals, alternative options),
• no “black box” decisions - operators must understand the basis of recommendations made by the AI advice.

d) Human factors and trust

• Performance standards will likely require:
• intuitive displays showing what AI sees and why,
• training requirements for SAR crews on AI strengths/limitations,
• safeguards against trusting AI too much (e.g. mandatory cross‑checks, “AI is advisory and supporting, only”).

5. Practical implications for SAR organisations

The paper relates to the use of AI on commercial vessels and related shore-based infrastructure, systems and processes. All these impact SAR vessels and, to some degree, SAR aircraft, because the same commercially developed systems are also fitted to these craft. The paper leads to some considerations for SAR:

• Policy & procurement
• Build into equipment specification: the requirement that Any AI‑enabled radio or navigation system must comply with IMO AI performance standards once adopted.
• Require disclosure of AI functions, AI-training data assumptions, and AI update policies from vendors.
• Operational doctrine
• Treat AI outputs as decision support, not authoritative decisions – AI should not be the final decision maker.
• Embed in SOPs: when to trust, when to cross‑check, when to switch off AI functions.
• Training & competency
• Add modules on AI behaviour, limitations, and failure modes to SAR crew and RCC training to ensure maximum understanding of the strengths and limitations of AI.
• Train specifically on AI mis‑classification scenarios (e.g. missed weak DSC, mis‑detected small target, incorrect collision avoidance resolution, etc).
• Testing & exercises
• Include AI systems in search exercises and comms drills:
• measure detection performance for small targets,
• test how AI handles high traffic, overlapping distress, degraded signals.
• Practice using AI supported systems and understand its limitations.
• Capture data to feed back into local performance baselines and to manufacturers.
• Governance & audit
• Maintain a configuration register of AI‑enabled systems (model versions, activation dates).
• Ensure incident reviews explicitly examine AI contributions (helped, hindered, neutral).

Relevance to IMRF AI Survey and AI in SAR Project

The International Maritime Rescue Federation (IMRF) is currently conducting a survey on the application of Artificial Intelligence (AI) in Search and Rescue (SAR) operations. Members of the community are encouraged to participate in this survey, as its findings will contribute valuable insights into the evolving use of AI within SAR contexts.

The activities undertaken by the International Maritime Organization (IMO) are particularly significant for the SAR community. Decisions made at the IMO directly influence the development of international standards and requirements that govern the design and manufacture of ship equipment. These standards, along with the regulations established by the International Telecommunications Union (ITU) for radio equipment and operational rules, as well as those set by other technical governance bodies, shape the technology available to leisure craft, fishing vessels, and SAR boats and aircraft.

By actively contributing feedback through the IMRF survey and related discussions, the SAR community is able to influence the direction and outcomes of international regulatory bodies. This participation ensures that the creation of international rules and regulations reflects the practical needs of SAR operations and supports the adoption of effective technology within the sector.

IMRF Members Briefing Paper – Summary

IMO MSC Paper 111-21-2: Performance Standards for AI Systems in Marine Equipment

Overview

This briefing paper, prepared for IMRF members, summarises IMO Maritime Safety Committee (MSC) Paper 111-21-2, which advocates for the development of formal performance standards for Artificial Intelligence (AI) systems embedded in marine equipment. The document highlights the relevance to Search and Rescue (SAR) organisations, whose vessels, aircraft, and shore-based systems increasingly rely on AI-driven technology.

Key Points from the IMO MSC Paper

• AI is already being integrated into navigation, radio, and marine decision-support systems.
• Current IMO regulations are based on deterministic systems, not the probabilistic, data-driven nature of AI.
• Without specific standards, embedded AI may act unpredictably in critical situations, be hard to verify, and undermine existing legal responsibilities.
• The paper recommends IMO develop clear, testable performance standards for marine AI, ensuring human-centred design, transparency, controllability, and rigorous testing for safety-critical functions.

Implications for SAR Units

• AI embedded in essential SAR equipment (e.g. radar, ECDIS, radios) needs standards covering accuracy, reliability, latency, and environmental limits.
• AI must support, not replace, human operators and be fully controllable and explainable for incident investigations.
• AI models require configuration management and documentation as they evolve over time.

Practical Actions for SAR Organisations

• Policy & Procurement: Ensure all AI-enabled equipment meets IMO standards and require transparency from vendors.
• Operational Doctrine: Treat AI outputs as advisory; embed procedures for cross-checking and overriding AI decisions.
• Training & Competency: Add AI-specific modules to crew training, focusing on limitations and failure modes.
• Testing & Exercises: Include AI systems in drills, measure performance, and use results to improve local and manufacturer standards.
• Governance & Audit: Keep detailed records of AI systems and incorporate AI impact evaluation in incident reviews.

Conclusion

The IMO MSC paper calls for robust, formal standards for AI in marine equipment to ensure safety, reliability, and human control—issues directly impacting the operation of SAR vessels and aircraft. SAR organisations are advised to follow IMO discussions closely and contribute feedback to ensure these standards address the unique requirements of search and rescue operations.