# AI and HAZOP: studying AI systems, and using AI in the study An AI system in the process adds HAZOP causes the guidewords do not prompt on their own, such as model drift, inputs outside the training range and plausible wrong output. In LOPA, an AI advisory or an AI-adjusted loop should generally not be credited as an independent protection layer, because its independence and its performance on demand are hard to demonstrate. When AI is used to draft the HAZOP itself, the study leader and team still own every node, cause, safeguard and recommendation. ## What new causes does an AI system bring to a HAZOP? Treat the model's output as a process variable and apply the guidewords to it. These causes need adding to the study. - The model drifts as the plant or feed changes. - An input falls outside the training range and the output stays plausible. - The output is wrong but credible, so nothing alarms. - The AI is lost through a server, network or data feed failure. - Data is stale, such as a frozen tag. - The vendor updates the model and its behaviour changes without a site change request. In the [Linmara](https://www.autonomous-operations.ai/cases/linmara/) case, a new well outside the training data and a failed analyser combine in the same week. The vendor update cause belongs in [management of change](https://www.autonomous-operations.ai/ai-management-of-change/) as much as in the HAZOP, as the [Quillmere](https://www.autonomous-operations.ai/cases/quillmere/) case shows. ## Can an AI advisory count as an independent protection layer? Generally not, without a strong justification. Reviewing LOPA studies after Buncefield, HSL and HSE authors summarised IEC 61511 as requiring a protection layer to be independent of the initiating cause and of other credited layers, and auditable, meaning designed so its protective function can be regularly validated ([Chambers and Pearson, Hazards XXII, 2011](https://www.icheme.org/media/9286/xxii-paper-75.pdf)). An AI model usually reads the same transmitters as the basic process control system, and it may set the loop whose failure starts the scenario. Its failures are systematic and depend on the data, so a periodic test does not show it will be right on demand outside the conditions tested. [ISO/IEC TR 5469:2024](https://www.iso.org/standard/81283.html) is a technical report giving guidance on AI in safety-related functions. It sets no requirements. Record the AI as a non-credited safeguard and credit only the layers that stand without it. The same paper expects an alarm layer to be described with the operator response and the means of stopping the event. If an AI ranks, filters or suppresses alarms, it sits inside that layer. ## How does operator trust in the AI affect the study? An AI that has been right for months becomes the reading operators act on. The HAZOP should record which indication the operator will believe when the AI and an instrument disagree, and the procedure should say which one governs. The NIST AI RMF notes that how far people are empowered to challenge AI output needs further study, and suggests collecting data on how often they overrule it ([NIST AI 100-1](https://doi.org/10.6028/NIST.AI.100-1)). Where operator response is credited, the task analysis should include the AI display. ## Can AI run or draft a HAZOP? AI can propose deviations, split nodes and record the session. OSHA PSM requires the hazard analysis to be performed by a team with engineering and operations expertise, including someone with experience of the process and someone who knows the method ([29 CFR 1910.119(e)(4)](https://www.osha.gov/laws-regs/regulations/standardnumber/1910/1910.119)). The study leader and team must still own: - node boundaries and design intent, checked against current P&IDs and site modifications; - each cause and safeguard, since a generated list reads as complete and omits what only the site knows; - risk ranking and every recommendation; - traceability, recording which entries came from the tool, which version, and who accepted them. P&IDs, cause and effect charts and incident history are confidential site data. Before any tool sees them, establish where the data goes, whether the provider keeps or trains on it, and whether site rules allow it. ## Where does this sit in the management system? HAZOP is one element of [AI in process safety management](https://www.autonomous-operations.ai/ai-process-safety-management/). Where an AI sits near a trip, see [AI and safety instrumented systems](https://www.autonomous-operations.ai/ai-safety-instrumented-systems/). If an AI system is going into your next HAZOP or revalidation, start with an [autonomous operations briefing](https://www.autonomous-operations.ai/#how-do-i-work-the-m-ops-service-range).