Executive takeaway

Augment analysis with accountable expertise

AI can add speed, relevance, and structured detail to risk assessment, especially when analyzing image data, but the paper reports gaps in hazard completeness, practical control design, and contextual understanding. Responsible use therefore pairs AI outputs with domain expertise, local validation, and accountable human decisions.

Why It Matters

Risk teams increasingly need to interpret large and varied data quickly. This study shows how image-capable AI can help surface hazards, evaluate risks, and organize control ideas, but speed does not remove the need to understand the operating environment or the consequences of an incorrect recommendation.

The comparison is a reminder that risk management is both analytical and contextual. AI can support structured first-pass analysis, while experts contribute practical judgment, local and regulatory knowledge, broader safety awareness, and accountability for what is accepted and implemented.

What Was Examined

The paper describes a comparative approach in which ChatGPT-4 (AIc-4) analyzes image-based risk scenarios to identify hazards, evaluate risks, and suggest control measures. It contrasts the AI output with subject-matter-expert risk assessment and evaluates completeness, relevance, practicality, response time, comprehensiveness, contextual understanding, and industry and safety concerns across three case studies, including an Annual Airline Gala.

Key Findings

01

The paper reports that AI can produce rapid, detailed, and structured risk-assessment outputs; response time was described as a couple of minutes, compared with at least a single day for human experts before additional approvals and meetings.

02

AI-generated risk evaluations were reported as highly relevant to standard risk-matrix practice, but the paper also notes missed hazards and incomplete coverage in the evaluated scenarios.

03

AI-proposed controls could be specific to identified hazards yet too general for local requirements such as the New South Wales context; subject-matter experts supplied more detailed and context-sensitive recommendations.

04

Human expertise remained stronger in practicality, completeness, contextual understanding, and the development of broader safety awareness and collective responsibility.

05

The paper presents AI and human expertise as complementary: AI can organize and accelerate analysis, while experts challenge assumptions, interpret context, and decide what is appropriate to implement.

06

The authors identify accountability, data dependency, bias, dynamic risks, interpretability, integration, privacy, ethical, and regulatory considerations as conditions for responsible adoption.

Evidence boundary: These findings describe the paper's evaluated AI system, image-based scenarios, case-study design, comparison criteria, and contextual assumptions. They are not universal performance estimates, and AI-generated controls require independent verification against the actual workplace, applicable standards, and jurisdictional requirements.

What This Means in Practice

For Engineers

  • Use image analysis as a documented first-pass input to hazard identification, risk evaluation, and control design, not as a completed engineering assessment.
  • Verify every material output against site conditions, system boundaries, applicable standards, local requirements, and physical constraints.
  • Test for missed hazards, incomplete coverage, and context errors before incorporating an AI-assisted assessment into a safety case, risk register, or design decision.

For Practitioners

  • Pair AI-generated findings with qualified subject-matter review and require reviewers to challenge omissions, assumptions, and generic controls.
  • Use structured prompts and comparison criteria to make relevance, practicality, completeness, response time, and contextual understanding visible.
  • Retain human responsibility for interpreting the work environment, communicating risk, selecting controls, and confirming implementation.

For Leaders

  • Assign accountability for AI-assisted risk decisions and define who may review, approve, reject, or escalate outputs.
  • Govern data quality, privacy, bias, interpretability, system updates, regulatory alignment, and integration with existing safety practices.
  • Measure whether AI improves decision quality without weakening professional capability, risk awareness, or independent challenge.
  1. 01

    Define the risk-management use case, operational context, jurisdiction, and decision boundary before using AI.

  2. 02

    Curate and validate image data, prompts, assumptions, and contextual information before analysis.

  3. 03

    Require qualified human review of hazard identification, risk ratings, and proposed controls.

  4. 04

    Compare AI outputs with expert assessment, applicable standards, local requirements, and the actual operating environment.

  5. 05

    Test for missed hazards, incomplete analysis, generic controls, bias, and poor contextual interpretation.

  6. 06

    Document the AI system, inputs, outputs, reviewer, corrections, approval decision, and unresolved uncertainty.

  7. 07

    Monitor performance, privacy, accountability, regulatory fit, and changing risks as the use case evolves.

These recommendations are decision-support guidance, not legal advice or a universal compliance standard. Applicable requirements must be established for the organization, jurisdiction, technology, and use context.

Evaluation Framework

  1. 01

    Define the use case and operating context

  2. 02

    Curate and validate image data and assumptions

  3. 03

    Generate AI-assisted hazard, risk, and control outputs

  4. 04

    Compare outputs with subject-matter-expert assessment

  5. 05

    Evaluate completeness, relevance, practicality, speed, and context

  6. 06

    Review local, ethical, privacy, regulatory, and accountability requirements

  7. 07

    Approve, restrict, revise, or reject the use case with documented human oversight

A responsible evaluation treats AI output as one input to a structured comparison. The organization defines the context, validates the data, reviews the results against expert judgment and applicable requirements, tests limitations, and records an accountable decision about whether and how the use case may proceed.

Methods and Limitations

  • The paper uses image-based risk scenarios and case studies rather than a longitudinal operational deployment.
  • The comparison considers completeness, relevance, practicality, response time, comprehensiveness, contextual understanding, and industry and safety concerns.
  • The paper reports missed hazards, incomplete AI feedback, and control recommendations that may be too general for local requirements such as New South Wales.
  • Contextual understanding is described as fair, so the evaluated AI system may not reliably incorporate operational nuances without additional expert input.
  • The findings are tied to the evaluated AI system, prompts, images, cases, scoring approach, and assumptions; they should not be treated as universal performance estimates.
  • The paper identifies accountability, data dependency, bias, dynamic risks, interpretability, integration, privacy, ethical, and regulatory challenges that require ongoing governance.

Funding, Conflicts, and Institutional Independence

Funding

The publication reports no external funding.

Conflicts of interest

The authors declare no conflicts of interest.

References to employers, universities, laboratories, publishers, professional organizations, credentials, and institutional roles are included solely for factual biographical and publication identification. They do not imply affiliation, sponsorship, endorsement, partnership, authorization, or responsibility for Risk Analytics AI or its services. RAAI content is prepared and presented in an independent professional capacity.