On this page we provide an overview of general risks of using AI in decision analysis, to help you ensure you are utilising AI responsibly, where it can it add value.
The risks related to the use of AI can be roughly divided into three categories:
Technical risks such as:
- Hallucinated facts
- Arithmetic errors
- Incorrect method descriptions
Process-related risks such as:
- Stakeholder exclusion
- Reduced participation
- Weakened ownership
- Overreliance on AI
Governance-related risks such as:
- Unclear accountability
- Undocumented AI use
- Hidden value judgments
- Lack of transparency
The extent to which these risks affect different parts of the decision-analysis process varies considerably. Technical risks are particularly relevant when AI is used to retrieve information, summarise evidence, perform calculations, or analyse data. Process-related risks are especially important when AI is used to support problem framing, stakeholder analysis, objective identification, or preference elicitation, where meaningful stakeholder engagement and learning are essential. Governance-related risks become increasingly important when AI influences comparisons of alternatives, policy recommendations, implementation planning, or communication of results, as these activities may affect the legitimacy, transparency and accountability of the decision process.
In general, the highest benefits and lowest risks of using AI are related to learning, explanation, communication, and exploration of decision problems. AI is particularly effective as a tutor, brainstorming partner, information synthesiser, communication assistant, and facilitator of structured thinking. In these roles, AI helps users broaden their understanding, consider alternative perspectives, and communicate complex information more effectively.
In contrast, the risks are greatest, and the benefits more limited, when AI influences one of the four foundations of the decision process: framing (defining the problem, stakeholders, objectives, and boundaries of the analysis), evidence (estimating impacts, interpreting data, and assessing uncertainties), values (eliciting preferences, assigning weights, and determining acceptable trade-offs), and decisions (identifying preferred alternatives, making recommendations, and selecting actions).
Interpreting the Risk Scores
The AIDA module uses a simple three-level scoring system to indicate the level of risk associated with using AI in different phases of the decision-analysis process.
! Low risk: A low score indicates that the risks are relatively limited and can usually be managed through verification and critical review.
!! Medium risk: A medium score indicates that AI should be used cautiously and that its outputs should be carefully examined and validated.
!!! High risk: A high score indicates that substantial risks exist and that strong human oversight is essential.
In general, the greatest risks arise when AI starts to influence the choices people make, rather than simply helping them to understand information and explore options
