ASD ISM 2085 Prevent exposure of AI model confidence scores
What it says
The exposure of exact AI model confidence scores in API outputs or user interfaces is prevented.
Quoted as published. Everything else on this page is written by Control Stack.
What it means
In plain EnglishDo not show AI model confidence scores in outputs to avoid revealing exact confidence levels.
This control advises against displaying how confident an AI model is about a decision or prediction. Sharing such confidence scores could allow others to figure out how your AI makes decisions, potentially leading to misuse or manipulation.
Free. Save your status, tasks and notes.
Why it matters
Revealing AI confidence scores could lead to exploitation, distorting the decision-making process or reducing system reliability.
How to implement
AI developers should avoid including confidence scores in user-facing outputs.
Instead, summarise results without specific numbers to prevent revealing internal working details.
Product managers should ensure user interfaces are designed to display decisions without confidence scores.
They can work with designers to present outcomes understandably without details that expose confidence.
IT teams must configure APIs to exclude confidence scores from responses.
This can be done by altering the API settings or data output format to suppress these numbers.
Project leads should regularly review AI outputs for unintentional exposure of confidence levels.
Set periodic audits or tests to confirm that outputs are free of these details.
Compliance officers should train staff to understand the importance of keeping confidence scores private.
Host workshops and create materials that outline risks of exposing this information.
Keep it running
Regularly check AI outputs to ensure confidence scores remain hidden, adapting practices as AI models evolve.
Evidence and audit
- Check 1 of 5What an auditor asks for
Ask for
user interface design documents: Verify that they specify no confidence scores shown.
Look at
screen mock-ups to ensure such details are not included.
Good evidence
is documentation showing outcomes without numeric confidence values.
- Check 2 of 5What an auditor asks for
Good evidence
is a clear indication that no confidence scores are provided.
- Check 3 of 5What an auditor asks for
Look at
adjustments made to suppress confidence scores.
Good evidence
is clear evidence of configuration changes.
- Check 4 of 5What an auditor asks for
Ask for
training materials on AI outputs: Ensure there's content educating staff on not sharing confidence scores. Training slides or manuals should highlight this control's requirements.
Good evidence
involves well-documented training sessions focused on privacy.
- Check 5 of 5What an auditor asks for
Look at
detailed records showing review dates and findings.
Good evidence
is consistent documentation of no breaches over time.
Mappings
Closest match in ISO/IEC 27001: A.8.28, Secure coding practices in software development, which this control partially meets.
ISO/IEC 27001: 1 mapping
This control partially meets A.8.28.
ISM-2085 requires organisations to prevent exposing exact AI model confidence scores in APIs and user interfaces. Annex A 8.28 requires secure coding practices, which would typically include designing API/UI responses to avoid leaking sensitive or security-relevant information such as exact confidence values. The ISO control is broader and applies to all secure coding concerns, while ISM-2085 is a specific secure-by-design output-handling requirement for AI systems.
Mappings show how controls relate across frameworks. They do not imply full equivalence or certification.