Evaluation of Automated Output Checking and AI Model Risk Assessment Community of Interest#
Leads: Jim Smith (University of the West of England), Jackie Caldwell (PHS)
Proposal#
Summary#
The session is intended to give an overview of the Community of Interest on automated risk assessment. This naturally includes the risk assessment of AI models, since this is not something that can be done manually.
The aim is to: (1) introduce the Community of Interest group, and its aim of reducing barriers to adoption of (semi) automated checking, (2) make some proposals about how we plan to move forward on this alongside UK-TRE, (3) to get feedback from people present about and how they would like to se the community develop and work
Preparation#
No required preparation beyond an open mind!
If people would like some perspective on where the project has evolved from, it might be useful to skim-read the first 5-6 pages of the SACRO project’s final report
But please note this Community of Interest has a broader remit than just SACRO - for example, by design we include projects such as DataShield, as well as other approaches for assessing ML privacy leakage.
Target audience#
No specific target audience in mind - anybody interested!
Session#
Summary#
THe workshop explored projects already exploring these issues, what the priorities of the community should be, and how to align everything already happening in this space.
A name was also chosen for the community! ReBOT, Reducing Barriers to Outputs from TREs.
Next steps include setting up a Jisc mailing list for the community, and a simple accessible guide, either written or video.
Raw notes#
Start of a community of people looking at these tools
Some projects have started tackling issues: DataShield, ACRO, GRAIMatter, SACRO
Defining best practice
Aligning finding of these projects with SDAP manual
Remove barriers to adoption by researchers
Weekly drop-in sessions for ACRO, AISDC, SACRO, etc
Email support service
Name of the community? ReBOT, Reducing Barriers to Outputs from TREs
What does everyone think?
Manual egress is just not scalable
What are we actually trying to protect? From who/what?
Statistical disclosure policy, flowchart to follow, if not straightforward then document decision making, take consensus of senior members of team
Next steps#
Create new JISC mailing list
Create simple accessible guide, either written or video