Identifying Stakeholders in ML Systems
What you'll learn: Who is affected by your ML system and why it matters to map all stakeholders, not just end users.
The Bigger Picture
When you deploy an ML model, it doesn't exist in a vacuum. Every system creates ripples that touch different groups of people in different ways. Identifying stakeholders means systematically asking: Who has something at stake when this system exists and operates?
Categories of Stakeholders
End Users
The people who directly interact with your system. For a loan approval model, this includes applicants who fill out forms and receive decisions.
Decision Subjects
People whose lives are affected by the model's outputs, even if they never touch the interface. In facial recognition systems, everyone whose face appears in the dataset or gets scanned is a decision subject—whether they consented or not.
Impacted Communities
Groups indirectly affected by deployment at scale. A hiring algorithm might disproportionately filter out candidates from certain neighborhoods, affecting entire communities' economic opportunities.
Operators and Maintainers
The humans who run, monitor, and troubleshoot the system daily. Their working conditions and decision-making authority matter for system outcomes.
Indirectly Affected Parties
Think broader: competitors whose business models shift, workers whose jobs change, or society experiencing norm shifts. A content recommendation system affects not just viewers but also creators, advertisers, and public discourse itself.
Why This Matters
Missing stakeholders in your analysis means missing risks, harms, and responsibilities. The loan model's developers might focus on bank efficiency (end users = loan officers) but overlook applicants and historically excluded communities who bear the real consequences of errors or biases.
Key Takeaway: Comprehensive stakeholder identification requires looking beyond immediate users to decision subjects, affected communities, operators, and those experiencing indirect impacts—only then can you assess who truly bears the risks and benefits of your ML system.