Responsible AI
Stakeholders, Affected People, and Power
Identify stakeholders, affected communities, power asymmetries, dependencies, and meaningful forms of participation.
By the end you can
- Explain why stakeholder analysis must distinguish influence, dependency, exposure, and the ability to refuse or seek remedy
- Distinguish Informing, Consulting, and Shared decision power
- Identify evidence that connects sponsors to public and civil society
- Design a review that moves from enumerate roles to close the feedback loop
Key idea
One representative does not make a process representative
Inviting one representative does not make a process representative. In 2022 four researchers made that the title of a peer-reviewed paper: Participation Is not a Design Fix for Machine Learning. They pull participation apart into three things: participation as work, participation as consultation, and participation as justice. They warn against "participation washing". Once the categories are separated, they write, "It also becomes possible to recognize where the discourse of participation fails to account for existing power dynamics and obscures the extractive nature of collaboration, openness, and sharing, particularly in corporate contexts." That is the shape of it. Taking part becomes extractive when people reveal risks and the organization keeps authority, deadlines and benefits unchanged. No consultation group speaks for every member of a heterogeneous community. Channels have to stay open after deployment, and dissenting or minority views have to be protected. No student other than Ogletree had ever objected to Cleveland State University's remote-exam room scans. The court held the practice unreasonable all the same.
Ask of any consultation not who attended but who could have refused. A process that looked thorough answers very differently.
Who can decide, who can refuse, who bears it
A stakeholder is anyone who can influence the system or be materially affected by it. A stakeholder list that flattens differences in power is not worth much. Responsible participation asks four questions instead: who can decide, who can refuse, who can contest, and who bears irreversible consequences. Those four rarely have the same answer. The gap between the first and the last is the whole subject of this lesson. The party that approves a system is almost never the party that has to live inside its output.
Consulting people is meaningful only when it changes information, design, or authority. A meeting after procurement cannot repair exclusion. Not if affected people lacked accessible materials, safe channels, compensation, or a credible route to challenge the project. Each of the cases below turns on one of those missing conditions. Three of them were eventually described in a court filing or a published letter. None of them was described in a project document.
Sort the people around a system by who decides and who bears the irreversible part. The asymmetry that shows up is what governance has to answer for.
Visual
Where stakeholder and power analysis enters the lifecycle
Sponsors, builders, operators, affected people and the wider public all touch the same system. Only some of them can walk away from it. In Ogletree v. Cleveland State University the last two bands included Ogletree's mother and two siblings, all of them home when the scan ran. They appear in the record of the case. They appear in no list the university drew up before deployment.
- 1
Sponsors
Set goals, budgets, timelines, and risk appetite.
- 2
Builders and vendors
Choose data, models, interfaces, and technical controls.
- 3
Operators
Interpret outputs, manage exceptions, and absorb workload.
- 4
Affected people
Receive benefits, burdens, surveillance, delay, or denial.
- 5
Public and civil society
Observe broader effects on norms, rights, markets, and institutions.
Example
Stakeholder and power analysis under operational pressure
A room scan before a remote exam is a Fourth Amendment search, and an unreasonable one. That is what the court held in Ogletree v. Cleveland State University on 22 August 2022. The record it worked from is short and specific. The scan lasted "less than a minute, and as little as ten to twenty seconds". Ogletree was warned only about two hours before the 12:30 pm exam on 17 February 2021. His mother and two siblings were all home at the time. No student other than Ogletree had ever objected to the scans. The court answered that fact head on: "Though schools may routinely employ remote technology to peer into houses without objection from some, most, or nearly all students, it does not follow that others might not object to the virtual intrusion into their homes or that the routine use of a practice such as room scans does not violate a privacy interest that society recognizes as reasonable, both factually and legally." What a room scan and an always-on camera expose is not evenly distributed: unstable internet, dark skin tones, religious head coverings, caregiving interruptions. For one of those categories the burden has been measured.
- Decision makers: The university set the requirement, the software and the notice period. Ogletree was told about two hours before the 12:30 pm exam on 17 February 2021 that his room would be scanned. Defining that as acceptable disruption was a decision nobody sitting the exam took part in.
- Operators: Faculty and support staff handle flags but may lack the time or authority to investigate context. That matters, because the flags are not distributed evenly. A study of proctoring output at the University of Louisville measured it: "Students with darker skin tones were flagged as missing from frame on average 4.79 times per assessment while students with medium and lighter skin tones were only flagged 1.39 times or 0.83 times, respectively".
- Direct users: Students must install software and comply with environmental requirements, and then the software has to see them. It does not see everyone equally. A 2022 study in Frontiers in Education analysed proctoring output for about 357 students in four STEM courses at Louisville. Faces were detected for an average of 78%, 87% and 92% of the assessment for students with darker, medium and lighter skin tones respectively (p < 0.001). Women with the darkest skin tones sat at 78% facial recognition, against 88% and 92% for medium and lighter skin tones. The authors say the software is used by at least 1,500 universities nationally.
- Affected non-users: Family members and housemates may be recorded without choosing to participate. In Ogletree's case they are named in the record: a mother and two siblings, all home when the scan ran, none of them a party to the exam, the contract, or any consultation about either.
- Power asymmetry: Students cannot easily refuse the tool without risking access to assessment. That is why an absence of complaints says so little. Cleveland State's scans had drawn no student objection but Ogletree's, and the practice was still found unreasonable.
Steps
Turn stakeholder and power analysis into an operating control
Power analysis becomes an operating control at the step where it records who was missing and what the organization did about it. For some deployers, the first three steps are no longer discretionary. Article 27 of the EU AI Act requires a fundamental rights impact assessment before most Annex III high-risk systems are deployed. It binds public-law bodies, private entities providing public services, and deployers of creditworthiness and life and health insurance pricing systems. One of its six required elements is "the categories of natural persons and groups likely to be affected by its use in the specific context", together with the specific risks of harm to those groups. The assessment must also set out the arrangements for internal governance and complaint mechanisms. Then the deployer notifies the market surveillance authority of the results, on a template developed by the AI Office. Enumerate roles, map power, identify missing voices: that is a list, with harms attached, filed with a regulator.
Step 4 is where the terms of taking part are written, and they can be written narrowly. Ofqual built a grading model for the 2020 A-levels. Its technical advisory group had too few independent external members — too few, that is, who were neither government employees nor current or former employees of the qualification regulators. That is the Royal Statistical Society's account, in a letter dated 14 August 2020 and published by the Office for Statistics Regulation four days later. The RSS had offered to nominate two distinguished Fellows. Here is what came back, in the letter's own words: "We eventually heard from Ofqual that they could consider these two Fellows, but only with a non-disclosure agreement that gave us real concern." The RSS read that agreement as precluding those Fellows from commenting in any way on the final choice of model for some years after results were released. Its offer to help was in the end not taken up. Ofqual's chair Roger Taylor replied on 21 August 2020. The RSS account was "incorrect both in terms of your reading of the confidentiality agreement and in relation to the correspondence with Ofqual", he wrote, and he made the confidentiality agreement public alongside his letter. Watch what step 5 is doing there. Both letters and the agreement itself were published. A reader can go and adjudicate the disagreement about the terms of participation.
1. Enumerate roles
Separate sponsors, builders, operators, subjects, bystanders, and downstream institutions.
2. Map power
Record who can approve, refuse, appeal, exit, and impose cost.
3. Identify missing voices
Look for people absent because participation is costly, unsafe, or inaccessible.
4. Design participation
Choose information, consultation, co-design, representation, or shared authority proportionate to impact.
5. Close the feedback loop
Publish what changed, what did not, and who decided.
Example
Evidence a stakeholder claim can be checked against
Each drill below asks for a record rather than an opinion: who holds power, what refusal costs, what participation was funded, and which objections were never resolved. The cases in this lesson supply the form each record takes when it survives.
- Stakeholder ledger: List each group, its interest, its exposure, and its decision power. Article 27 of the EU AI Act demands the same list of covered deployers — affected categories, and the specific risks of harm to them — and then requires the result to be notified to the market surveillance authority. That is a ledger with an external reader.
- Dependency test: Ask what happens if an affected person refuses the system. Ogletree's answer is on the docket. Warned about two hours before a 12:30 pm exam, he objected, then sat the scan. Whether he could have refused took until a ruling on 22 August 2022 to settle.
- Participation budget: Allocate time, translation, accessibility, compensation, and independent facilitation. Record the conditions attached, too, because conditions are where a budget is quietly spent. The RSS offered Ofqual's technical advisory group two distinguished Fellows and reported that the non-disclosure agreement offered in return gave it real concern. The offer was in the end not taken up.
- Dissent record: Preserve unresolved objections rather than summarizing them away. Treat a single objection as a finding, not an outlier. Cleveland State had heard from no student other than Ogletree, and the court held that it does not follow from routine acceptance that others might not object.
Comparison
Informing, Consulting, or Shared decision power?
Informing, consulting and shared decision power differ on one question: whether anything the participants say can still change the outcome. The three bands compress a named typology; this lesson did not coin them. An eight-rung ladder of citizen participation, set out by Sherry Arnstein in 1969, opens with the distinction this section rests on: "There is a critical difference between going through the empty ritual of participation and having the real power needed to affect the outcome of the process." The paper has 15,982 citations in OpenAlex. On her ladder, rungs (3) Informing and (4) Consultation are not participation at all. They are "degrees of tokenism". Only rungs (6) Partnership, (7) Delegated Power and (8) Citizen Control count as citizen power. Two of the three columns below have carried that label since 1969.
A court has drawn the same line inside a proctoring dispute. The University of Amsterdam's Central and Faculty Student Councils, together with one student, asked the Rechtbank Amsterdam to stop exams proctored with Proctorio. On 11 June 2020 the preliminary relief judge refused the injunction. No student-council right of consent (instemmingsrecht) applied, because invigilation rules fall under the exam board's exclusive competence. The Central Student Council had already sent an unsolicited advice on 29 April 2020, objecting to online proctoring and asking that no room scan be performed and that an alternative be offered. The judgment records what that advice was worth: "Dat de UvA ervoor heeft gekozen bij haar beslissing wel de studentenraden te betrekken getuigt van zorgvuldig handelen, maar maakt niet dat zij daardoor gehouden zou zijn hun instemming te vragen." Involving the councils was careful conduct. It created no duty to ask them for consent. The students were ordered to pay UvA's costs of EUR 1,636.00. Consulting and shared decision power are separated by exactly that: a duty, not a courtesy.
Informing
The organization explains a largely fixed plan.
- Useful for basic transparency
- Offers little influence over design
- May arrive too late for alternatives
- Appropriate only for low-impact changes
Consulting
People provide evidence and critique.
- Requires accessible and safe channels
- Needs a response to submitted concerns
- Influence can still remain advisory
- Useful during design and evaluation
Shared decision power
Affected groups shape scope, controls, or approval.
- Recognizes asymmetric consequences
- May include veto, representation, or co-design
- Requires resources and conflict rules
- Best for high-dependency or high-impact systems
A consultation with no record cannot be audited
Consultation that leaves no trace in the record cannot be audited afterwards, and an unauditable consultation is hard to tell apart from a decorative one. The Ofqual dispute is usable at all because two letters and a confidentiality agreement were published. The Amsterdam and Cleveland State disputes are usable because a court wrote down what each party had and had not been owed.
Define when stakeholder and power analysis requires the reviewer to redesign, restrict, remedy, or retire the system.
European law now writes one of these duties down, and writes it narrowly. The EU AI Act, Regulation (EU) 2024/1689, has been in the Official Journal since 12 July 2024. Article 26(7) reads: "Before putting into service or using a high-risk AI system at the workplace, deployers who are employers shall inform workers' representatives and the affected workers that they will be subject to the use of the high-risk AI system." The information is to be provided in accordance with Union and national law and practice on information of workers and their representatives. That is an information duty and nothing more. The people told are the people who never chose the tool. Notice is not consent. On Arnstein's ladder, informing is rung three.
Naming that limit is part of the analysis.
Key takeaways
- Stakeholder lists are incomplete unless they distinguish power, dependency, exposure, and recourse. Article 27 of the EU AI Act turns that list into a filing: covered deployers must record "the categories of natural persons and groups likely to be affected by its use in the specific context", along with the specific risks of harm to them.
- Direct users are not the only people affected by data collection or automated decisions. The record in Ogletree v. Cleveland State University includes a mother and two siblings who were all home when the room scan ran.
- Meaningful participation must occur early enough to change scope, controls, or approval. Informing and consultation were filed under "degrees of tokenism" on Arnstein's 1969 ladder precisely because neither guarantees that.
- Accessibility, compensation, translation, and psychological safety determine who can participate, and so do the terms attached. The Royal Statistical Society's two nominated Fellows were considered only with a non-disclosure agreement, and the offer was in the end not taken up.
- Token representation can hide an unequal distribution of authority rather than correct it. The Rechtbank Amsterdam found that involving the student councils showed careful conduct but created no duty to seek their consent, and the students who sued paid EUR 1,636.00 in costs.
- Governance should preserve dissent and keep participation channels open across the lifecycle, because silence measures cost rather than consent. No student but Ogletree had ever objected to Cleveland State's scans, and the practice was still held unreasonable on 22 August 2022.