Artificial intelligence (AI) governance has become a common concern among administrators. For example, a ministry deciding whether to automate benefit cheques, a city buying a hiring tool, a health agency testing a triage model and a regulator reviewing a facial-recognition system now face the same practical question: Is this AI system fit for public use? In McKinsey’s 2025 survey on the state of AI, 88 percent of respondents reported regular AI use in at least one business function, with roughly one-third reporting scaled enterprise use. In other words, adoption has outrun many accountability routines. Accuracy, speed, uptime and cost belong in the dashboard, but public legitimacy requires a broader base of evidence that considers affected people, legal authority, contestability, vendor responsibility, fiscal exposure and the environmental burden.
The existing policy baseline is complex, and incomplete. The Organisation for Economic Co-operation and Development (OECD) AI Principles promote trustworthy AI that respects human rights and democratic values, while the United Nations Economic, Scientific and Cultural Organization Recommendation on the Ethics of AI links AI governance to human rights, fairness and environmental sustainability across 194 member states. The National Institute of Standards and Technology AI Risk Management Framework Core gives organizations a working cycle: govern, map, measure and manage; and the EU AI Act creates a risk-based legal framework for high-risk systems, including requirements for risk management, data governance, technical documentation, logs, transparency, human oversight, accuracy, robustness and cybersecurity. ISO/IEC 42001 adds a management-system approach, with the Global Digital Compact placing AI governance inside a wider international agenda for inclusive digital cooperation. These approaches may suggest that a policy framework is in place, but the more important question is how that framework can be translated into traceable metrics.
The proposed Prosocial AI Index tries to make those principle-oriented baselines usable for policy makers, boards, regulators and public agencies. It translates governance duties into a 16-cell dashboard using color ratings: four Ts (Tailored, Trained, Tested and Targeted) read against four Ps (Purpose, People, Profit and Planet). Each cell asks for auditable evidence: green means the evidence is documented, monitored and owned; amber means the evidence is partial or there is an unresolved dependency; and red means evidence is missing, there is acceptable exposure or there is a deployment condition that requires a named decision maker.
On the T side, Tailored asks whether the system fits the setting: language, workflow, infrastructure, accessibility, law and affected groups. For a public agency, this would translate to local validation before procurement approval. Trained asks whether data, labels and optimization goals reflect the decision at hand. Tested asks for evidence before and after release: performance, bias, robustness, cybersecurity, usability, human-factors assessment, incident reporting and complaint routes. Lastly, Targeted asks whether the system is used for an authorized purpose, within clear boundaries, with escalation, appeal, audit rights and off-switch authority.
The four Ps aim to widen the accountability frame. Purpose covers mission fit and legal basis. People covers citizens, patients, students, workers, applicants, communities and frontline staff. Profit covers durable economic value; in the public sector it reads as stewardship, including total cost, service quality, procurement risk, liability, vendor dependency and resilience. Planet covers energy, water, hardware, model size and lifecycle effects. The International Energy Agency report Energy and AI, the OECD work on AI compute and environmental impacts and the United Nations Environment Programme’s lifecycle note on AI make the environmental cell a mainstream governance question.
A Policy-Maker-Friendly Scoring Rule
This index should sit inside the same workflow regularly used for budgets, procurement and regulatory sign-off. It can be completed in a 90-minute review for low-risk systems and expanded into a formal assurance file for high-risk uses. The rule is simple: every green cell needs a document link; every amber cell needs an owner and a deadline; and every red cell needs scaling blocked until the accountable official supplies evidence, narrows the use case or accepts the risk in writing.
The following table outlines examples of what this entails:
| Dashboard Cell | Policy Question | Evidence to Attach | Decision Trigger |
|---|---|---|---|
| Targeted x Purpose | Is this the right use of AI under the legal mandate? | Use-case approval; legal basis; human oversight; prohibited-use check | Red triggers policy review or withdrawal |
| Tailored x People | Can affected groups be harmed through errors, bias or poor usability? | Bias analysis; accessibility test; user testing; complaint route | Red blocks launch; amber requires remediation date |
| Tested x Profit | Will the system remain fiscally and operationally viable? | Total-cost model; audit rights; incident notice; exit plan; liability terms | Red triggers procurement escalation |
| Trained x Planet | What compute, energy, water and lifecycle evidence exists? | Vendor footprint data; cloud region; model-size comparison; low-compute option | Amber requires lower-impact option for comparison |
A city that buys an automated employment decision tool, for example, can use the index before tender award. Tailored x People would require accessibility for candidates with disabilities and language access for applicants who may use translated instructions, while Trained x People would require evidence that historical hiring data has been checked for bias. A few examples already exist, such as New York City’s Local Law 144 that now requires bias audits for covered automated employment decision tools. The US Equal Employment Opportunity Commission guidance for workers makes clear that federal employment discrimination protections apply when AI is used in workplace decisions. A prosocial AI score could turn these legal duties into procurement clauses and release gates.
A ministry of health comparing triage tools should, for example, include benchmark accuracy, frontline conditions and maintenance capacity in the same review. The World Health Organization regulatory considerations for AI in health emphasize risk-benefit assessment, evaluation and performance monitoring, but the external validation of the Epic Sepsis Model found poor discrimination and calibration in predicting sepsis, illustrating why local testing before wide deployment matters. In this case, the index would reward a slightly less complex model that frontline staff can use safely and maintain over time.
In benefits, debt recovery and fraud detection, a welfare agency using AI to flag suspected fraud needs stronger evidence than administrative efficiency. The Australian Royal Commission into the Robodebt Scheme report and the Dutch system risk indication judgment summarized by the Library of Congress show what happens when automated risk logic enters welfare administration with insufficient legality, transparency and human safeguards.
At the same time, when exams, school places or scholarships are influenced by algorithmic tools, Tested x People should require distributional analysis before results are released. Targeted x Purpose should require a clear appeals path and a communication plan that families can understand. The UK Office of Qualifications and Examinations Regulation interim report on 2020 exam grading offers a useful policy lesson: a technically motivated standardization model loses legitimacy when students, teachers and universities experience the decision as opaque, unfair and rushed. A dashboard review would have forced an earlier discussion of cohort effects, small schools, private candidates and remedy design.
Lastly, implementations for public safety or retail surveillance should score red when false-positive risk, demographic performance and remedy routes are undocumented. The Federal Trade Commission action against Rite Aid banned the company from using facial recognition for surveillance purposes for five years after allegations that the system falsely tagged consumers, particularly women and people of colour, as shoplifters. In Prosocial AI Index terms, Tested x People and Targeted x Purpose would have required demographic performance evidence, use restrictions, human verification, incident logs and a mechanism for affected people to correct the record.
For ministers and agency heads, the index supplies a plain language brief: which AI systems are ready, which need conditions and which require a pause. For procurement officials, it converts AI ethics into tender clauses, evidence requests, service-level terms and audit rights. For regulators, it creates a common framework for comparing systems across sectors. In other words, transparency becomes a lever for hybrid trust.
The practical question is straightforward: What value does this AI create, for whom, under which legal and policy duties, at what cost, with which safeguards, and with which evidence of improvement over time? The Prosocial AI Index gives policy makers a way to answer that question before the public carries the risk.
