Mapping AI value alignment practices: an evidence-based approach for safer public service AI
About the project
The rapid development and increasing use of advanced artificial intelligence (AI) systems have raised ethical and societal concerns among the public and policymakers, including risks of bias and discrimination, threats to privacy, labour displacement and negative environmental impacts. These challenges are intensified by the complexity of general‑purpose AI systems—such as generative AI—where limited insight into how these systems function makes it difficult to assess and manage potential risks. Value alignment, one approach to risk mitigation, ensures that AI systems operate in line with human intentions and values. For instance, aligning AI-enabled services with privacy requires intentional system design that accounts for the use of confidential, context-specific information about recipients.
This project critically assesses the state of knowledge on AI value alignment as an essential aspect of AI governance. Our systematic review identified concrete practices for aligning AI systems with human, organizational and societal values. Additionally, it identifies which institutional changes are necessary to support these practices and prevent harm from direct or indirect exposure to AI systems, particularly in the public service context. Our analysis was informed by the Values and Ethics Code for the Public Sector to ensure value propositions grounded in the behaviours expected of Canadian public servants.
Key findings
Some values are more congruent than others
- Key values like openness, equity, fairness and accountability are commonly listed as necessary for the responsible development of AI systems that protect human dignity.
- Values such as effectiveness, reliability and professionalism are often linked to practical rationales for the implementation of AI tools within the public sector.
- Sustainability as a value is overlooked, with only a couple of expert articles listing this value as necessary for a responsible AI system. This indicates a general disconnect between environmental values and AI tools.
Barriers to AI value alignment
- Organizational and institutional barriers are the most frequently identified obstacles to successful value alignment in practice. These challenges stem from differing organizational interpretations of implementation goals, conflicting value priorities and regulatory gaps.
- Geographical and cultural contexts are also a noticeable barrier. For example, Indigenous communities engaging with AI-facilitated services have noted a lack of cultural sensitivity to their norms and values.
- Knowledge and digital literacy gaps hinder successful implementation of value-aligned AI systems. These challenges manifest in expert recruitment, the availability of training and design practices for AI-enabled tools and services.
Ways to ensure value alignment
- Surveys are the most common tool for identifying shared public values, along with multi-criteria decision frameworks and methods for combining preferences. However, experts recommend shifting toward more interactive and deliberative processes that bring stakeholders together to discuss trade-offs, hear diverse perspectives, and develop more informed and nuanced views.
- Frameworks and implementation models are viewed as robust and actionable methods to ensure public values are embedded in public sector AI systems. For example, design frameworks like value-sensitive design (VSD), 4-steps framework, and the Human-Computer trust model, all integrate values in their design from the project outset.
- Meaningful human oversight of AI systems is required to preserve human agency and accountability. Accordingly, explainable AI is preferred over “black-box” models because it supports informed decision-making.
Policy implications
Stronger governance on value alignment
- Currently, the Government of Canada provides various guides and directives for the responsible use of AI in government. While both the Guide on the use of generative artificial intelligence and the Algorithmic Impact Assessment (AIA) tool mandate a risk-mitigation approach, they should be refined to include target values that are sensitive to the intended uses and affected populations.
- The Government of Canada could also consider embedding value alignment within the Policy on Service and Digital as a requirement for all digitally-enabled projects.
Ensuring citizen buy-in
- To build trust, departments could adopt transparency practices such as publishing “value scorecards” that document metrics, targets for achieving public value alignment and known system limitations. Periodic external audits can verify these commitments, while public access to decision logs and system records can enhance public buy-in.
Workforce and capacity-building around AI roles
- A significant barrier to successful AI implementation has been the uneven level of digital literacy across public sector roles. While large-scale AI literacy training for the public service would require substantial investment, a centralized knowledge repository could quickly standardize understanding of AI tools and systems. Embedding key public values into these resources would help normalize value-sensitive design across the public service.
Contact the researchers
Principal investigator
Adegboyega Ojo, Professor and Canada Research Chair in Governance and Artificial Intelligence, School of Public Policy and Administration, Carleton University: adegboyega.ojo@carleton.ca
Research assistants
Nicolas Ferraiuolo, PhD Candidate, Arizona State University: nferraiu@asu.edu
Kasia Polanska, PhD Candidate, School of Public Policy and Administration, Carleton University: katarzynapolanska@cmail.carleton.ca
Further information
Read full report (Coming soon)