Trust and AI in Manitoba’s public sector

The Kinew government has embraced new technology as the basis for innovation and enhanced productivity in the economy, including the modernization of government operations. It established a new department for innovation and new technology, created a “blue-ribbon” advisory task force on the use of technology to support the economy, and launched public consultations on how AI systems could be used to promote the rights and opportunities of citizens.

This is part of the background to the Public Sector Artificial Intelligence and Cybersecurity Act (Bill 51) which is about to be sent to a committee of the legislature for detailed study. The bill represents a cautious first step to set some guardrails on the design, application and outcomes of AI in the public sector broadly defined.

Some brief, incomplete comments on AI and its potential impacts set the stage for the analysis of Bill 51.

AI is global in its reach, is evolving rapidly and is largely under the control of a small number of major technology companies. This means regulation of the private-sector use of AI must come mainly at the national level, with the provincial government potentially supplementing those rules.

An important distinction must be made between early AI systems designed for specific tasks ( like processing information) and the emergent large language systems, like ChatGPT or Google Gemini, which have been built to perform “thinking” and adapt their operations. Studies indicate that such systems make mistakes — called “hallucinations” — one to two per cent of the time.

Embedding newer generative AI systems into institutions offers the promise of efficiency and enhanced productivity. It potentially allows for services to be designed around citizens rather than through silo-based departments and programs. With a modern digital infrastructure and the appropriate authority and skills, public servants can offer citizens more timely and customized service.

There are also potential negatives associated with AI such as job losses, built-in biases, mistakes and unwanted outcomes. Operating like a “black box,” AI can lead to a loss of transparency and accountability. Talk of “superhuman” AI systems implies an even more fundamental loss within human relationships of such core values as autonomy, dignity, privacy, empathy, respect, compassion and fairness.

Trust matters more to the successful application of AI in the public sector than the technology itself. Put simply, trust involves positive, confident expectations about the motivations and competence of the individuals and institutions that control AI systems. Such sentiments are slow to develop and can be quickly lost, especially in the public sector which is under constant scrutiny.

To promote trust AI systems must be designed and operated on the basis of ethical reasoning and conduct that reflects shared values and rules within society. Commitment to ethical standards in AI can be based on personal moral judgments, on the requirement to comply with binding legal rules, and/or because compliance is beneficial to individuals and institutions financially or otherwise.

Trust also requires transparency and accountability. Decision-makers should not be allowed to hide behind opaque filters and algorithms in computer-based programs. Accountability requires monitoring, the duty to answer for actions and consequences for inappropriate or mistaken use of AI.

Turning to Bill 51, it begins with a preamble of six “whereas” statements intended to outline the principles/values which will provide the governance framework for AI in Manitoba’s public sector. The final “whereas,” calling for human oversight, risk assessment, transparency and accountability, is by far the most important of the six statements.

Consistent with the need for flexibility in a dynamic, risky field, the bill is drafted in general terms with the detailed rules left to the minister. The bill also recognizes that different public sector entities (which includes departments, non-departmental bodies and municipalities) will need some latitude to customize the use of AI based on their individual tasks and related operational requirements.

The freedom to experiment with new AI systems is balanced by a requirement for each entity to develop an accountability framework. What this entails is not specified in the bill, but presumably a framework would identify what functions will be AI-based, who is in charge, how performance will be monitored, how success will be defined and measured, and what might happen as a result of abuses or mistakes.

The bill also promises to set cybersecurity “standards,” to deal with such growing concerns as harmful fake images online and privacy violations. It authorizes the minister, with the approval of cabinet, to issue to a given entity a binding directive to develop such standards.

Directives must be specific to a particular entity, not applied generally, and must be made public, “in any manner the minister considers appropriate.” It is not clear why the minister needs to have complete discretion over how such directives will be made public. It would be better to require that directives be posted online and ultimately tabled in the legislature when it next meets.

Another provision of the bill grants the minister broad general authority to make regulations on any matter, a provision which governments routinely include to grant themselves maximum flexibility when it comes to applying a law. Unfortunately, decades ago, for many reasons, the legislature abandoned any systematic attempt to oversee the exercise of such delegated legislative authority granted to governments.

Bill 51 includes valuable requirements for public consultation on draft regulations and the automatic review of the effectiveness of any regulation after three years.

How consultation will occur is not prescribed. My suggestion would be a 90-day “notice-and-comment” period, which would allow for timely decisions while limiting potential dominance of the process by narrow interests. The review of regulations sensibility avoids a mechanistic “sunset” provision which would cause regulations to lapse automatically if not reviewed/renewed by a certain date on the calendar.

Bill 51 seeks to assure the public that AI will be used in the public sector in a safe and responsible manner. Revising the public service code of values and ethics to take account of AI would be another good step.

Paul G. Thomas, professor emeritus of political studies at the University of Manitoba, received helpful advice on AI from middle son Neal Thomas, PhD.