The global AI governance landscape is complex and rapidly evolving. Key themes and concerns are emerging, however government agencies should get ahead of the game by evaluating their agency-specific priorities and processes. Compliance with official policies through auditing tools and other measures is merely the final step. The groundwork for effectively operationalizing governance is human-centered, and includes securing funded mandates, identifying accountable leaders, developing agency-wide AI literacy and centers of excellence and incorporating insights from academia, non-profits and private industry.

The global governance landscape

As of this writing, the OECD Policy Observatory lists 668 national AI governance initiatives from 69 countries, territories and the EU. These include national strategies, agendas and plans; AI coordination or monitoring bodies; public consultations of stakeholders or experts; and initiatives for the use of AI in the public sector. Moreover, the OECD places legally enforceable AI regulations and standards in a separate category from the initiatives mentioned earlier, in which it lists an additional 337 initiatives.

The term governance can be hard to define. In the context of AI, it can refer to the safety and ethics guardrails of AI tools and systems, policies concerning data access and model usage or the government-mandated regulation itself. Therefore, we see national and international guidelines address these overlapping and intersecting definitions in a variety of ways. For all these reasons AI governance should begin at the level of concept and continue throughout the lifecycle of the AI solution.

Common challenges, common themes

Broadly, government agencies strive for governance that supports and balances societal concerns of economic prosperity, national security and political dynamics, as we’ve seen in the recent White House order to establish AI governance boards in U.S. federal agencies. Meanwhile, many private companies seem to prioritize economic prosperity, focusing on efficiency and productivity that drives business success and shareholder value and some companies such as IBM emphasize integrating guardrails into AI workflows.

Non-governmental bodies, academics and other experts are also publishing guidance useful to public sector agencies. This year the World Economic Forum’s AI Governance Alliance published the Presidio AI Framework (PDF). It “…provides a structured approach to the safe development, deployment and use of generative AI. In doing so, the framework highlights gaps and opportunities in addressing safety concerns, viewed from the perspective of four primary actors: AI model creators, AI model adapters, AI model users, and AI application users.”

Across industries and sectors, some common regulatory themes are emerging. For instance, it is increasingly advisable to provide transparency to end users about the presence and use of any AI they are interacting with. Leaders must ensure reliability of performance and resistance to attack, as well as actionable commitment to social responsibility. This includes prioritizing fairness and lack of bias in training data and output, minimizing environmental impact, and increasing accountability through designation of responsible individuals and organization-wide education.

Policies are not enough

Whether governance policies rely on soft law or formal enforcement, and no matter how comprehensively or eruditely they are written, they are only principles. How organizations put them into action is what counts. For example, New York City published its own AI Action plan in October 2023, and formalized its AI principles in March 2024. Though these principles aligned with the themes above–including stating that AI tools “should be tested before deployment”–the AI-powered chatbot that the city rolled out to answer questions about starting and operating a business gave answers that encouraged users to break the law. Where did the implementation break down?

Operationalizing governance requires a human-centered, accountable, participatory approach. Let’s look at three key actions that agencies must take:

1. Designate accountable leaders and fund their mandates

Trust cannot exist without accountability. To operationalize governance frameworks, government agencies require accountable leaders that have funded mandates to do the work. To cite just one knowledge gap: several senior technology leaders we’ve spoken to have no comprehension of how data can be biased. Data is an artifact of human experience, prone to calcifying worldviews and inequity. AI can be viewed as a mirror that reflects our biases back to us. It is imperative that we identify accountable leaders who understand this and can be both financially empowered and held responsible for ensuring their AI is ethically operated and aligns with the values of the community it serves.

2. Provide applied governance training

We observe many agencies holding AI “innovation days” and hackathons aimed at improving operational efficiencies (such as reducing costs, engaging citizens or employees and other KPIs). We recommend that these hackathons be extended in scope to address the challenges of AI governance, through these steps:

• Step 1: Three months before the pilots are presented, have a candidate governance leader host a keynote on AI ethics to hackathon participants.
• Step 2: Have the government agency that is establishing the policy act as judge for the event. Provide criteria on how pilot projects will be judged that includes AI governance artifacts (documentation outputs) including factsheets, audit reports, layers-of-effect analysis (intended, unintended, primary and secondary impacts) and functional and non-functional requirements of the model in operation.
• Step 3: For six to eight weeks leading up to the presentation date, offer applied training to the teams on developing these artifacts through workshops on their specific use cases. Bolster development teams by inviting diverse, multidisciplinary teams to join them in these workshops as they assess ethics and model risk.
• Step 4: On the day of the event, have each team present their work in a holistic way, demonstrating how they have assessed and would mitigate various risks associated with their use cases. Judges with domain expertise, regulatory, and cybersecurity backgrounds should question and evaluate each team’s work.

These timelines are based on our experience giving practitioners applied training with respect to very specific use cases. It gives would-be leaders a chance to do the actual work of governance, guided by a coach, while putting team members in the role of discerning governance judges.

But hackathons are not enough. One cannot learn everything in three months. Agencies should invest in building a culture of AI literacy education that fosters ongoing learning, including discarding old assumptions when necessary.

3. Evaluate inventory beyond algorithmic impact assessments

Organizations that develop many AI models often rely on algorithmic impact assessment forms as their primary mechanism to gather important metadata about their inventory and assess and mitigate the risks of AI models before they are deployed. These forms only survey AI model owners or procurers about the purpose of the AI model, its training data and approach, accountable parties and concerns for disparate impact.

There are many causes of concern about these forms being used in isolation without rigorous education, communication and cultural considerations. These include:

1. Incentives: Are individuals incentivized or disincentivized to fill out these forms thoughtfully? We find that most are disincentivized because they have quotas to meet.
   
   
2. Responsibility for risk: These forms can imply that model owners will be absolved of risk because they used a certain technology or cloud host or procured a model from a third party.
   
   
3. Relevant definitions of AI: Model owners may not realize that what they are procuring or deploying meets the definition of AI or intelligent automation as described by a regulation.
   
   
4. Ignorance about disparate impact: By putting the onus on a single person to complete and submit an algorithmic assessment form, one could argue that accurate assessment of disparate impact is omitted by design.

We have seen concerning form inputs made by AI practitioners across geographies and across education levels, and by those who say that they have read the published policy and understand the principles. Such entries include “How could my AI model be unfair if I am not gathering PII?,” and “There are no risks for disparate impact as I have the best of intentions.” These point to the urgent need for applied training, and an organizational culture that consistently measures model behaviors against clearly defined ethical guidelines.

Creating a culture of responsibility and collaboration

A participatory and inclusive culture is essential as organizations grapple with governing a technology with such far-reaching impact. As we have discussed previously, diversity is not a political factor but a mathematical one. Multidisciplinary centers of excellence are essential to help ensure that employees are educated and responsible AI users who understand risks and disparate impact. Organizations must make governance integral to collaborative innovation efforts, and stress that responsibility belongs to everyone, not just model owners. They must identify truly accountable leaders who bring a socio-technical perspective to issues of governance and who welcome new approaches to mitigating AI risk whatever the source—governmental, non-governmental or academic.

For more on this topic, read a summary of a recent IBM Center for Business in Government roundtable with government leaders and stakeholders on how responsible use of artificial intelligence can benefit the public by improving agency service delivery.