Webinar: Why responsible AI could help government move beyond experimentation

Public sector organisations looking to move artificial intelligence (AI) projects beyond experimentation need to consider governance and accountability from the outset, according to SAS.

Antti Heino, principal AI advisor at SAS Institute, said responsible AI ultimately comes down to whether the technology can be trusted by the people affected by it.

“Responsible AI is about developing and deploying AI in a way that citizens, regulators, and public servants can trust,” said Heino.

Implementing responsible AI is the topic of an upcoming webinar, Using AI Responsibly in the Public Sector, which will examine how public sector organisations can adopt AI while maintaining transparency, accountability and public trust.

For public bodies, that means asking questions about an AI system before deployment: who is accountable for it, whether its decisions can be understood, what safeguards exist against unintended harm, and whether it is secure, resilient and lawful.

Organisations should also be able to demonstrate that the technology improves public outcomes.

Getting AI out of the experimentation stage

Experimentation has an important role in helping organisations understand what AI can and cannot do, according to Heino. But problems can emerge when organisations fail to consider what will be required to operate a system over the longer term.

“The biggest risk usually comes from the lack of understanding,” he said. “AI is already incredibly capable, but it has its flaws. When we understand where and how we can use AI safely, the risk of getting it wrong decreases significantly.”

Those risks include AI producing incorrect answers and biased or unfair outcomes. Heino said the data used both to train an AI system and when it makes a decision can make a significant difference to the trustworthiness of its outputs.

There is also the danger of investing in AI without a sufficiently clear business or public-service need.

“There should be clear value and need to use AI that originates from an organisation’s strategy,” he said. “Feasibility to maintain accountability, transparency and trust should be evaluated beforehand to avoid wasting resources.”

Organisations also need to consider the costs associated with running AI after the initial development phase.

Governance, model retraining and ongoing maintenance can represent a significant proportion of the overall cost, depending on whether an organisation develops its own system or buys an existing solution.

Building reusable responsible AI capabilities can therefore help remove some of the barriers to taking subsequent use cases into operational deployment, Heino argued.

Governance and innovation are not competing priorities

This becomes particularly important in the public sector, where organisations need to balance the potential benefits of AI against their responsibilities to citizens.

“Innovation and governance should be viewed as forces that strengthen each other. They are not competing priorities,” said Heino. “In the public sector, trust is what enables innovation to scale.”

One way of achieving that balance is through risk-based governance, where the level of scrutiny and controls reflects the potential consequences of an AI system.

Lower-risk applications can be subject to lighter controls, while higher-risk uses – including those involving benefits, healthcare and policing – require stronger governance, documentation, testing and oversight.

Rather than applying the same requirements to every AI project, this approach allows organisations to concentrate their strongest safeguards on applications where the potential consequences are greatest.

Five steps towards responsible AI

So what can public sector leaders do now?

Heino recommends beginning with an inventory of the AI models and use cases already operating across an organisation. As well as giving leaders greater visibility, this can identify duplicated work and provide a clearer picture of the risks that need to be managed.

From there, organisations can establish risk tiers and apply appropriate controls to different types of AI.

Monitoring is another important component. Heino said organisations need model and decision-management capabilities that can monitor potential bias and provide transparency into decisions.

He also recommends considering whether synthetic data – artificially generated data designed to mimic characteristics of real-world data – could reduce exposure to some data-related risks.

Finally, responsible AI cannot be treated purely as a technology or compliance exercise. Organisations need to invest in skills and culture so that employees have a shared understanding of both AI’s opportunities and its risks.

“Organisations that embed clear responsible AI practices from the beginning will move faster than those that try to retrofit them later,” said Heino.

The practical challenges of putting these principles into practice will be explored further in SAS’s upcoming webinar, Using AI Responsibly in the Public Sector on September 17. Register your place now.