The RSS AI Task Force has published a new paper, AI Regulation Needs Statistics, setting out some of the ways in which statistical considerations should be impacting the way that AI is regulated.
The paper builds on our recent paper AI is Statistics, where we argued that AI is fundamentally statistical – and appreciating the statistical nature of AI is essential to effectively and ethically using and evaluating the technology. A major focus of that work was around the challenges of evaluating AIs, given that they are dynamic systems that continue to evolve once they have been deployed in the real world.
Without recognising the statistical nature of AI, many important regulatory questions cannot be properly addressed. For example: how accurate and reliable outputs are, how performance varies across groups and settings, how sensitive systems are to shifts in data or context, how the output can be explained and interpreted and how uncertainty is measured and communicated. If AI is to be regulated well, it must be evaluated well – and this needs statistics.
Our paper looks at three sectors where AI is increasingly used – healthcare, education and finance – and shows some of the ways in which statistical evaluation would support regulation in those sectors. From those case studies we have drawn five key lessons that inform our recommendations:
-
Government should establish clear responsibility for identifying and addressing gaps in AI regulation
-
Regulators should ensure that the evaluation of the AI systems deployed in their sector reflects how AIs are operating in practice.
-
Government should ensure regulators have the right powers to require evidence generation for high-risk AI systems
-
Regulators should build their statistical capability and embed statistical thinking within their organisations.
-
Regulators should provide clear statistical guidance to organisations on the evaluation of AI to inform their procurement processes.
Read the full report.
We will now focus our attention on using this report to start discussions with government and regulators, aiming to ensure that statistical considerations properly inform regulatory practice.