A recent paper published on arXiv cs.CY examines the challenges public service AI governance frameworks encounter with general-purpose AI (GPAI). The authors argue that the characteristics of GPAI, such as its generality, accessibility, and low deployment cost, undermine the traditional approaches to AI safety. These frameworks, often designed for narrow, purpose-built AI, rely on concepts like accuracy, bias, explainability, and accountability that become less tractable with GPAI's unbounded outputs and free-text judgments.
Key Points
- Public services are under increasing pressure to adopt AI to address rising demand and resource limitations.
- This pressure has intensified with the advent of general-purpose AI (GPAI), which is built on large language models.
- GPAI can be directed by prompts to perform an effectively unbounded range of tasks.
- The generality, accessibility, and low deployment cost of GPAI challenge historical AI safety approaches.
- Traditional safety concepts like accuracy, bias, explainability, and accountability, which were manageable with narrow AI, are less applicable to GPAI.
- Accuracy is difficult to quantify for unbounded outputs generated by GPAI.
- Bias is harder to disaggregate when GPAI produces free-text judgments instead of categorical predictions.
- Explainability in GPAI may present the appearance of explanation, and accountability can diminish as outputs are optimized for persuasion.
Context
According to the arXiv paper, public service governance frameworks have historically presupposed conditions that GPAI removes. The mitigations prescribed in guidance documents are often based on the characteristics of narrow, purpose-built AI, where concepts like accuracy and bias could be more readily quantified and addressed. The paper develops its argument through the case of policing, highlighting the potential consequences of these governance gaps.
Why It Matters
This research indicates that policymakers and public service organizations need to re-evaluate and potentially redesign AI governance frameworks to effectively manage the unique properties and risks associated with general-purpose AI. The current reliance on frameworks designed for narrow AI may leave public services vulnerable to unforeseen challenges in areas like fairness and transparency.
What To Do
- Review existing AI governance frameworks within public service organizations to identify areas that may not account for GPAI's characteristics.
- Note the specific challenges highlighted regarding accuracy, bias, explainability, and accountability in the context of GPAI.
- Watch for further research or policy guidance that proposes new approaches to governing general-purpose AI in public service.
- Consider how the case of policing, as mentioned in the paper, might inform governance strategies in other public service sectors.