Drawing inspiration from the "bug bounties" used to incentivise developers to spot errors in sortware, researchers from institutions including Alan Turing Institute, Google and Cambridge University last week called for a "robust toolbox" of mechanisms to support the verification of claims about AI systems and development processes. Ethical frameworks for AI have proliferated in recent years over growing concerns that algorithmic bias could result in discriminitory outcomes for certain groups.
From Principles to Mechanisms
“In order for AI developers to earn trust from system users, customers, civil society, governments, and other stakeholders that they are building AI responsibly, there is a need to move beyond [ethics] principles to a focus on mechanisms for demonstrating responsible behaviour,” the executive summary reads. “Making and assessing verifiable claims, to which developers can be held accountable, is one crucial step in this direction.”
The paper outlines 10 key recommendations of which bias bounties is one. It notes that bounties for other areas, such as security, privacy protection or interpretability, could also be explored.
In the case of a bias bounty, a developer, government or other organisation would offer financial rewards to people who discover and report bias in AI systems – addressing the risk that AI exacerbates existing race and gender prejudices.
While the paper notes that bounties alone cannot ensure that a system is safe, secure, or fair – “some system properties can be difficult to discover even with bounties, and the bounty hunting community might be too small to create strong assurances”, it says – it asserts that bounties might increase the amount of scrutiny applied to AI systems.
It encourages developers to pilot bias and safety bounties for AI systems. In establishing a bounty programme, the paper recommends they consider setting compensation rates for severity of issues discovered; determine processes for soliciting and evaluating bounty submissions; and develop processes for reporting and fixing issues discovered via bounties.
Commentaires