Ethical AI (or Responsible AI) is about exercising a duty of care in how AI is being used. This starts with a clear understanding of the limitations and blind spots of AI both among the creators/developers and the operators of such AI applications. Based on this, the right scope and boundaries of AI can be determined. Safeguards such as mechanisms for exception handling, human interventions, the ability to override AI decisions, human oversight, and governance are all part of Responsible AI.

Another dimension of Responsible AI is how much it is trusted by the stakeholders. Deep learning AI systems are not intuitive and there is a “because AI said so” angle to their automated decisions. Even if AI is performing better than the current baseline, a few highly-publicised errors can undermine confidence in their reliability. Furthermore, if the decisions are challenged, there needs to be an easily understood rationale.

To increase the transparency and explainability of Deep Learning systems, technical and non-technical solutions are emerging. Documentation of datasets, specifying the boundary conditions for AI usage, relying on AI techniques that are more explainable, and deconstructing how AI models arrived at the decision (aka explainability) are some example approaches to increase the transparency and trust. Trust in AI systems is a necessary condition for their adoption.

As to the how of AI regulation, a wide range of regulatory tools and approaches are available. There is self-regulation where industry participants voluntarily establish standards and best practices and agree to abide by them. Organizations declaring their support for Responsible AI principles and standards is an example. Independent audits and third-party certifications of compliance to standards define the next level.

We already have product safety, consumer protection, and anti-discrimination laws, which also apply to products and services that use AI. When AI makes mistakes, the consequences depend on the context and use case. For example, autocorrect not working properly carries low stakes, while getting charged for a crime because of an AI error, has a massive impact and must be avoided. The bar for AI should be set higher when the cost of errors and consequences of mistakes is high. That’s exactly the level-of-risk approach to regulation currently being considered in the EU. Additionally, sector-specific regulation can bring contextual granularity to regulation.

AI is not only a horizontal technology with broad applications but also a dual-use technology, meaning that it can be put to both good and bad uses. AI requires updates to our regulatory approach and upgrades to our risk architectures. As a final note on regulation, AI easily crosses borders and we also need global treaties and conventions, particularly for AI used in warfare and military applications.