AI Bias Detection — A New Normal from NIST

NIST has been in the forefront of a lot of technology initiatives (years ago I used to participate in the PKI initiative — even co-chaired the 3rd annual research workshop then) and so there is no surprise that NIST has taken a leadership in AI.

They have a set of initiatives starting from the AI Risk Management Framework [Here] and also in the bias in AI. Their document Identifying and Managing Bias in Artificial Intelligence is the topic of this blog.

Let me just dive in … I will write about the implications, impressions and some thoughts. A very short summary of ~45 pages plus definitions and references.

Implications

While the publication covers a lot of ground, the impliucations are very interesting.

1. We should question Data Driven Decisions … more precisely the data might not be qualified to make those decisions — focus on what data should be used rather than what is available !
2. Societal value implications will triumph computational/accuracy metrics to evaluate an AI Model; in the future, this will become a regulatory policy
3. It takes a village to develop AI models — Multi-stakeholder Engagement & Impact Analysis is essential for a robust AI practice. For example, Technology or datasets that seem non-problematic to one group may be deemed disastrous by others
4. The document acknowledges the realities — AI is neither built nor deployed in a vacuum, sealed off from societal realities of discrimination or unfair practices
5. While there are lots of principles, the policy and practice side of AI Bias is still murky. This is the most challenging part — We need precise, concise and actionable best practices. May be things like Model Bias Score Card or similar mechanisms are need to be defined and transparently available, which I hope NIST will work on. May be I can help.
6. Bias Mitigation is still untouched — Whom do we call when we see indications of negative impacts ? And, what do we do then ? What are the ways of increasing fairness ? Another challenge

Now, onto 1st impressions …

They have a few very informative and detailed diagrams.

As I mentioned earlier, the guidance part is a little sketchy. It is also a little verbose and could use a tad more organization. Of course this is a draft and the guidance will be concise only after the other parts are in good shape.

In short …

In short, good work on framing the issues of Bias in AI with a broad stroke beyond the normal statistical and computational vectors. But need more work on how to actually being the ideas to practice — Best Practices, Mitigation Strategies et al. Probably that requires additional input from practitioners from different industries…