I recently had the pleasure of returning to my alma mater attending the CMU-K&L Gates Conference on Ethics and AI. It was refreshing to hear from “expert”[1] leaders across domains ranging from computer science and finance to public policy and community organizations. Particularly as the uses and implications of AI continue to be developed and understood, it is an important first step in introducing a more multi-faceted and socio-technical approach in thinking about AI.

Some key takeaways I got from the conference:

• User input and engagement should be considered as important, if not more, as key inputs into every step of the AI lifecycle when considering the ethical implications of AI

• Rather than thinking about the more long-term existential (i.e., often the “doom-and-gloom” questions of AI), how about actually asking the grassroots community what it really needs right now? No shortage of projects (including those with AI) can be discovered from there

• Data…and its subsidiaries (“Data, LLC”)[2] should honestly be treated and assessed more rigorously as assets and liabilities are done in financial institutions. Frameworks like EU’s GDPR on data privacy or the OCC’s SR 11-7 on risk management frameworks in finance can be good starting points in creating such standards. The AI Act (2023) preliminarily passed recently by the EU also serve as a good reference for overall AI governance framework.

As mentioned above, the question of AI will extend across disciplines and generations, requiring take up from every walk of life. However, for my own academic purposes and interest, I specifically tie this back to implications on finance. Particularly, in their keynote on ethics and computational technologies, Manuela Veloso, head of AI Research at JP Morgan, discussed the role of AI/ML in finance. In particular, they mentioned the rise in the use of synthetic data in model development. The generation of synthetic data via AI is intriguing in that it can “auto generate ground truth” while keeping its comparative performance cost low versus traditional methods, i.e., having the costly manual human labeling of data, particularly when generating sufficient amount of training data for models. For asset pricing and finance, this may help generate and increase the number of observations, a limitation to the adoption of machine learning in finance models as identified by Israel, Kelly, and Moskowitz (2020).

Of course, with AI come increased importance and value of data, good data and training data, to be precise, something that was reiterated in multiple times throughout the conference. The privilege of data ownership should also be coupled with the responsibility of the enrichment of the data and research subjects impacted (or at minimum, not leave the subjects worse off nor perpetuate existing societal issues than before the intervention). And a key driver of this lies in the incentives and value of such data. How this value is quantified is for sure worthy of future research.

[1] A term, suggested during the conference, to be reimagined, broadened in its definition to more fully reflect the “contextual expertise” that is out in the world especially as we think about the interactions between humans and AI.

[2] Ironically, of course.