Written by Helen Brooks, NERC-funded CSaP policy intern (March-July 2018)
This session highlighted the potential opportunities provided by data science and AI, and how these might influence or aid future policymaking. Potential negative effects of data science and AI were also highlighted.
You can listen to the recording here:
The session was chaired by Hetan Shah (Director of the Royal Statistical Society), who was joined by three panel members:
- Professor Helen Margetts, Professor of Society and the Internet, University of Oxford and Director, Public Policy Programme, Alan Turing Institute for Data Science and Artificial Intelligence.
- Dr Emily Shuckburgh, Head of Data Science Group and Deputy Head of Polar Oceans Team, British Antarctic Survey.
- Dr Karen Croxson, Head of Research & Deputy Chief Economist at the Financial Conduct Authority
Initially, the panel highlighted the potential opportunities provided by data science and AI.
Prediction: Data science will likely provide new ways of predicting the future. This could be used to forecast demand for services (e.g. schools, hospitals or shopping), or to assess which individual person is going to need to be put in the care system or the justice system. This would allow targeted planning of services and intervention and, arguably, better provision of resources.
Modelling: Data science and AI could help us better understand the impact of future policies, before we implement them.
Data use and analysis: At present we produce lots of data. AI would allow us to tap into more of this data than we are currently able to. This analysis could then be used to improve models.
However, it was also recognised that these opportunities also bring issues which need to be addressed.
Predictability: Data makes us easier to predict, which means we are more exploitable. This is a particular concern for the most vulnerable members of society.
Ethical (Data Sharing): Who will the data be shared with? How will the data be used? Once collected, the data could become de-anonymised and could be used for purposes for which it was not originally intended.
Ethical (Data Use): We may end up being able to predict things on an individual level. If we can reasonably assume that a child may be taken into care, what do we do? This likely has implications for individual circumstances. Similarly, we don’t currently know if algorithms could learn to collude data (e.g. to set prices higher than would happen normal).
Transparency/reproducibility/interpretability: Are the data and the model open access? We need good data and models to be made available. This might require consideration of things like new publishing practises (e.g. a way to publish a model). By understanding the model, we will be more likely to understand why we produce these predictions, and the uncertainty associated with them.