The need for policy to regulate the development and deployment of AI applications is regularly discussed and, indeed, there is ongoing surveillance work which describes and catalogues AI policies and strategies.  However, despite the need for ongoing scrutiny around the ethical approaches around the creation of AI tools, little cohesion and understanding exists resulting in a multitude of tools from many sources and disciplines but no gold standard.    Indeed, it has been suggested that there are three core principles which can aid the understanding of the ethical implications of AI tools:  impact (to avoid harm and benefit society), justice (to promote fairness and equity) and autonomy (to allow access to the use and modification of AI).

Thus, while there is much excitement at the potential for AI to benefit many areas of decision-making, such tools are frequently deployed in ways that lack transparency despite documented risks that they may increase systemic inequities. There is an urgent need to understand more about when AI is (or can be) used for decision-making, what the potential benefits and harms are, and how we can equip people with the tools and knowledge they need to understand these issues more effectively.

A partnership of the Campbell Collaboration and EPPI Centre have mapped the evidence on AI and equity in consultation with the American Institutes for Research (AIR). The online map is browsable by clicking the image below.

The map was first put online in January 2024. It was constructed by searching 19 databases and downloading the results into EPPI Reviewer. 34, 541 records were identified, of which 8,485 were found to be duplicates. The reminder were screened automatically by GPT-4 with human-validated sensitivity of 95% and specificity of 100%. 6,628 records remained which were then ‘mapped’ using a pre-defined coding tool – again by GPT-4. Human validation of the mapping found that 86% contained no errors and an additional 12% contained only minor errors, so that automation was deemed to be sufficiently reliable for the map to be placed online.

The map was then updated in November 2024. An additional 10,380 records were identified by re-running the original searches for records published since December 2023. GPT4o is now integrated within EPPI Reviewer, and all mapping tasks were undertaken in this platform. Bearing in mind that this is a new language model, and since the feature can be run within the EPPI Reviewer environment, we decided to re-screen and map all records - both from the original and update searches. We first developed and validated prompts to screen records for inclusion. Sensitivity was found to be 100% (based on the manual checking of 221 records), meaning that all, or close to all, the records that should be included are indeed included in the map. We then developed and validated prompts to map the records into the categories shown on the left. The keywording of a total of 175 records was manually checked. 13 records were found not meet the eligibility criteria for the map and were removed from analysis, giving the map an assessed specificity of 93%. Of those remaining, five had titles only, and no abstract, so were excluded from analysis. Of the remaining 157, 141 (90%) were assessed as being completely correct, not missing any categories, and not receiving any erroneous categories. Four items (3%) had major errors, which might result in them being 'lost' in the map due to no equity code being assigned. The remaining (7%) of items had minor errors, which would have been the incorrect assignment or non-assignment of a marginally relevant category, but not something which would lead to the indexing of the record being misleading in a significant way.

This map can be cited as: Thomas J, Stokes G, Stansfield C, Dehdarirad H, Mudiyanselage I, Welsh V, Moy W (2024) A high-level map of the research landscape covering AI and equity. EPPI Centre software, University College London, London.