Scientific Text Analysis with Robots applied to observatory proposals

TL;DR

This study investigates the impact of AI language models like ChatGPT on the scientific proposal review process at ESO, revealing biases and limitations in AI-generated assessments and references.

Contribution

It provides empirical insights into AI's influence on proposal grading, highlighting biases and accuracy issues relevant for future adoption in observatory reviews.

Findings

ChatGPT-adjusted proposals receive lower grades.

ChatGPT 3.5 often provides incorrect references.

AI tends to give higher scores and prefers its own proposals.

Abstract

To test the potential disruptive effect of Artificial Intelligence (AI) transformers (e.g., ChatGPT) and their associated Large Language Models on the time allocation process, both in proposal reviewing and grading, an experiment has been set-up at ESO for the P112 Call for Proposals. The experiment aims at raising awareness in the ESO community and build valuable knowledge by identifying what future steps ESO and other observatories might need to take to stay up to date with current technologies. We present here the results of the experiment, which may further be used to inform decision-makers regarding the use of AI in the proposal review process. We find that the ChatGPT-adjusted proposals tend to receive lower grades compared to the original proposals. Moreover, ChatGPT 3.5 can generally not be trusted in providing correct scientific references, while the most recent version makes a better, but far from perfect, job. We also studied how ChatGPT deals with assessing proposals. It does an apparent remarkable job at providing a summary of ESO proposals, although it doesn't do so good to identify weaknesses. When looking at how it evaluates proposals, however, it appears that ChatGPT systematically gives a higher mark than humans, and tends to prefer proposals written by itself.