Will The Last Human Researcher Please Close the Door After Them?
The peer review process in science and technology is the bedrock of research. Without it, anyone could publish any old rubbish. We see this with paper mill journals, and they often say that they can publish papers within just a few weeks of submission. Overall, there is virtually no credibility in publishing in these journals, as the peer review process is often so weak. But there are problems. Firstly, there is no payment for the peer review process, and, secondly, a good review often takes a considerable amount of time. One must also worry that the peer reviewer is not the best person in the world to review a specific paper.
Now, a paper outlines that AI could be playing an increasing role in the peer review process [1]:
The authors used a GPTZero LLM detector to detect the presence of AI assisted reviews for the ICLR (International Conference on Learning Representations) 2024 conference. For this, they split their work into three main experiments: to determine the prevalence of AI-assisted reviews, analyse the effect of AI-assisted reviews, and determine the effect of AI-assisted reviews on the acceptance rate.
The LLM detector provides a range of methods for detecting AI generation. This includes using words which are more common in ChatGPT than in human written words. These words include underscores, delves and bolster:
Overall, they found that the ICLR 2024 conference showed an increase in 15.8% in likely AI generated reviews:
When it came to scoring, the conference allowed the grading with:
- 1: strong reject.
- 3: reject, not good enough.
- 5: marginally below the acceptance threshold.
- 6: marginally above the acceptance threshold.
- 8: accept, good submission
- 10: strong accept, should be highlighted at the conference.
When it came to the effect of an AI-assisted review related to the acceptance of the paper, in (A) they looked at the effect of AI-assisted reviews against human reviews. For this they found that there was a considerable variation on those papers with very weak and very strong human reviews, but, in general, the papers in-between these boundaries were generally more positive than the human reviews. This showed an increase of 4.9% in increased acceptance rates for AI assisted reviews for those between the boundaries, and a 3.1% increase overall in AI-assisted reviews.
Conclusions
It will be a sad world of research if we move towards an AI-driven peer review process, especially in using an LLM approach. While ChatGPT is good at knowledge-based approaches, it struggles in the analysis of creativity and in creating and analysing novelty. On the flip side, we could see a rise in AI-driven papers, which contain no real novelty but will match well with the AI review process.
Many, too, are observing the increasing number of AI reviews [here]:
References
[1] Giuseppe Russo Latona, Manoel Horta Ribeiro, Tim R. Davidson, Veniamin Veselovsky, Robert West, The AI Review Lottery: Widespread AI-Assisted Peer Reviews Boost Paper Scores and Acceptance Rates https://arxiv.org/abs/2405.02150
