Interesting how ChatGPT had a vastly higher tendency to "overcorrect" sentences (relative to Grammarly and GECToR), completely changing sentence structures to produce proper sentences rather than fixing specific grammar mistakes.
I wonder how the results of this study would have differed if the team looked at GPT-4?
It seems they used the web UI to do this. I wonder how it would have been if they did use the API and changed some parameters to reduce this modification effect.
honestly, seems like a paper without sufficient rigor
Would love to see the differences vs the Language Tool offering.