> Research is supposed to be new ideas. If much of your research paper can be written by AI, I call into question whether or not it represents actual research.
One would hope the authors are forming a hypothesis, performing an experiment, gathering and analysing results, and only then passing it to the AI to convert it into a paper.
If I have a theory that, IDK, laser welds in a sine wave pattern are stronger than laser welds in a zigzag pattern - I've still got to design the exact experimental details, obtain all the equipment and consumables, cut a few dozen test coupons, weld them, strength test them, and record all the measurements.
Obviously if I skipped the experimentation and just had an AI fabricate the results table, that's academic misconduct of the clearest form.
I am not an academic, so correct me if I am wrong, but in your example, the actual writing would probably only represent a small fraction of the time spent. Is it even worth using AI for anything other than spelling and grammar correction at that point? I think using an LLM to generate a paper from high level points wouldn't save much, if any, time if it was then reviewed the way that would require.
My brother in law is a professor, and he has a pretty bad opinion of colleagues that use LLMs to write papers, as his field (economics) doesn't involve much experimentation, and instead relies on data analysis, simulation, and reasoning. It seemed to me like the LLM assisted papers that he's seen have mostly been pretty low impact filler papers.
> I am not an academic, so correct me if I am wrong, but in your example, the actual writing would probably only represent a small fraction of the time spent. Is it even worth using AI for anything other than spelling and grammar correction at that point? I think using an LLM to generate a paper from high level points wouldn't save much, if any, time if it was then reviewed the way that would require.
Its understandable that you believe that, but its absolutely true that writing in academia is a huge time sink. Think about it, the first thing your reviewers are going to notice is not results but how well it is written.
If its written terribly you have lost, and it doesnt matter how good your results are at that point. Its common to spend days with your PI writing a paper to perfection, and then spend months back and forth with reviewers updating and improving the text. This is even more true the higher up you go in journal prestige.
One would hope the authors are forming a hypothesis, performing an experiment, gathering and analysing results, and only then passing it to the AI to convert it into a paper.
If I have a theory that, IDK, laser welds in a sine wave pattern are stronger than laser welds in a zigzag pattern - I've still got to design the exact experimental details, obtain all the equipment and consumables, cut a few dozen test coupons, weld them, strength test them, and record all the measurements.
Obviously if I skipped the experimentation and just had an AI fabricate the results table, that's academic misconduct of the clearest form.