Hello Everyone,
I would like to receive your viewpoints on ethical use of AI Language Models in research.
Here are one aspect that I would like to draw your attention:
AI models are good at searching information that already exists, a researcher need not to waste so much time on reading unnecessary literature, a AI model saves lot of time, energy and resources by giving you the results in seconds which otherwise might take you to read -a whole book, few research papers, blogs, etc., and this will eat up so much time. Isn't it fair to use AI language models to search for information and minor questions that pops up during an research journey? In my opinion, it is fair, it is just like using an calculator to do calculations quickly. What are your opinions? and What should be the most ethical use of AI Language Models to do internet search related to minor research questions? Whether we should give an AI model an credit, citation, or acknowledgement and How?
Kindly provide your opinions to enrich my understanding.
Currently language model AI are basically plagiarists which mash together what humans have previously wrote... I would suggest they be treated as high-level versions of search engines, so using them to do information gathering is fair, though they should probably be included in the "Methods" section as tools.
Writings generated by AI should definitely be banned from research, though, and should not be treated as reliable and citable resources, because:
1. They're potentially inaccurate
2. They do not reference resources themselves, so incorporating them into the world of research is detrimental to the consistency of research ethics
The way it is trained is the way it replies. It can be trained for good purpose or malicious
Sahil Sharma ,
You have two different topics that are very different
Scientific writing
Chat GPT(Bing AI) and BARD (Google's) have no business in scientific writing as Shiran Yuan stated, the probability of plagiarism is >0 and it is unreliable (see comment below). Though, I do draw a distinction between search engines and these NN models.
Research
These models are trained at scale through reinforcement learning. That means that while trained, the scale of training makes for poor supervision on the quality of the resources. For more information on the training that ChatGPT uses see[1] (I also have posted some comments on the workings of ChatGPT in[2] ). Now, the scale is enough for poor supervision and not enough to be a full scale search engine. Google's scope of documents falls on the billions of documents. Far above of the capacity for NNs. That said, NNs have been use as part of search engines for a while doing query expansion and image to text relations.
In terms of the quality of the output that ChatGPT generates, they are of poor quality due to the absence of fact checking of the sources. Further, due to the nature of the NN training, the NN will be more likely to have bias than a search engine. It would not be able to return results in the way that Google returns thousands of links per query (at least not as currently being designed and without external help from a search engine).
In summary, it is too early to even rely on this technology.
References
[1]Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C. L., Mishkin, P., ... & Lowe, R. (2022). Training language models to follow instructions with human feedback. arXiv preprint arXiv:2203.02155.
[2]https://www.researchgate.net/post/How_is_Chat_GPT_created_and_how_powerful_it_is
Sin duda hay que ponerlo en prueba y como investigador relacionar con resultados o conocimientos establecidos. lamentablemente las respuestas son aparentemente buenas, pero sin evidencia científica....
- Badges
- Science topic