how would LLMs help with uncovering "bad studies"? I don't think LLMs are yet sophisticated enough to figure out potential plagiarism or Adobe Photoshop copy-and-paste (as mentioned in the article above).
Presumably the same can be done for tabular data and genomic or "omics" data. At the very least statistical techniques can be used. I imagine high throughput imaging modalities will be the main target.
That being said the only way the -omics data will have utility is via AI models which are trained on some task... which present their own problems..
You don't need a language model. The effort is in collecting and chopping the data to find snippets to compare. NNs can help with that.
Newer fraud will probably use generative diffusion AIs to make "realistic western plots" on demand .. there's probably a paper in that!