These are quite compelling responses for the prompts. I have a feeling that widespread falsifying of written works will become a huge problem for colleges very soon (probably already the case but small enough to be under the radar). This will become the next iteration of "just Google it." With concerns about social media controlling what's allowed to be said and by whom. Soon, an algorithm will be able to speak to you with infinitely better capabilities to manipulate and entrance. Amazing and scary times we're living in.
no. and in fact i'd propose that models like gpt-3 can potentially exhibit catastrophically self-defeating effects. consider that nlp models are trained on a corpus of text, much of which is found online. but in the early stages, it is presumed that much of the text content is in fact human-generated, thus providing the model with an essentially human prior to use as its 'ground truth' for natual language parsing and generation.
ok now fast forward to a point in time when the amount of text content online that is generated by gpt-3-like bots is non-trivial, or perhaps even majority. now when the 'ground truth' corpus on which the model is trained is collected, it can no longer be assumed to still maintain a 'human' prior. rather, it will have something more akin to a 'self' prior, with the original 'human' portion dwindling away over time (assuming the rate of generation of bot content out-paces human content in a monotonically increasing way).
so what happens next to the models? well either humans develop better methods, or my best guess is that eventually the models' ability to generate content deteriorates asymptotically towards a very non-diverse language. in a similar process to hebbian learning, such bots released in the wild and used to generate content that they (future generations of) themselves use to train on, generate a large-scale positive feedback loop. and tiny random fluctuations in the models weights eventually grow larger and more pronounced over time, resulting in incrementially more uniform content. potentially to the point where eventually the models only spit out "lorem ipsum dolor sit amet" on continuous repeat. (obviously not actually, but i think you get the point. it would probably be more like "the cat ran and the cat ran and the cat ran...")
could try to combat this by working to identify and only train on 'human content', and that may or may not be effective, insofar as model develpers will generally aim to have their bot content be indistinguishable from human content. a bit like GANs. at that point though, it becomes a little harder for me to reason on/make predictions about
It will not be possible to detect a bot from the isolated peace of text in the near future. Only interactive tests will do the job, like those in the Blade Runner movie.
