For a machine, those would be nice works of association, but that is not unheard of, as there are known paper generators such as SCIgen. The occasional typos work against a clear diagnosis.
Now I wouldn't mind if an AI posed a clear and interesting question. But those are not clear (to me), and if those questions are indeed machine-generated, this means that someone only needs to turn a knob to increase MSE's junk inflow far beyond manageability.
We may need increased machine assistance. Are there tools in place or in development that could automatically send posts like those to a review queue?
Update: Another post (#2810926, now deleted) (Screenshot) that might have been machine-generated has been identified in the comments. This time it is an answer.
Until a bot attack or a publication in the spirit of SCIgen occurs, we might remain in the dark about whether those posts have been machine-generated or not.
Currently, what matters are the characteristics that help those posts pass existing automated detectors and superficial human reviews.
Characteristics that should be machine-testable:
- No block-displayed math, very few inline math snippets
- Frequent references to, or quotes with superficially relevant keywords from Wikipedia / ArXiv / other sites known to host math content
- No other hyperlinks
- Generous use of inline markup (bold or italic portions)
- Many blocks (paragraphs/quotations/list items), but all short
This all serves to distract from the post's thin content while creating an appearance of thorough composition. I'd like to see some scoring mechanism work against such pretension.
Further characteristics, perhaps not machine-testable, but annoying:
- In questions: Typos and poor grammar except when reciting
- In answers: Telegram style only
- No complete thought in the non-recited text portions
- Following the references yields no relevance.
I find this use of fake references particularly frustrating because it multiplies the waste of every dedicated reviewer's time and effort.
Can we get the existing detectors sharpened, or new detectors introduced, to take into account at least the machine-testable features listed above?
Update: I suppose that there is some scoring tool in place already, perhaps a linear classifier. It would apply a number of metrics, dot-multiply the resulting vector with a vector of weights, and obtain a score. Interestingly, adding more metrics often helps to better separate good examples from bad ones. Therefore, this is mostly a matter of adding new metrics and refining existing ones. I'd propose to:
- Count links to Wikipedia and ArXiv separately from other hyperlinks
- Measure density of block-displayed math
- Measure markup density outside math, blockquotes, and code sections
and append those to the other existing metrics. Then recompute the optimal separating hyperplane (i.e. weights and threshold) from training data. Make sure to include gibberish posts and cross-validate. I'd be interested in the results.