Submissions:2025/Hidden Vandalism: Characteristics and Detection via Machine Learning
This submission has been noted and is pending review for WikiConference North America 2025.
Title:
- Hidden Vandalism: Characteristics and Detection via Machine Learning
Type of session:
- Lecture (15-30 min)
Session theme(s):
- Future of Wikipedia
Abstract:
It is remarkable -- and perhaps taken for granted -- that Wikipedia is not replete with vandalism. Most of Wikipedia can be edited by any Internet user, even without creating an account, so where is the onslaught of spam, test edits, and juvenile humor? In fact, the onslaught is there, but much of it is blocked by edit filters or immediately removed by bots like ClueBot NG. What survives is usually caught by human editors monitoring recent changes or article watchlists. Nevertheless, a small percentage of vandalism slips through and can go undetected for months or even years, becoming "hidden vandalism." It is not known how many articles contain hidden vandalism, nor are there cost-effective means for detecting it. In this session, I will explore whether large-language models (LLMs) can help address these issues. While commercially available LLMs achieve some success "out of the box," they are expensive to apply at scale and suffer from a high rate of false positives. By contrast, I will demonstrate that fine-tuning significantly improves detection performance, to the extent that even "small" LLMs become effective. These "small" LLMs incur relatively low computational costs and can be applied at scale to the entirety of English Wikipedia. I will also present statistics on 2000+ instances of vandalism that were identified by a fine-tuned model, with the aim of defining the scope of the hidden vandalism problem and informing preventative measures. Possible implications for future editing practices on Wikipedia will be discussed.
Author name(s):
Wikimedia username(s):
- MaxwellMolecule
Affiliated organization(s):
Estimated length of session
- 20-25 minutes
Will you be presenting remotely?
- I will present in-person
Okay to livestream?
- Livestreaming is okay
Previously presented?
Special requests: