This came up with my step mom who writes fan ficton on several other websites. All her stories recently keep tripping AI detectors when she is not using it! She has a doctorate in english and teaches english and technical writting at a state college. To test the system she had other professors submit stories and they to were flagged. While her department is struggling with the school issue, are we left with knowing how to write means a claim of AI?
Many AI detectors are not that good. See (for instance): https://lawlibguides.sandiego.edu/c.php?g=1443311&p=10721367
Interesting quote:
False positive rates vary widely. Turnitin has previously stated that its AI checker had a less than 1% false positive rate though a later study by the Washington Post produced a much higher rate of 50% (albeit with a much smaller sample size). Recent studies also indicate that neurodivergent students (autism, ADHD, dyslexia, etc…) and students for whom English is a second language are flagged by AI detection tools at higher rates than native English speakers due to reliance on repeated phrases, terms, and words.
Or see: http://marcushere.substack.com/p/is-turnitins-ai-detector-accurate
Thirty-four of the fifty papers, or 68%, scored below 20% AI probability. That's the outcome you'd want: the system largely treating human writing as human.
Nine papers, 18% of the sample, scored between 20% and 50% AI probability. Seventeen papers total, or 34%, scored above 20%.
Nine papers - 18% - scored above 50% AI probability. Three papers, 6% of the sample, scored above 80%.
Let me be explicit: those nine papers scoring above 50% AI probability were written entirely by humans, verified through their pre-2022 provenance, and already reviewed by faculty members who found nothing anomalous about them. Turnitin's detector assessed them as more likely AI-generated than not.
This is a problem for universities that rely on such products, and a problem for students as well. There are numerous reports out there of students whose work was questioned or outright rejected, sometimes with loss of credit or academic repercussions, who subsequently demonstrated that their work was their own. But some of those required them to retain an attorney and sue. Even if the university was forced to pay attorney fees, that means a waste of time and money and potentially lingering disparagement.
Meanwhile, other reports (e.g. https://teaching.temple.edu/sites/teaching/files/media/document/Evaluating%20the%20Effectiveness%20of%20Turnitin%E2%80%99s%20AI%20Writing%20Indicator%20Model.pdf) show that it's reasonably possible for a person to edit AI-generated work in such a way as to sidestep AI detectors. I've seen anecdotal evidence (no link) that people are having success by using a second AI and a carefully tailored prompt to 'humanize' the output of a first AI.
One last quote, from https://www.bestcolleges.com/news/analysis/testing-turnitin-new-ai-detector/
(which had better news on false positives, though it wasn't entirely convincing):
According to a recent release from [Turnitin], about 11% of those papers indicated at least 20% AI writing present, while 3% indicated more than 80%.
These figures correspond almost exactly to what Turnitin found in July 2023, only a few months after the launch. At the time, Turnitin had reviewed more than 65 million papers.
Assuming the tool is reliable, these data suggest student behavior hasn't changed much over the past year.
Maybe, or maybe it suggests the Turnitin developers think they know what the rate is and are tailoring the tool to match the rate they expect. If they're wrong, problems ensue.
TLDR: it's a mess, and it's going to stay a mess for quite a while. The AIs adapt to the detectors, the detectors adapt to the AIs, and it's often not clear whether enough is being done (or can be done) to prevent false positives.