3 ms·

I wouldn't be surprised if the headline is accurate, but AI detectors are widely understood to be unreliable, and I see no evidence that this AI detector has ov

I wouldn't be surprised if the headline is accurate, but AI detectors are widely understood to be unreliable, and I see no evidence that this AI detector has overcome the well-deserved stigma.

SoftTalker 11mo ago
In particular, conference papers are already extremely formulaic, organized in a particular way and using a lot of the same stock phrasings and terms of art. AI or not, it's hard to tell them apart.
- Jensson 11mo ago
  Its the reviews that were found to be AI, not the papers themselves. The papers were just 1% AI according to the tool, so it seems to work properly. > AI or not, it's hard to tell them apart. Apparently not for this tool.
maxspero 11mo ago
Co-founder of Pangram here. Our false positive rate is typically around 1 in 10,000. https://www.pangram.com/blog/all-about-false-positives-in-ai-detectors https://www.pangram.com/blog/all-about-false-positives-in-ai.... We also wanted to quantify our EditLens model's FPR on the same domain, so we ran all of ICLR's 2022 reviews. Of 10,202 reviews, Pangram marked 10,190 as fully human, 10 as lightly AI-edited, 1 as moderately AI-edited, 1 as heavily AI-edited, and none as fully AI-generated. That's ~1 in 1k FPR for light AI edits, 1 in 10k FPR for heavy AI edits.
- Fuzzwah 11mo ago
  Give your final sentence a re-read there....
  - maxspero 11mo ago
    Thanks, fixed.
Jensson 11mo ago
The conference papers were 1%, peer reviews 20%, is there another reason for that big difference than more of the peer reviews being AI generated than the papers themselves? We can't use this to convict a single reviewer, but we can almost surely say that many reviewers just gave the review work to an AI.
- bjourne 11mo ago
  Peer reviews are much shorter than papers. A bad review may only be six to eight sentences along the lines of "Math is confusing" and "Too little related work". 20% AI generated seems high if for no other reason than writing a bad review takes very little effort.

I wouldn't be surprised if the headline is accurate, but AI detectors are widely understood to be unreliable, and I see no evidence that this AI detector has overcome the well-deserved stigma.

SoftTalker 11mo ago

In particular, conference papers are already extremely formulaic, organized in a particular way and using a lot of the same stock phrasings and terms of art. AI or not, it's hard to tell them apart.

Jensson 11mo ago
Its the reviews that were found to be AI, not the papers themselves. The papers were just 1% AI according to the tool, so it seems to work properly. > AI or not, it's hard to tell them apart. Apparently not for this tool.

maxspero 11mo ago

Co-founder of Pangram here. Our false positive rate is typically around 1 in 10,000. https://www.pangram.com/blog/all-about-false-positives-in-ai-detectors https://www.pangram.com/blog/all-about-false-positives-in-ai.... We also wanted to quantify our EditLens model's FPR on the same domain, so we ran all of ICLR's 2022 reviews. Of 10,202 reviews, Pangram marked 10,190 as fully human, 10 as lightly AI-edited, 1 as moderately AI-edited, 1 as heavily AI-edited, and none as fully AI-generated. That's ~1 in 1k FPR for light AI edits, 1 in 10k FPR for heavy AI edits.

Fuzzwah 11mo ago
Give your final sentence a re-read there....
- maxspero 11mo ago
  Thanks, fixed.

Jensson 11mo ago

The conference papers were 1%, peer reviews 20%, is there another reason for that big difference than more of the peer reviews being AI generated than the papers themselves? We can't use this to convict a single reviewer, but we can almost surely say that many reviewers just gave the review work to an AI.

bjourne 11mo ago
Peer reviews are much shorter than papers. A bad review may only be six to eight sentences along the lines of "Math is confusing" and "Too little related work". 20% AI generated seems high if for no other reason than writing a bad review takes very little effort.