Interesting methodology. The results for Pangram are surprisingly good.<p>These tools are definitely not 100% perfect, which is the primary complaint used to dismiss them. However the error rate is also getting impressively low.<p>In most cases I see socially and online, there is a high suspicion that the content is AI generated before someone thinks to submit it to Pangram. It’s used on-demand as a tool to confirm suspicions. I have seen several cases where Pangram had some false negatives where the content was judged to be likely human written but the author later admitted it was written by an LLM.<p>Pangram is very interesting in the context of Substack because the platform was a target for lazy AI newsletters. People realized they could start 10 (or maybe many more) newsletters and spend only a few minutes getting ChatGPT to write posts for them. Starting an AI generated substack and trying to get paid subscribers for it was becoming one of the popular ways to use AI to try to get a little cash. Having a tool that makes it a little bit harder, at least until the LLMs get good enough to evade it, was important for the platform.
> I have seen several cases where Pangram had some false negatives<p>Yep, Pangram says they think false positives are way worse than false negatives, so if their AI isn't sure enough they prefer a "not ai" output to an "is ai" output.
On the other hand, see <a href="https://freddiedeboer.substack.com/p/i-wouldnt-say-pangram-is-broken-but" rel="nofollow">https://freddiedeboer.substack.com/p/i-wouldnt-say-pangram-i...</a> where Freddie deBoer shows some instances where Pangram is definitely giving wrong or misleading results.<p>(Somewhat-plausible-to-me explanation: It's looking for various stylistic features; older writing very rarely has the most AI-like features, or perhaps almost always has some non-AI-like features that outweigh whatever signs of AI-ness might be there. Present-day writers are more likely to get wrongly flagged as AI.)
> Pangram boasts a false positive rate of 1 in a 10,000. That is, if Pangram says a block of text is AI there is only a one in ten thousand chance that it was written by a human.<p>That'd be if they had a <i>false discovery rate</i> of 1/10,000.<p>If for instance:<p>* 100,000 samples are tested<p>* 100 of which are AI-generated, the rest human-written<p>* Pangram flags 50 of the AI-generated samples (true positives)<p>* Pangram also flags 10 human-written samples (false positives)<p>Then the FPR is 1 in 10,000, but the chance that a flagged sample isn't actually AI (FDR) is 1 in 6.
So the result here is that you fed a few thousand segments of old books into Pangram and zero of them were classified as AI. Seems impressive?
I had to read this sentence twice:<p><pre><code> Right now many niche book communities are facing an onslaught of unlabeled AI works, mine included.
</code></pre>
On first reading, I thought the author meant that they themselves had submitted 'unlabeled AI works' to one or more niche book communities.
Pangram is amazing. I hope they eventually provide something like a browser plugin that scans text for being AI generated as you read stuff. I’d pay good money for that.
Note that any good binary classifier (like "this is AI writing" for Pangram) should maximize the following four probabilities:<p>1. <i>P(actually true | predicted true)</i> (= true positive rate, recall, sensitivity)<p>2. <i>P(actually false | predicted false)</i> (= true negative rate, specificity)<p>3. <i>P(predicted true | actually true)</i> (= positive predictive value, precision)<p>4. <i>P(predicted false | actually false)</i> (= negative predictive value)<p>If you try to minimize the share of false positives, you are only maximizing 1.
I find it's surprising that people doubt the effectiveness of Pangram when I feel like its job has gotten easier ever since Claude models got RLHF to death with its bizarre punchy negatives and tortured analogies, making AI writing much more noticeable than the GPT 4o era.