Hacker News Reader

Comment by pjmorris | Hacker News Reader

Dr. Casey Fiesler recently tweeted

> The @huggingface GPT detector works very well on ChatGPT-created text. I ran 5 student essays and 5 ChatGPT essays for the same prompt through it, and it was correct every time with >99.9% confidence.

How about adding a %human/%GPT statistic to posts and comments?

View on HN

I would argue that this test isn't particularly informative. Given 5 attempts and 5 successes, even though the point estimate of accuracy is 1, the 95% CI ranges from 0.48 to 1:

    > binom.test(5,5,0.5)

     Exact binomial test

    data:  5 and 5
    number of successes = 5, number of trials = 5, p-value = 0.0625
    alternative hypothesis: true probability of success is not equal to 0.5
    95 percent confidence interval:
     0.4781762 1.0000000

In other words, we don't have enough data in that small sample to reject the possibility that the model is 50% accurate, much less 99.9% accurate.

I think the message was claiming something else, specifically that each classification was given a score of how confident the model was in the answer and the answers were given 99.9%+ in those cases.

See the app: https://huggingface.co/openai-detector/ - it gives a response as % chance it's genetic or chat bot.

Seems to have major biases against who knows what sentence structures. Even without trying to make it say fake, some of my messages and text I write in it are pretty confident I'm GPT-2...

With 5 samples, we have no way to assess whether the app’s 99.9% self-assessment is remotely well calibrated. (As noted above, 5/5 is also consistent with a model that is right 50% of the time.)

Not that the HN team isn’t capable of solving this, but I think that would be a pretty significant increase in server cost to run every comment through a classifier.

We don't have millions of comments here. I think a $5 additional vcpu can solve it easily.

My gut tells me there is no way a single cpu model can quickly spit out useful scores here

The GPT detector requires a significant amount of text for accuracy. Works well on student essays, not so much on short comments.

https://news.ycombinator.com/item?id=32447928 is marked as nearly 100% fake, whereas I can assure you it was written by a human.

Maybe I was just unlucky with the comment I tried it with (took the longest one I saw in my history), but I don't think I would have liked seeing it either removed or spat at for being considered as "AI generated"...

The detector also thinks this comment is fake. Seems influenced by flavors of mistakes.

Idiomatic ones. Spelling ones. Grammar. All non-native speakers will easily get flagged. Does not look spot-on for now. Checked all those assertion live-typing on the demo. 0.09% real.

It would work, for like 1 day

HN doesn't update anything. That will obviously never happen.

Im sure we will start seeing reddit /twitter bots that reply to users that have high gpt confidence.

We update things all the time! But we prefer to be subtle about it.

Do you have an announce list somewhere for significant or interesting changes?

Reply on news.ycombinator.com