←back to thread

Do AI detectors work? Students face false cheating accusations

(www.bloomberg.com)
427 points JumpCrisscross | 10 comments | 20 Oct 24 17:26 UTC | HN request time: 0s | source | bottom
Show context
greatartiste ◴[21 Oct 24 06:46 UTC] No.41901335[source]▶
For a human who deals with student work or reads job applications spotting AI generated work quickly becomes trivially easy. Text seems to use the same general framework (although words are swapped around) also we see what I call 'word of the week' where whichever 'AI' engine seems to get hung up on a particular English word which is often an unusual one and uses it at every opportunity. It isn't long before you realise that the adage that this is just autocomplete on steroids is true.

However programming a computer to do this isn't easy. In a previous job I had dealing with plagiarism detectors and soon realised how garbage they were (and also how easily fooled they are - but that is another story). The staff soon realised what garbage these tools are so if a student accused of plagiarism decided to argue back then the accusation would be quietly dropped.

replies(14): >>41901440 #>>41901484 #>>41901662 #>>41901851 #>>41901926 #>>41901937 #>>41902038 #>>41902121 #>>41902132 #>>41902248 #>>41902627 #>>41902658 #>>41903988 #>>41906183 #
1. ClassyJacket ◴[21 Oct 24 07:03 UTC] No.41901440[source]▶
How are you verifying you're correct? How do you know you're not finding false positives?
replies(1): >>41901515 #
2. Etheryte ◴[21 Oct 24 07:15 UTC] No.41901515[source]▶
Have you tried reading AI-generated code? Most of the time it's painfully obvious, so long as the snippet isn't short and trivial.
replies(1): >>41901632 #
3. thih9 ◴[21 Oct 24 07:36 UTC] No.41901632[source]▶
To me it is not obvious. I work with junior level devs and have seen a lot of non-AI junior level code.
replies(1): >>41901914 #
4. llmthrow102 ◴[21 Oct 24 08:28 UTC] No.41901914{3}[source]▶
You mean, you work with devs who are using AI to generate their code.
replies(4): >>41901987 #>>41902164 #>>41902461 #>>41904228 #
5. ◴[21 Oct 24 08:44 UTC] No.41901987{4}[source]▶
6. ben_w ◴[21 Oct 24 09:15 UTC] No.41902164{4}[source]▶
Not saying where, but well before transformers were invented, I saw an iOS project that had huge chunks of uncompiled Symbian code in the project "for reference", an entire pantheon of God classes, entire files duplicated rather than changing access modifiers, 1000 lines inside an always true if block, and 20% of the 120,000 lines were:

//

And no, those were not generally followed by a real comment.

replies(1): >>41903546 #
7. michaelt ◴[21 Oct 24 10:00 UTC] No.41902461{4}[source]▶
Actually some of us have been in the industry for more than 22 months.
8. tonypace ◴[21 Oct 24 12:41 UTC] No.41903546{5}[source]▶
And yet, I have an unfortunately clear mental picture of the human that did this. In itself, that is a very specific coding style. I don't imagine an LLM would do that. Chat would instead take a couple of the methods from the Symbian codebase and use them where they didn't exist. The God classes would merely be mined for more non-existent functions. The true if block would become a function. And the # lines would have comments on them. Useless comments, but there would be text following every last one of them. Totally different styles.
replies(1): >>41903997 #
9. ben_w ◴[21 Oct 24 13:33 UTC] No.41903997{6}[source]▶
Depends on the LLM.

I've seen exactly what you describe and worse *, and I've also seen them keep to one style until I got bored of prompting for new features to add to the project.

* one standard test I have is "make a tetris game as a single page web app", and one model started wrong and then suddenly flipped from Tetris in html/js to ML in python.

10. max51 ◴[21 Oct 24 13:58 UTC] No.41904228{4}[source]▶
I saw a lot of unbelievably bad code when I was teaching in university. I doubt that my undergrad students who couldn't code had access to LLMs in 2011.