Comment by KineticLensman | Hacker News Reader

Comment by KineticLensman | Hacker News Reader

Parent Full thread

(source: am a literature student - UK open university)

I just asked get ChatGPT a question about use of language in a specific book (China Miéville's Kraken). It gave a plausible answer but used a common English word ('knackering') instead of the word 'knacking' that is actually used (the word 'knackering' isn't in the text). I then asked it to give me a quote using 'knackering' and it gave me a quote with the word 'knacking'. I asked it to put the quote into Harvard reference style which it did, but didn't provide a page number, which it said I'd have to look up myself.

I think relying on ChatGPT for an essay would be fairly frustrating, and if you used ChatGPT to develop your thesis, you could very quickly get a generic plagiaristic answer (based on the quotes everyone else used) rather than one than that captured your own response to the text, and which contained factual errors.

[Edit] It also said that "gods inhabiting USB sticks are presented matter-of-factly" in Kraken which is false (there are no mentions of USB). When I asked it for a quote to support its assertion, it replied "While the novel doesn't explicitly depict gods inhabiting USB sticks..."

Mieville is a wonderfully challenging prompt for ChatGPT. In the same vein, I would imagine 'The City and The City' would end up with a bunch of half-hallucinated forum-posting and reviews regarding Disco Elysium.

> Mieville is a wonderfully challenging prompt for ChatGPT.

Indeed. I'm doing a close reading of Kraken and have been amazed at the number of 'weird' entities and plot events that Miéville introduces [0]. It would be very easy to hallucinate a few more.

There are more than 100 passages in Kraken where Miéville names people, places and events that occur in other media or culture. For example, in a list of people associated with submarine technology, he throws in the name of the underwater stuntman [1] who played the Creature from the Black Lagoon. Again, it would be easy to hallucinate additional content.

[0] For example, a secondary character who was originally an ancient Egyptian burial slave-statue who went on strike in the afterlife, and in modern London is now the trade-union convenor for the Union of Magicked Assistants.

[1] Ricou Browning, if anyone is interested.

Did you upload the text of the novel? It’s under copyright so wouldn’t be in the training data or retrievable via search.

I didn't upload the text; I was trying to see what happened if one asked a question about a specific book, as if I was someone trying to do an essay in a hurry. I know the text quite well so could easily spot the hallucinated answers but someone with less knowledge might have been fooled. The experience reconvinced me not to rely on this specific LLM for help with my essays.

Well, you have to give the tool a chance. If it can’t even see the text you’re asking about, it’s going to rely on whatever quotes and references it can crib together from reviews and blogs and comments.

Try it for a text that is in the training data, or the public internet, or can be put into the context window - then it might help.

Well, yes, I'm sure if I uploaded a full text and spent time on the prompts I could get better results, although these might be similar to other students doing the same thing, which might get me a plagiarism fail.

My point was that I would be very hesitant to rely on ChatGPT as an assistant in a real literature task. Many of the texts on my Eng Lit course are in copyright, as is all of the module material (the OU's course-specific textbooks). The hallucinations are a real show-stopper.

Your position seems to be “if you use the tool like a total dumbass, you might get a bad outcome.” Like… maybe use it intelligently?

On the topic of hammers: "if you don't hold the hammer at exactly 2.6cm from the end of the handle and strike the nail with 6N of force at an angle of 58 degrees then of course you won't get a good nail strike into the wood. Oh and you must only use acacia sourced from the subtropics".

Give me a break, it's a hammer. This is a perfectly normal "use" of ChatGPT and a good example of how a literature student may opt to try and use AI to make their work easier in some way. It also conveniently demonstrates some of the shortcomings of using a LLM for this sort of task. No need to call them a dumbass.

Reply on news.ycombinator.com