A new report from plagiarism detector Copyleaks found that 60% of OpenAI’s GPT-3.5 outputs contained some form of plagiarism.
Why it matters: Content creators from authors and songwriters to The New York Times are arguing in court that generative AI trained on copyrighted material ends up spitting out exact copies.
The individual GPT-3.5 output with the highest similarity score was in computer science (100%), followed by physics (92%), and psychology (88%).
And that’s why this claim is mostly bullshit. These use cases are all sciences, where the correct solution is usually the same or highly similar no matter who writes it. Small snippets of computer code cannot be copyrighted anyway.
Not surprisingly, softer subjects like “English” and “Theatre” rank extremely low on this scale.
Not to mention that a response “containing” plagiarism is a pretty poorly defined criterion. The system being used here is proprietary so we don’t even know how it works.
I went and looked at how low theater and such were and it’s dramatic:
The lowest similarity scores appeared in theater (0.9%), humanities (2.8%) and English language (5.4%).
Pun intended?
Yeah, anyone who has written a thesis knows those tools are bullshit. My handwritten 140 page master’s thesis had a similarity index of 11%.
So, if the Ai gives you a correct answer to a science question, it’s “infringing copyright” and if it spits out a bullshit answer, it’s giving you wrong, and unsupported claims.
Right? Nod doubt that output can be similar to training data, and I would believe that some of it is plagiarism, but plagiarism detectors are infamous among uni students for being completely unreliable and flagging pronouns, dates and citations. Until someone can go “here’s an example of actual plagiarism” (which is obvious when pointed out), these claims make no sense.
If it’s plagiarizing, so are Google search results summaries.
It’s not like it doesn’t cite where it found the data.
Eh, kinda. It’s not like a science paper is just going to be an equation and nothing else. An author’s synthesis of the results is always going to have unique language. And that is even more true for a social science paper.
Are those “best matches” paper-sized, or snippet-sized?
Article mentioned 400-word chunks, so much less than paper-sized.
But also, there is far less training data to mix and match responses from, so naively I would expect a higher plagiarism rate, by its very nature.
Less than 2% of the world’s population has a doctorate. According to the US Census Bureau, only 1.2% of the US population has a PhD.
“Only” 1 in a hundred Americans are PhDs? Thats far higher than I would have expected.
Surely many who have them received them from elsewhere before immigration to America, and likewise the proportion of immigrants who have them I would expect to be oversized. Americans tend to be more greedy than anything else and don’t put in the effort required for such small (financial) rewards.
Also, those with PhDs tend to congregate into certain areas that support those jobs, i.e. cities but not even a goodly number of those so much; plus smaller college towns too ofc. As such, many in the general populace might rarely if ever run into one for the largest majority of their lives, unless traveling specifically to those areas for some reason?
And ofc rural areas are far larger, geographically speaking, than places where a person with a PhD would (likely) go. So you could randomly pick a spot on a map 100 times and never manage to find someone with a PhD anywhere within tens of miles, I would expect - although that line of thinking reveals my own biases: do most educated farmers stop at like an MS and just follow up with their own (possibly even extensive) self studies, or go all the way to PhDs while working their actual farms? (I doubt it bc it does not sound practical, and that is a hallmark of farmers afaik, but I could be wrong…) Anyway, I expect the unequal distribution is a contributing / exasperating factor to the general rarity.
The most recent data is 2.1% for people over 25. https://www.census.gov/data/tables/2022/demo/educational-attainment/cps-detailed-tables.html
Ironically, in the article, the link to the original Census source of the 1.2% datum is now dead.
Also, it’s 2.1% now (for people over 25), according to the Wikipedia article’s source: https://www.census.gov/data/tables/2018/demo/education-attainment/cps-detailed-tables.html
Edit: the Wikipedia citation is from 2018 data. The 2023 tables are here: https://www.census.gov/data/tables/2022/demo/educational-attainment/cps-detailed-tables.html
Citation party!
I think the issue is more about HOW they wrote it, rather than WHO wrote it.
You can’t write a paper covering scientific topics without plagiarism. A human would be required to. Generative AI should be held to at least as high of a standard.
Turns out ChatGPT isn’t writing a scientific paper though, it’s conversing with the user.
If it’s regurgitating other people’s work then it needs citations.
This looks like an ad. They go on about what their proprietary detection method found without any details about how it came to these conclusions or even how they generated the test data. They give 0 actual examples for any of their claims.
Here’s the original blog post the article is referencing: https://copyleaks.com/blog/copyleaks-ai-plagiarism-analysis-report
Yep, Axios straight-up printed an ad as news.
They should show a small, but representative sample of questions they gave it.
Also they should compare the scores to similarity scores for a flesh and blood smart human that answers the questions.
Well, I tried it. So here’s an example.
this may soon be a thing of the past as
This fragment was flagged as plagiarism.
Don’t worry, It’s only piracy if a poor person does it.
“Plagiarism detection company claims LLM conditions plagiarism according to their detector.”
I wonder how many student written essays also contain ‘plagiarism’ according to their tool.
Probably very few. The bias for these companies is in false negatives, not false positives, since false positives create controversy when students appeal a ruling.
The bias here was certainly to come up with a lot of false positives for advertising; kinda like anti-virus companies do it.
100% iirc, there are only so many ways to write about how the blue curtains indicate the character is feeling depressed or something.
Ai outs ai. Also, haven’t these ai anti-plagiarism tools shown to have very high false positive rates?
Isn’t that basically what the current LLM AI fundamentally does? Just digests a bunch of text and gives a summary back in response to queries?
No. It’s not really clear what LLMs do, but it certainly depends on context.
What they fundamentally do is continue a text. That’s what they were originally trained to do. Then they were fine-tuned to continue a chat log or respond to an instruction. To be able to do that, they have learned a lot. Unfortunately, we do not know what.
If you ask for a summary of some text, it will give you one; regardless of whether the text even exists.
The summary could be one written by a human that it has memorized. Or it could be complete nonsense, that it is making up on the fly. You never know.
A genuine question: How well do chatgpt & others add citations if asked?
ChatGPT itself doesn’t know where it got the info from, so it makes up links and names - it’s a language model, not a search engine.
On the other hand, if you manage to find a reputable source and give it relevant metadata, it can format a nice citation for you, saving you time on that instead.
Badly. This burns my laziest students every semester. Chatgpt just adds nonsense citations.
Microsoft’s copilot adds them, it’s why I prefer to use it.
Copilot is GPT under the hood, it just starts with a search step that finds (hopefully) relevant content and then passes that to GPT for summarization.
I know, but gpt doesn’t do it, so I won’t use it.
There is custom gpts for that. ScholarAI and Consensus are OK.
It depends on how they’re using it behind the scenes. Chatbots like ChatGPT can’t cite sources, because they are just generating text on the fly. However, some approaches (if links/sources are provided) use an approach called Rag (Retrevial Augmented Generation). This approach uses similarity in search terms to find sources first, then uses the sources to augment/generate its answer.
That being said there are pros and cons to both approaches.
Perplexity AI includes citations every time.
One AI company throwing accusations at another AI company, and the evidence on both sides is to point their fingers at their own black-box LLMs like they’re magic…
10% are bullshit legal case citing made up for lawyers who are lazy
Only 60%?
59.7%
ok, so? plagiarism is a meaningless, tenuous call that can be avoided simply with some quotation marks and a link. isn’t this supposed to unite humanity’s knowledge rather than nitter and ditter around about meaningless technicalities? it’s not writing a fucking paper (and if you copy and paste it for a paper that’s already plagiarism anyway)
if it had to remember a citation for everything it knew, it’d only be able to remember half as much information because its memory would be cluttered with useless citations that you could easily find by googling if you really cared to know. most people just want quick facts
Is it using quotation marks and a link?
