GPTZero Case Study – Exploring False Positives

https://gonzoknows.com/posts/GPTZero-Case-Study/

by sungage • 3 years ago

181 113 3 years ago

Maken2• 3 years ago

Alternative title: Most academic papers are indistinguishable from AI generated babble.

sigmoid10• 3 years ago

...to AI. It's kinda funny how this is yet another area where these models suck very much in the same way that most humans do. LLMs are bad at arithmetic? So are most people. Can't tell science from babble? I already wouldn't ask a non-expert to rate any aspect of an academic paper. Trusting the average Joe who has only completed some basic form of education would be tremendously stupid. Same with these models. Maybe we can get more out of it in specific areas with fine tuning, but we're very far away from a universal expert system.

brookst6• 3 years ago

The best was the Ted Chiang article making numerous category errors and forest/trees mistakes in arguing that LLMs just store lossy copies of their training data. It was well-written, plausible, and so very incorrect.

jejeyyy77• 3 years ago

I mean, humans can't distinguish AI written text either - which is why this tool was built?

I don't see how it will be possible to build such a tool either as the combination of words that can come after another is finite.

Maken2• 3 years ago

I do agree that the most likely reason is that scientific papers tend to be highly formulaic and follow strict structures, so a LLM is be able to generate something much more alike to human writing than if it tries to generate narrative.

But it's still fun to deduce that the reason is that the quality of technical writing has sunk so low, that is even below the standards for AI generated text.

jldugger• 3 years ago

...to AI.

Perhaps we can call it the "Synthromorphic principle," the bias of AI agents to project AI traits onto conversants that are not in fact AI.

nobu-mori• 3 years ago

That's great! All we need to do is negate the output and it will be more accurate.

explaininjs• 3 years ago

Non-editorialized title: GPTZero Case Study (Exploring False Positives)

@dang

icapybara• 3 years ago

Right, I mean, how could it know?

I could write in the tone of ChatGPT if I tried hard enough. It's an intractable problem and a tool like this probably does more harm than good.

JoeAltmaier• 3 years ago

If it's like many of these 'AI' engines, it's a statistical map of text. I'd expect it to produce no better than what it's trained on, then muddy that by randomly combining different versions of similar statements.

I'm impressed it is ever accurate.

brookst6• 3 years ago

I’ve got a fun little side project that uses GPT. I tested gptzero against 10 of my projects’ writings and 10 of my own. It detected 6 out of 10 correctly in both cases (4 gpt-written bits were declared human, 4 human-written were declared gpt).

Which is better than 50% but not nearly good enough to base any kind of decision on.

d0mine• 3 years ago

which is better than 50%

Unrelated: p-value for getting 12 from 20 correct just by chance is ~0.4 that is there is not enough data for the conclusion "better" in this case.

Null hypothesis: 50%/50%, the result random, normal distribution:

  H0: p=1/2
  H1: p!=1/2 (two-tail) 


  import statistics
  
  p0 = 0.5  # proportion of successes according to null hypothesis
  n = 20   # sample size
  p_sample = 12/n  # 12 from 20 are correct
  
  sigma = (p0 * (1 - p0) / n)**.5  # std according to H0
  z_score = (p_sample - p0) / sigma  # test statistic
  p_value = 2*statistics.NormalDist().cdf(-abs(z_score))  # prob. two-tails
  # p-value -> 0.4

panarky• 3 years ago

As millions of people interact with ChatGPT, their writing will subtly, gradually, begin to mimic its style. As future versions of the model are trained on this new text, both human and AI styles will converge until any difference between the two are infinitesimal.

onos• 3 years ago

Interesting idea, but isn’t there variance in the output? Eg I’ve seen people ask it to “write in the style of x” etc and different people also clearly have different writing styles.

Filligree• 3 years ago

You can try, but it isn’t very good at that. The style remains very ChatGPT.

moralestapia2• 3 years ago

We train AIs but they also train us.

ominous• 3 years ago

Some related idea, in case you like to see that thought explore: https://medium.com/@freddavis/we-shape-our-tools-and-thereaf...

brookst6• 3 years ago

One of the big complaints with LLMs is the confident hallucination of incorrect facts, like software APIs that don’t exist.

But the way I see it, if ChatGPT thinks the Python list object should have a .is_sorted() property, that’s a pretty good indication that maybe it should.

I work in PM (giant company, not Python), and one of these days my self-control will fail me and I will open a bug for “product does not support full API as specified by ChatGPT”.

xiphias2• 3 years ago

Or rather you can just catch method missing in the runtime and patch it with a chatgpt call

matthewdgreen• 3 years ago

ChatGPT also thinks certain coding algorithms should violate well-known information-theoretic bounds ;) I’ll put a ticket in with Claude Shannon.

howlin• 3 years ago

LLMs is the confident hallucination of incorrect facts

This is a very common feature of delirium in people. Chatting with an LLM seems a lot like what it would be to talk to a clever person with encyclopedic knowledge, who is just waking up from anaesthesia or is sleep talking.

dr_dshiv2• 3 years ago

it, if ChatGPT thinks the Python list object should have a .is_sorted() property, that’s a pretty good indication that maybe it should.

Yes! And when it hallucinates references for articles, often times those articles probably should exist…

visarga4• 3 years ago

if ChatGPT thinks the Python list object should have a .is_sorted() property, that’s a pretty good indication that maybe it should.

Hahaha, Python language fixing itself!!!

wankle2• 3 years ago

AI will turn half the people into Eloi (https://en.wikipedia.org/wiki/Eloi) while the rest of us will become the Morlocks.

andyjohnson02• 3 years ago

Prediction #1: Once enough ChatGPT output gets posted online, it will inevitably find its way into the training corpus. When that happens, ChatGPT becomes stateful and develops episodic memory.

Prediction #2: As more people discuss ChatGPT online, by late 2023 discussion of Roko's Basilisk exceeds discussion of ChatGPT. (half /s)

ClumsyPilot2• 3 years ago

Or. ChatGPT will overtrain on it's own data and go to shit the way google search did

LelouBil• 3 years ago

Did you see the new Bing chat ?

#1 is already happening !

See here (other HN thread) : https://twitter.com/tobyordoxford/status/1627414519784910849

fullshark• 3 years ago

Sounds accurate and horrifying, I don't get the enthusiasm for this at all beyond a desire to be there first and make a ton of money. All manuscripts get a run through an AI editor, all business writing is even more soullessly devoid of purpose beyond accomplishing task X, all blogposts are finetuned for maximum engagement and therefore ad/referral revenue.

That's already happening I know but it will be amplified to the point that all humanity in writing in lost. All ideas in writing will be a copy of a copy of a copy and merely resemble something once meaningful. Time to go touch grass.

visarga4• 3 years ago

Too much AI for you? You can fix your problem with even more AI! Get your own AI, running on your hardware, loyal only to you. It will act like a firewall between you, the vulnerable, hackable human, and the wild internet full of other AIs. To go out on the internet without an AI is like going for a walk during COVID without a mask, or browsing without AdBlock. Their AI will talk to your AI, and that's how you can be safe.

zdragnar• 3 years ago

So long as ChatGPT is forbidden from communicating in certain ways (swearing, speaking ill or positive of controversial people or topics, etc), convergence will never happen. People interact with other people more than they do ChatGPT, so the majority force will remain dominant.

joe_the_user• 3 years ago

An opposite possibility is that the commonness of ChatGPT will cause people to adopt a style as distinct from it as possible.

Of course, this might mean future Chatbots would successfully emulate that. But it's not impossible an "adversarial style" exists - this wouldn't be impossible to emulate but it might be more likely to cause the emulator to say things the reader can immediately tell are false.

One idea is to "flirt" with all things that people have come up with that AI chokes on. "Back when the golden gate bridge was carried across Egypt..."

wand3r• 3 years ago

I saw this[1] interview with Sam Altman touching on interim AI impact. I really agree with his point that basically detecting output from LLMs is basically going to be futile and only really relevant in the near term. Accuracy is obviously going to improve in models and detection isnt that difficult now but will be in the future, especially if output is modified or an attempt to obfuscate origin is made.

[1]https://youtu.be/ebjkD1Om4uw

sebzim45002• 3 years ago

detection isnt that difficult now

I would have thought this, but every attempt I've seen at detecting chatGPT generated text has failed miserably.

Closi2• 3 years ago

It fails on false positives, but you don't often get false negatives, which might be enough at the moment for quite a few use-cases.

Also false positives are typically "this text is likely to contain parts that were AI generated" rather than "This text is higly likely to be AI generated" (which is what GPT-generated content generally produces).

When I've tried to prompt-engineer GPT to produce text that GPTZero will flag as negative it has been pretty tough!

RC_ITR3• 3 years ago

Accuracy is obviously going to improve in models

Well, to be clear, they can put rules based filters and other things on top of the neural net, but the core GPT will never get more accurate since it has no mechanism to understand what words mean.

valine2• 3 years ago

GPT3 is far more accurate than GPT2. Seems reasonable that larger models trained on more data will continue to improve accuracy. I'd also expect larger models to be better at summarizing text, ie potentially fixing the Bing issues where it hallucinates numbers.

Our models sizes are a product of our scaling and hardware limitations. There's no reason to believe we are anywhere near optimal.

martinflack• 3 years ago

It's probably a short-term social phenomenon. We don't bother detecting mathematical output from calculators or spreadsheets; we just like that folks give us the right answer, even if they had easy tooling to produce it. However, watching someone do things the old way would seem bemusing. If you watched a manager notating all over a physical spreadsheet with a pencil (as was commonly done at one time) it would seem quaint or backwards depending on context. Likewise, waiting for someone to write a letter and taking more than 90 seconds because they didn't co-author it with AI might seem slow.

mtlmtlmtlmtl3• 3 years ago

"Write an email that says I did X and they should do Y, but if Z then W, and we should schedule a meeting with P and Q."

I feel like for most emails I write, information density is close to a maximum. This means there's no actual gain to be had from a language model. The email I would write myself is going to be about the same length as the prompt I'd have to write anyway.

scotty79• 3 years ago

I hope gpt detectors will evolve into general bs detectors.

Nevermark• 3 years ago

I wonder how good the best language models are at detecting bs.

So many experiments to run...

stevenhuang• 3 years ago

It sounds feasible to detect cases of low temperature output, as the text output this way sounds very generic (like an average of all content seen near the prompt's latent space).

However once you prompt the LLM with a higher temperature, or tell it to roleplay as someone with elaborate personas, or suggest to use certain linguistic styles, or train it on example text... then it becomes much harder.

I imagine pathological cases of formulaic word use, sentence/paragraph structure will only be detectable in longer form text. After all text is already pretty low-resolution, not much for adversarial models to work with.

e2le• 3 years ago

Perhaps instead, the solution is to hold people to a higher standard. AI as a tool to help people in identifying logical fallacies (special pleading, anecdotal, the fallacy fallacy) written or otherwise could be very valuable. Maybe in the same way that ChatGPT can identify and explain the problems a piece of code has. It could even be integrated with moderation tools. This could in-effect make individuals less vulnerable to convincing nonsense said by professional bullshitters or parroting individuals.

wongarsu2• 3 years ago

This is testing on abstracts of medical papers. I wouldn't be surprised if the way abstracts are carefully written and reviewed by > 3 people is very different from normal text, and in some ways more similar to LLM output.

jmfldn2• 3 years ago

It's important to shine a light on the limitations of AI detection software, and this case study on GPTZero does just that. False positives can have serious consequences, particularly in sensitive areas such as healthcare.

MonkeyMalarky• 3 years ago

Thank you for your heroic effort in copying and pasting a chat log, I am left in awe by it.

...nice edit...

jmfldn2• 3 years ago

Lighten up mate.

bentcorner• 3 years ago

Random thought: In the future the simplest method to reduce your chance of GPT usage being detected is to pretend your paper was written in 2022 or earlier.

k__• 3 years ago

Yes, it's really bad.

False positives and negatives all over the place.

I really wish it worked, but generally, it doesn't.

lwhi3• 3 years ago

Perhaps AI generated text should be created with a specific signature in mind _specifically_ to be identifiable?

mattnewton• 3 years ago

isn’t this essentially asking anyone who runs a model to flip the evil bit[0]? People who want to misrepresent model output as human written output will trivially be able to beat this protection by removing the signature or using a version of the model that simply doesn’t add it.

[0] https://en.m.wikipedia.org/wiki/Evil_bit

lwhi3• 3 years ago

I think you're correct, but it would promote the idea that what's produced is a basis or source for further work .. and would mean that effort is required.

Feels like a good basis tech for something like ChatGPT.

wongarsu2• 3 years ago

There's a large body of research into invisible text watermarking, so this would certainly be possible. Maybe the simplest to implement in LLMs would be to bias the token generation slightly, for example by making tokens that include the letter i slightly more likely. In a long enough text you could then see the deviation from normal human text characteristics.

lwhi3• 3 years ago

Yep, I think it's a good idea.

PeterisP2• 3 years ago

All the main scenarios for detecting generating text that I can imagine do have to assume that the LLM isn't "cooperative" but actually is specifically designed (or fine-tuned) to avoid detection.

wankle2• 3 years ago

"Tell me more about your perspective customers."

https://projects.csail.mit.edu/films/aifilms/AIFilms.html

netfortius• 3 years ago

FoxNews is accurate 10% of the time, at best, and it's allegedly produced by humans, so I'm not really seeing the problem here...

ClumsyPilot2• 3 years ago

you do it all wrong, it's 90% accurate and 90% sucessfull - its goal to be as inaccurate as possible without its audience realising.

If fox news deletws the last 10% of reality from it's broadcasting, the people might start catching on. Although these days i am not sure

dsign2• 3 years ago

I had some fun yesterday when ChatGPT hallucinated a bibliographic reference to an article that didn't exist. But the journal existed, and it had plenty of articles that made ChatGPT's hallucination plausible. I think that at least this use case can be fixed with some pragmatic engineering[^1].

[^1]: Which may take a bit to happen, because our current crop of AI researchers have all taken "The bitter lesson"[^2] to heart.

[^2]: http://www.incompleteideas.net/IncIdeas/BitterLesson.html

FPGAhacker• 3 years ago

Did chatgpt post the reference as a footnote (or parenthetical)?

At least for now,I was thinking it didn’t do that, and maybe the lack of references would be an indicator of unedited gpt output.

dsign2• 3 years ago

It did:

...

However, according to a study published in the Journal of Dairy Science, the diacetyl content of butter can range from approximately 0.5 to 14 parts per million (ppm) (https://doi.org/10.3168/jds.S0022-0302(03)73775-3).

The doi I could not find, so I'm pretty sure is bogus.

I asked it to produce a full reference:

Sure, the full reference for the study I mentioned is:
Yvon, M., Chambellon, E., & Bolotin, A. (2003). Effect of pH on diacetyl and acetoin production by Lactococcus lactis subsp. lactis biovar diacetylactis. Journal of dairy science, 86(12), 4068-4076.

I went to the index of the journal 86(12), and that article is not there.

photochemsyn• 3 years ago

Just wrote this myself, although I did try to chatGPT-style it a bit. I thought the final third would serve to identify it as non-AI as it goes off on a tangent about isotopes...

"The periodic table is a systematic ordering of elements by certain charcteristics including: the number of protons they contain, the number of electrons they usually have in their outer shells, and the nature of their partially-filled outermost orbitals."
"Historically, there have been several different organizational approaches to classifying and grouping the elements, but the modern version originates with Dmitri Mendeleeve, a Russian chemist working in the mid-19th century."
"However, the periodic table is also somewhat incomplete as it does not immediately reveal the distribution of isotopic variants of the individual elements, although that may be more of an issue for physicists than it is for chemists."

GPTZero says: "Your text is likely to be written entirely by AI"

Now I'm feeling existential dread... perhaps I am an AI running in a simulation and I just don't know it?

thomastjeffery• 3 years ago

Technical writing, in order to be relatively unambiguous - the "technical" part - defines itself as a subset of English with a constrained grammar and vocabulary.

You just illustrated technical writing. Naturally, your writing style is very similar to that of other technical writing.

Take one guess what kind of writing exists in most of the text GPT is trained on.

catchnear4321• 3 years ago

How has your “success” been with chatgpt? Qualitatively, generally, positive/negative.

Being able to speak to machines would likely correlate with “sounding like one.”

There could be a different (mis)categorization here but also por que no los dos.

LoveMortuus• 3 years ago

I mean, technically speaking you are a man made intelligence, thus it would be fair to say that you are artificial intelligence. ^^

sdwr• 3 years ago

You made a few unforced errors that move it away from the quietly authoritative, mirror-sheen AI voice.

The colon in line 1 is clunky, the combination of "but" and "with" in line 2 reads as passive, and line 3 is full person.

robocat3• 3 years ago

Perhaps you just write very predictably.

A partisan writing cliched slogans and regurgitating tired political statements has a “temperature” closer to 0, likewise an engineer stringing together typical word combinations. A poetical writer with surprising twists and counterintuitive mixtures of words has a temperature closer to 1.0.

I can see a few clichés in your writing, also I did a search on some fragments of sentences which showed a number of results.

If you want to be less “robotic” then you could add whimsical or poetic wording, and less common turns of words (be a phrase rotator).

mt_• 3 years ago

Isn't the GPTZero based on model detector by OpenAI for GPT-2? In the initial preview it was exaclty like the demo, a text box styled diferently.

[1] https://github.com/openai/gpt-2-output-dataset/blob/master/d...

gtsnexp• 3 years ago

Detecting text generated by large language models like ChatGPT is a challenging task. One of the main difficulties is that the generated text can be highly variable and can cover a wide range of topics and styles. These models have learned to mimic human writing patterns and can produce text that is grammatically correct, semantically coherent, and even persuasive, making it difficult for humans to distinguish between the text generated by machines and the ones written by humans.

Another challenge is that large language models are highly complex and constantly evolving. GPT-3, for example, was trained on a massive dataset of text and can generate text in over 40 languages. With this level of complexity, it can be challenging to develop detection systems that can keep up with the ever-changing text generated by these models.

To implement a reliable detection system like GPTZero, which is designed to detect text generated by GPT-3, several challenges need to be addressed. First, the system needs to be highly accurate and efficient in identifying text generated by GPT-3. This requires a deep understanding of the underlying language model and the ability to analyze the text at a granular level.

Second, the system needs to be scalable to handle the vast amounts of data generated by GPT-3. The detection system should be able to analyze a large volume of text in real-time to identify any instances of generated text.

Finally, the system needs to be adaptable to the evolving nature of large language models. As these models continue to improve and evolve, the detection system needs to keep up and adapt to the changing landscape.

ericmcer• 3 years ago

It was weird how easy this was to identify if you have read any amount of ChatGPT content. It has a particular writing style that is pretty obvious.

I am not sure how you would code something to detect an author based on writing style. It feels like something people would have tried to do before. Probably using a similar approach that LLMs use but with a separate predictor for specific authors.

mds• 3 years ago

ChatGPT in particular writes in middle school essay format: introduction, point 1, point 2, point n, conclusion.

throwaway675309• 3 years ago

It's because it reads exactly like every middle schooler writing their first MLA formatted five paragraph essay.

sdflhasjd• 3 years ago

Ironic, because to me this was so obviously ChatGPT to me from literally the first sentence.

I think this is probably because it doesn't match the conversational style of a forum discussion.

janalsncm• 3 years ago

GPTZero cannot work because they don’t have all of the logits used to judge perplexity. They only have the top five or so. Even OpenAI which has the full set of logits is not able to correctly classify all texts. It’s a futile effort.

jaimex2• 3 years ago

Who didnt see this coming when GPTZero was announced by clueless media?

I didn't even have to look at the internals to know GPTZero and OpenAI's solutions would be pointless. I'm still very surprised OpenAI made snake oil.

air7• 3 years ago

I read an interesting paper about an idea of watermarking LLM output text in such a way that makes detection very accurate for a long enough text. This is done by subtly changing the probabilities of the next word to be generated based on the last word that was outputed. Circumventing it by manually changing words post hoc would potentially require almost as much work as writing it from scratch.

The idea seems quite roboust to me and I can envisage a future where companies that provide access to LLMs would also publish a detection tool for their models.

tluyben2• 3 years ago

It's not hard to make a model that rewrites the text without changing the meaning which fails this. Our model[0] which is based on feeding chatgpt random things from the interwebs from before 2020 and letting it wobble on about it is pretty good and nice to play with, but it's pretty easy to change the score radically with just changing a few words. This is whack-a-mole no matter how it's done. For now you can be sure people are too lazy to do it, but there will be many tools in the future to evade tools like this.

[0] https://filteroutai.com/validate/a07e081b71b294ba2de236441be... https://filteroutai.com/validate/2c3fa6de32845df02be7a4ff185...

GPTZero Case Study – Exploring False Positives

Related Stories