Yes, but can it prove it? The old fashioned plagiarism problem was solved by finding the original it was copied from. With AI, the falsely attributed work IS original. There in lies…
AI-generated content can still be traced back to its source. Even though the content may be original, AI-generated works can be identified by their algorithmic patterns, which can be used to determine the source of the work.
ChatGPT wrote that, not me. How do we know we aren't simply reading ChatGPT arguing with itself?
> It is certainly possible for AI models to generate text that appears to be self-referential or self-reflexive. However, it is also possible to trace the source of AI-generated text by examining the specific patterns and characteristics of the text. For example, you can analyze the language and style used in the text, as well as the context in which it was generated, to determine whether it was likely produced by a specific AI model or not. In addition, you can also examine the metadata associated with the text, such as the date and time it was created, to help identify its source.
ChatGPT wrote that. But I'll say that this is a lot easier said than done. This will always be an increasingly difficult cat and mouse game.
Without knowing the false positive and false negative rate this article is meaningless IMHO. Furthermore, even if it had a 0% false negative and false positive rate, it then would be used to train the next iteration of language models.
A good idea superficially, but ultimately solves the wrong problem.
Literally everything. They’re using it to help with math homework. Putting together copy for a slide presentation about the history of Pepsico (why? I have no idea). Writing a silly story about an ape man who dominates the world through vicious conquests. “Now write a sequel” apparently works really well.
In the end, (ideally) it's the students that are losing out on education, though. That's unless there's a competition dynamic among students. Maybe embracing it is removing competition among students.
EDIT: On second thought, one can just put more emphasis on exams taken in-premises with supervision.
No, even ideally it's still wrong. Educators are the arbiters of competence. It's their subjective opinions of you (marked with grades) that literally decide if you eat lobster thermidor or if you eat raman through your life.
The greeks had an ideal. It's been so corrupted by the modern system that even the ideal is dead.
> It's their subjective opinions of you (marked with grades) that literally decide if you eat lobster thermidor or if you eat raman through your life.
I don’t think so? Maybe in some rare edge cases. If you were born to a rich family even with midling grades you will remain rich. If you were born to a poor family even with excelent grades you likely won’t raise above middle class, and that if you are lucky. If you were born into a middle class family and you are lucky enough to avoid catastrophies you will likely stay okay, but not lobster termidor every day rich.
Honestly it is much more likely that if they give you great grades and put the idea into your head that you should be a poet or a philosopher or similar that they send you down the ramen path.
Look at Paris Hilton. Do you think she cares about subjects she doesn't know whilenshe gets chauffered around on her rolls Royce, eating meals made by a personal chef, in her Beverly hills home?
The answer is no. A life of material richness vs being poor and highly educated. Ask a postdoc this question and I bet the majority would take lifetime wealth to having to slave away in a lab.
> would you rather be stupid? think of all the things you would have wished to know before you die.
If someone takes the approach that school is required for learning, then they'll stop learning after finishing university (as most people do) and within a decade they'll be stupid compared to someone who believed learning is a personal thing and kept learning after their formal schooling had ended.
You have few really distorted, cynical perceptions and biased ideas about school.
First you're projecting your personal experience in your (I guess crap) school in a crap educational system to the norm. It's not.
Second, jobs having higher educational standards is a reflection of the booming of educated youngs, not the other way around.
Third, skilled uneducated professionals (e.g. plumbers, cooks, electricians, mechanics) make much more than graduates on average, it's a night and day difference, comes also debt free and you can start compounding much earlier. If money is your goal an education is actually not necessarily needed if not dangerous.
No, I am not projecting. I went to good schools and had positive experiences in school. It's still an arbiter of competence. Its still failed to hold up even a vestigial "ideal".
And no, the stats say that tradespeople are massively underpaid on avg compared to BS holders. Ofc, no one believes this here since HNs demographics are highly clustered in the bay area where trades people basically just tell you how hard you're going to bend over for them...
The US does not have a good education system and you're projecting US system to the norm.
I'm European, know our school systems in various countries and it has no resemblance to the ideas you have about education and we don't really have such gatekeeping at college at all.
Third, about trades, I've never lived in SF, but I lived in italy/poland/switzerland and even 6 months at OSU in columbus and I've seen that all the time: skilled tradesmen in all these places make MUCH more than average graduate. There's no comparison. I'm 100% sure the average decent carpenter or electrician makes more than a stem graduate in Italy or Switzerland.
This is just cynical and wrong. School is about ideological indoctrination and conditioning the political subject via propaganda during childhood development away from the prying eyes of disobedient parents. Read your Plato and Dewey.
my partner (a professor in a European university) and I had a chat with gpt3 for 5 minutes tops, even though we were having some beers while we were doing that, we noticed instantly the “rhythm” of the output.
Oh that is just great. We will have all kind of teachers and professors running around with half baked ideas of what ChatGPT does or does not sound like, deducting points and being hardass based on a hunch. What do you think your false positive rate is? How did you quantified that based on a five minute interaction (which only contained true positive samples, and only in response to your prompting)
Ok, so I'll have chatgpt write my 15 page research paper with citations. I'll then spend 1.5 hours rewriting and proofing it as opposed to the multiday researching, outlining, and writing I used to have to do.
And with all of that extra time, I'll study for your "exams".
Just replace chatgpt with Wikipedia, did this for 4 years of uni and another 2 at an ivy n school. No one was the wiser and now I have a cushy MBB job to show for it.
Teachers will adapt. Some are already giving students a question, allowing the students ample time to research, then writing papers with a pencil in class.
ChatGPT is still supremely usefull there. I can borrow whole phrases and structure from it if I want to.
I get my question, research it, provide chatgpt with an outline and ask it to fluf it out in a nice style and maybe copy it down once. Even if I only remember 80% of what I wrote then that is still a win.
"Homework robs children of their childhood". Hell yeah!
I too think that's exactly what's wrong with the education system nowadays - too much homework and not enough time for drugs, Call of Duty, $120 apparel from LuluLemon, browsing TikTok all day in bed, and sending nudes over SnapChat.
If "wasting kids' time so they can't do worse things" is meant to be the point of the education system, it's a stupid idea, as it just means when they're adults with more free time they'll catch up on what they missed, and as a bonus they'll carry a lifelong hatred of learning from being trained to associate it with mind-numbing, pointless busywork at school.
I took Spanish for 7 years and passed all the classes. Countless hours of homework involved every week.
As of my 30s, I can now mostly read warning labels and some of the Spanish manual if I happen to grab the wrong one out of the box. Conversations with a native speaker? No way.
If that wasn’t a complete waste of my time, I don’t know what is.
It seems like you wrote this in support of the argument that homework is a waste of time, but it doesn't seem like you're describing a problem with the concept of homework or how it applies to learning a subject. You're describing a situation when the subject itself was not useful to you in retrospect.
Presumably, from your point of view, all of the time invested into it was a waste of time, not just the homework part. That said, it doesn't follow from this that any homework is a waste of time. It doesn't even follow that Spanish homework is generally a waste of time.
One cannot generalize from your example at all because you might be an outlier. You'd have to show that learning Spanish provided no value to the majority of kids who studied it, or that class time is sufficient to learn Spanish so homework isn't really needed, which would be a very hard thing to argue for a human language.
For what it's worth, I took Bulgarian folk music - I'm not Bulgarian, nor do I live in Bulgaria, nor can I play any musical instruments. Was it a waste of time? Acquiring knowledge is never a waste of time for me. Typing comments on HN might be though.
i would've been that student, and now i teach creative writing to kids in my free time. we use pen and paper to write poems.
but if i were an english teacher assigning research-like papers, i'd definitely incorporate chatgpt responses into my assignments. make them write a draft with chatgpt3 then revise it themselves in class with pen and paper. then write a version in their own words without plagiarizing the chatgpt version, but they can only use google docs or some format that can track changes, so the teacher can see the process. or something like that — it would need some iterating and real-world feedback.
>but if i were an english teacher assigning research-like papers, i'd definitely incorporate chatgpt responses into my assignments.
now THAT is a creative thing to do. I like that. I suppose the calculator version would be to do something like: "here is a simple graph plot, create an equation that produces a similar plot"
Wish there were more teachers like you embracing this stuff as a tool rather than doing your best to design a curriculum to 'defeat' it.
Ah, there's the true business model. Students can use ChatGPT for free, but schools get charged to be able to detect it. Brilliant!
Similarly, OpenAI's tools can be used to create all sorts of content marketing and other webspam. They they'll create a monetizable search engine-killer that knows how to ignore all this fluff (which will probably constitute 99% of the internet within a few years).
Oh I know — that's why I assume the web will be choking on AI-generated content marketing in the foreseeable future. The killer app will be the search/information/answer engine that can suitably ignore the AI-generated tripe.
I wonder if we'll look back one day and say "remember back before 2020, when the internet was filled with junk, but at least it was human-written junk?".
I can’t see how this will be useful unless the false positive rate is 0%. Can you imagine how demoralizing it would be to legitimately write a term paper and have it flagged as being AI-written? How do you disprove it?
Some people are putting a lot of hope in AI text detectors to prevent ChatGPT usage in contexts like student essays, etc.
While these detectors are of course interesting from the hacker/academic points of view, I don't think they will really solve much in practice.
On the one hand, our (teachers') adversary is not ChatGPT. It's ChatGPT + a human, which will be aware of these techniques and can tweak the text to avoid detection.
On the other hand, even if we could detect generated texts with high accuracy, are we going to fail a student just because a black-box AI model says so? Doesn't seem very ethical to me. It's funny how some people worried about ChatGPT's ethical implications seem to be OK with this, just because it goes against their personal demon.
I don't think there is any realistic alternative to just assuming that these systems will be used wherever text is required and you can't physically watch the writer, and living with it.
I went through college using Wikipedia to not only make my outline but also provide citations. I'd then rewrite the articles in my own word.
This greatly reduced the amount of preparation needed prior to writing.
No shame, the majority of research papers done in unis are silly and regurgitation.
Really technology is just getting good at eliminating busy work which will be the death kneel of liberal arts.
Double rant: college is largely a waste of time for liberal arts classes.
It's lazy writing, and the people who take these detectors at face value are the very same people who are easily fooled by AI text generators.
I like ChatGPT and spend quite a few hours a week working with it; it's a lot more interesting than just a text generator. That said, people using it to do schoolwork for them are short-changing themselves, because they're opting out of developing core cognitive skills.
> On the other hand, even if we could detect generated texts with high accuracy, are we going to fail a student just because a black-box AI model says so? Doesn't seem very ethical to me.
We already have that, today.
With fantastic tools such as TurnItIn[0], that regularly flags my own work as plagiarizing my own work. And we see with those tools, already, there is a presumption of guilt. The professor will believe the tool over the student, often without the slightest glance to see why said tool is saying the student is guilty.
Well, that's bad practice, and I hope it's not too common. I teach at a non-elite university at a peripheral place and we use TurnItIn, but we would never fail anyone without a human check to see exactly why the work was flagged.
Just my opinion, but the kinds of writing assignments ChatGPT excels at seem to me the least useful ones for assessing student understanding.
Why are we making kids write entire persuasive essays and research papers when in the real world these are almost always the result of collaborative effort?
ChatGPT is similar to a calculator. We don't let 3rd graders use calculators on their arithmetic homework, even though there's no way to detect if they used one. The point of a 3rd grader adding fractions is not because we care about the result of 3/4 + 1/3, but because the 3rd grader should learn how to add fractions.
If you think of ChatGPT as a fancy calculator, you can just apply the same techniques we've been using for a while: tell students not to use them on homework, make homework only a small % of the grade, and rely on in-class, paper only exams for most the grade.
The best class I had in high school (Mr. Hecker’s English class), required writing an essay in class every week. All my college essays were easy by comparison.
To each their own. I didn’t really learn how to write until in grad school, when I had things I really wanted to write about. I can imagine how much joy would have been sucked out of writing if I just had to write some essay every week that I didn’t have much interest in.
I don’t enjoy arithmetic or typing much either, but they are skills you have to learn in school so you can do things you _do_ enjoy. I think “writing an essay” is about the same.
I went to a high school when and where typing was required as a pre-requisite to programming. The only C I got in my four years of high school.
I think we make kids hate school via "sucking it up" when we really don't have to. Learning to write in the right context is much more effective than learning in the wrong context. But our educational system is so standardize that we have to fit most kids into the same interest peg.
Hah, I took a "keyboarding" course before programming was taught in my school. I typed a stretch at 60wpm. The teacher dryly remarked that I had just taken the 5wpm test and I'd need to work my way up to 60. Good times, good times...
I was just surprised we were graded on speed. Like, the girls (who...had thinner fingers on average?) would all get As by reaching 80 WPM, I'm stuck with fat fingers at 20-30 WPM (mind you, just copying text from one screen to another)...which is plenty of speed for programming and even writing comments on hacker news.
These days, that would be considered insane! But in 1989...
Take a look at Itzhak Perlman's hands[1] and listen to him play Paganini's 24 Caprices... I am positive that you could type faster if you tried. But, I agree, one does not need to be a champion speed typist to write good code. Think more, type less.
Yes, thin dexitorious finger build isn’t necessary, but at 14 would have definitely been helpful. The job they were training us for (transcription) was mostly already on the way out anyways.
I’m a big proponent of doing difficult things regularly.
If we habitually do hard things, then we generally have no problem, stepping up to the plate, for when it’s showtime.
Same reason that musicians practice scales, all the time.
I write code, every day. I walk 5K, every day. I solve problems, every day (detect a pattern, here?). It makes showtime a lot less stressful, and folks seem happy with my work.
First thing I do, is a 5K brisk walk. Gets the exercise quota satisfied. I tried running, but kept injuring myself, so walking, it is. I can do that, seven days a week. Takes a bit less than an hour (I walk about 10min/Km).
I usually "triage" my day, on these walks. I mentally sort through the tasks ahead, and scope out my plans. Sometimes, it's straightforward, other times, I have to develop a plan.
When I get home, I hit the shower, and generally start coding, right away. I code throughout the day, taking breaks as I need/want. I have no issue, stopping work for a couple of hours, and watching some TV or even taking a short nap.
I'll code at any time. I don't have a "set" workday.
I have found that I am most productive in the morning, and most creative at night, but I sometimes find myself losing focus, towards the end of the day. When that happens, I generally wrap things up, as I do more damage than good. I'll often fix a bug in my head during my walk, that was confounding me the night before.
I believe in F^3 - Finite, Focused, and Finished.
- Finite: I know what "done" looks like. Tasks have a discrete beginning and end, though the schedule may be fuzzy.
- Focused: No distractions. I stay on beam, and devote full attention to the task at hand.
- Finished: I don't stop, until I have reached a point that I can wrap things. I usually punctuate this with a final tagged commit.
It's also important to scope things reasonably. Stretch, but not too much.
I had a teacher like that in middle school. It was terrifying but after her class writing was never a sweat until 5-paragraph essays were no longer acceptable.
A competitive admissions high school I know requires prospective students to write the admissions essay in a set amount of time when they come to tour and interview in order to avoid parents/tutors/etc. from writing these essays, and leveling the playing field…think it’s a fantastic system
Yes, we had this all the time in English class in high school. It would be a prompt about the book we were reading or some texts he would print out and give us then and there.
When I was in third grade, I had to show the intermediate steps I used to formulate an answer to prove that I knew how to do the task. This has two benefits: (1) it gets students in the habit of working through a problem methodically, and (2) it lets teachers assign partial credit when students get the methodology correct but make a simple error.
Similarly, this is why we students have been required to cite sources instead of simply rendering an opinion on a topic. Most of these AI bots are just fancy plagiarism machines.
For queries in my field (materials science), chatGPT actually seems pretty good about citing real sources, often one of the canonical sources for the topic
An acquaintance who works in the NLP space says this is done in post processing, where GPT probably hallucinates a reference but they run a web query or something to try to fix it up.
Or create escape rooms for kids. In each room is all the information you need to solve the puzzle and escape. The final room you get to within an allotted time period determines your grade.
Edit: by Escape Room, I mean like the fun kind you would go to with friends or colleagues. I'm not saying we should torture children... Google if you haven't heard of one.
The problem is, any idea like this stops being fun the moment there are meaningful consequences attached to it. About the only way I can imagine such Escape Room working is if it's either ungraded, or the lowest outcome is an A, and anything above just gets praise or some "karma points" that have no value beyond giving kids something they can tease or challenge each other about.
The problem with "no graded homework" (which has to include "no sneaking homework into school"), is that it means that it's impossible to grade anything substantial like a thesis project, so grades only cover trivialities, and have a huge bias for rewarding speed over depth and quality.
I think text generators will do two things: reduce the average quality of published text, and reduce the average text writing skills. It thus increases the value of writing skills. I count that as a win.
I don't think cheaters will use any text from ChatGPT verbatim. The majority of "cheating" is going to be similar to how students use Wikipedia as a reference.
This might be useful to filter spam from places that initially got bombarded, like stack overflow. I agree with other comments, it's hard to judge the value of this tool without false positive/negative stats. The output isn't completely deterministic though so I wonder how "certain" this tool can be.
You'd be surprised how many students turn in Wikipedia-articles run through software that replaces words with (often wrong) synonyms. I've tutored students who paid such find-and-replace services.
Zero effort for the majority of students that cheat, very easy to detect, rarely punished (a zero on an essay here or there, no expulsion - students are customers after all).
There will be paid ChatGPT services for cheating students, the ideal solution would be to rethink and rehaul how student performance is evaluated, but universities will instead pay something like Turnitin because throwing money at a problem is always easier than thinking.
I taught college for a brief time. Catching cheaters is hard work, and unpaid, and it's perceived that doing it too much affects your chances of getting your contract renewed. "Let the employers or the next course level sort them out" seems to be a more pragmatic strategy.
Seeig a fellow TA pursue a cheater who copied a major math assignment symbol for symbol from another student, mistakes and all, and fail to break through the bureaucracy and achieve actual punishment was quite an eye opening experience. If this is how brazen cheaters are handled in one of the top Canadian engineering schools, what hope is there?
That's not how its handled in American universities. TA refers the suspected evidence to their superior then they hear nothing about it because its no longer their business, its between the student and the administration at that point, and students certainly get expelled.
Nothing to do with that at all. This is a top notch school, almost all students were domestic at the time (not sure about now, this took place in 2009).
Regardless of whether students should be using ChatGPT, I strongly disagree that a ChatGPT-written essay is "plagiarism".
ChatGPT is not an author, it's a computer program. We don't expect bibliographies to include Microsoft Word, Grammerly, or iOS Keyboard Predictive Text in the list of references. If I use a piece of software to create an essay, the essay that's created belongs to me.
I guess it depends if you consider the act of writing to be something that can only be done by a human. If a program can produce a work that passes as "writing", then that program could be called the "author" of the work.
So a student that takes said output and hands it in is plagiarizing by claiming credit for something written by a different "author".
It'll probably take some time for language to evolve around non-human agents. There's no intrinsic reason to prefer separating (or combining) that terminology by humanness of agent, afaik. At least, for common usage. One will win and only time will tell.
Plagarism is passing off others' ideas or words off as your own without attribution. IMO there is an argument to be made whether conjuring up text from a model of others' words should qualify as "others' ideas or words", but if it isn't technically academic plagarism I reckon it is a distinction without a difference. The concern with academic plagarism isn't primarily IP, so I reckon it doesn't inherently matter whether "others" being borrowed from is a particular author.
Also, There's a commonly understood threshold for plagiarism where a single word won't qualify and a phrase probably won't qualify, but a whole paragraph does. That is why Grammarly isn't plagarism, not because it is software.
It's not plagiarism in the sense that you took stole else's work, it's plagiarism in the sense that you didn't produce it yourself and passed it off as your own.
So in an academic context, for grading your understanding, it becomes meaningless to grade something written by a machine instead of by you. In other contexts, it's totally fine.
In this context, the initial prompt is the one from the teacher to the student that describes an essay title or topic. Adding prompts like 'academic writing' or 'avoid excessive detail' don't constitute any real creative or intellectual contribution. Do you think buying a cheeseburger makes you a chef?
> Do you think buying a cheeseburger makes you a chef?
No, because someone else was the chef in that case. But if I bought a conveyer belt machine that made cheeseburgers automatically at the press of a button, and I used that machine to cook dinner—maybe!
If we lived in a world where machines like that were common, the meaning of "cooking" would change...
I don't think so. I used to cook for a living and still make meals regularly at home, I'm pretty good at it. But sometimes I just want cup noodles, and when I make them I don't consider it cooking.
Maybe a better example is popcorn. I can make it in a microwave, or I can make it in a pot on the stove from scratch. It takes about the same amount of time, and the quality is about the same. But doing it by hand involves many more small decisions, like the amount of oil, corn kernels, heat, cook time, knowing when to stop, what flavor ingredients to add and how much etc. Now I can do it on autopilot and never have a single burned or unpopped kernel, but it took maybe 100 less-than-perfect efforts to get to that.
False. A craft requires deliberate effort and practice otherwise to proudly claim the "act of creation" (I cooked X, I drew Y, I wrote Z) has no meaning.
You are making 'letter of the law's arguments for a very much 'spirit of the law' issue. If we just grant you are absolutely correct for argument's sake, it would plainly indicate the letter of the law needs to change.
Students need to learn how to produce the best writing they can. If their best writing comes from feeding a prompt into ChatGPT, they should use ChatGPT. Presumably this student will also use ChatGPT in their writing post-graduation.
Of course, I would expect such a student to get marked down for factual errors, logical inconsistencies, and a lack of citations!
Some students, however, may take a hybrid approach, in which ChatGPT writes certain passages which the student manually cleans up. IMO, this is completely acceptable, and probably a good skill to practice! Students should learn how to use technology to produce the best work they can!
There are absolutely exceptions, when you'd want a student to practice writing independently. But writing the whole thing off as "plagiarism" is going much too far!
ChatGPT will certainly affect teaching methodology. But there's a reason why we teach students arithmetic and basic algebra without calculators, long after calculators were already ubiquitous.
The same goes for composing an essay. It's not just about the ability to produce a cogent-looking essay, in the same way it isn't good enough that a kid can punch 17*13 into a calculator.
^ I actually think we spend too much time on arithmetic, and should let students use calculators starting earlier than we do.
But, yes, calculators shouldn't always be allowed, and neither should ChatGPT. But, well, "plagiarism" is never allowed, so IMO the distinction matters!
At the end of the day, its not a skill, its a crutch. There will be many times where you are asked to contribute substantially to something without having time to go and craft a prompt and return output and make sure chatgpt isn't just spitting out bullshit like its ought to do. Imagine asking someone in a meeting, "what do you think about x, can you share your perspective?" and they tell you to hold on while they pound something into their phone then read it off their screen.
I could write a prompt and send it to someone for cash to write me an essay. It's still not my essay. Chatgpt is the same here. Grammerly isn't doing research for your essay. Most teachers for those sorts of papers don't care about how well you write anyway, unless its english class, they care about the arguments you've presented and the evidence you use to support them.
What's the difference between this and say something like stable diffusion to generate a piece of artwork?
By your logic could I not claim that I drew that ensuing picture because I "wrote the prompt"?
Let's take it a step further, what happens if I take my prompt and hire an artist on fiverr? We go back-and-forth via email as I ask them to make various changes and then wind up with a final drawing.
Am I still allowed to make the claim that I drew that picture?
> By your logic could I not claim that I drew that ensuing picture because I "wrote the prompt"?
I feel like "drawing" is a very specific action involving a writing utensil. I don't think I'd consider an image made in Adobe Illustrator to be "drawn", unless you were using a Wacom or similar.
But, yes, I'd say you could claim you "made" or "designed" the ensuing picture.
> Let's take it a step further, what happens if I take my prompt and hire an artist on fiverr? We go back-and-forth via email as I ask them to make various changes and then wind up with a final drawing. Am I still allowed to make the claim that I drew that picture?
No, because the artist on fiverr drew the picture.
Let's tweak this analogy a bit. Let's say you had a photograph of two people holding hands on a crowded beach. You want to retouch the photograph to remove the crowd, so the beach appears empty.
If you're lucky, you may be able to use Photoshop's Content Aware Fill to remove the crowd in less than a minute. In this case, you could absolutely claim that you retouched the photograph, even though Photoshop's algorithm did all of the work. A computer program can't be the creator of anything—the person using the program can.
If you're unlucky, Photoshop's Content Aware Fill will produce bad results, and you'll have to painstakingly go through the photo by hand. If you're short on time, you could hire a retoucher on Fiverr. In this scenario, however, you wouldn't be able to rightly claim that you retouched the photograph, because another human did that.
---
In every other case I can think of, the person using a tool gets to take credit for what they've made with the tool. Use an electronic screwdriver to fix a car? The screwdriver didn't fix the car, you did! Use a calculator to add two large numbers? The calculator didn't find the sum, you did.
None of this should change just because we have better tools. Schools may place additional limitations on students for certain assignments—math classes may restrict the use of calculators, art classes may restrict the use of drawing—but these limitations (1) should not always be in place for every assignment and thus (2) should be stated explicitly up-front.
ChatGPT is like hiring someone to write your essay for you. It's not your words and unethical to claim they are. If you used chatgpt and you think its up there with Microsoft Word or spell checkers in terms of tooling, then go ahead and tell your teacher you've used chatgpt for the assignment and see how that goes for you.
> Separately, burstiness compares the variations of sentences. Humans tend to write with greater burstiness, for example, with some longer or complex sentences alongside shorter ones. AI sentences tend to be more uniform.
Although I thought it would be trivial to work around this heuristic by specifying "use some short sentences and some long sentences", that didn't really work for me. When I said "use only short sentences", the average sentence length was a bit shorter. But I was surprised that this type of meta instruction didn't work better (at least in this version).
I tried defeating GPT detectors recently by prompting it something like "write with high perplexity" (https://towardsdatascience.com/perplexity-in-language-models...) and although it worked, I also couldn't get it to stop literally using the word "perplexity" in the output.
The problem isn't chatgpt, it's curriculum. If students were motivated to learn to write (and the skill is genuinely useful, this shouldn't be a hard task) this wouldn't be a problem. But instead for whatever reason we've adopted a coercion based system of education where we assume students innately dislike learning. Personally I think it's a good thing that chatgpt is allowing students to ignore the coercive education system.
None of this will work if you run the generated text through a smaller model designed to rewrite it.
I'm betting on schools using expanded remote proctoring software to monitor and record students' homework process and flag potential irregularities for teachers to check. Whether it's positive or negative for students' education, it will be less disruptive to schools than eliminating homework or significantly altering the curriculum.
But you can fool it by first asking chatGPT if it can write in the style of some known author, letting it output an example, then asking it to write about a given topic in the same style.
Here's generic output to "write a one paragraph description of Genghis Khan". The Hugging Face output detector says it's 99.98% likely to be fake (from GPT):
> Genghis Khan was a Mongolian warrior and ruler who founded the Mongol Empire, one of the largest empires in history. He is credited with uniting the nomadic Mongol tribes and establishing a powerful and efficient military force that conquered much of Asia and Europe in the 13th century. Genghis Khan is known for his strict code of law and his innovative use of cavalry and military strategy. He is also remembered for his brutal treatment of conquered peoples, as he and his armies were responsible for the deaths of millions of people during their conquests.
But this response, in the style of Charles Bukowski, gets a 98.66% real (not GPT) rating by the checker:
> Genghis Khan was a man of fire and fury, a conqueror who swept across the land like a storm. His armies were vast and unstoppable, crushing all who opposed him beneath their hooves. He was a man of vision and determination, who united the scattered tribes of the steppes and forged them into a mighty nation. But he was also a man of blood and death, who left a trail of destruction in his wake. Villages were burned and pillaged, their inhabitants slaughtered or taken as slaves. The world trembled at the mention of his name, and even the greatest kings and emperors bowed before him in fear. Such was the power of Genghis Khan, the greatest warrior the world had ever known.
If I can figure it out in a couple of minutes, everyone else would figure it out too.
By the way, here's the response to "can you write in the style of Charles Bukowski?" (the first prompt in the session):
> Sure! Here's a little something I came up with:
"I woke up this morning with a hangover and a broken heart. The bottle was empty, just like my bank account. Love had come and gone, leaving nothing but pain and regret in its wake. I stumbled out of bed and into the shower, hoping the water would wash away the memories of her. But as I stood there, staring at the peeling paint and the mildew on the walls, I knew that the only thing that would cure me was time and more whisky. So I dressed myself in my finest rags and headed out into the city, searching for the next fix, the next distraction, the next moment of fleeting pleasure that would take me away from the pain. Because in the end, that's all we really have: the present moment and the hope that tomorrow will be better, or at least no worse than today."
Thematically it’s pretty much on spot but stylistically nah. It needs to relax a bit and be a lot more informal / speechy and throw in some expletives here and there I suppose.
I wonder what would happen if you fed the Robert Jordan novel Brandon Sanderson had to finish and asked GPT to rewrite them in Robert Jordan's style. Then review the diffs and use that to do an edit.
What strikes me is that even OpenAI has been working on a way to "watermark" GPT-generated text with an "unnoticeable secret signal" to identify its source.
Disclaimer: I'm working on an app to solve the impending wave of generative content, so I'm somewhat biased.
As I'm sure many of us did, I tested out the app here on GPT-3 output. Unmodified, it detected it was GPT-3. Great! However, I added about 10 additional words to the output provided by GPT-3, and it shot up my "human" score by like 60 points, and determined it was human-generated.
This is going to be the problem underscoring _any_ model that is trying to identify "AI" generated text. A human can modify it slightly, or subtract words, and it throws the entire thing off. There are other paths that we need to explore to this problem.
As a generative modeling researcher, thank you for the new adversarial training methods.
In all seriousness though, this is really a cat and mouse game. Working in CV I might be biased to think that images are going to be easier to detect than text, but we are going to get better. I'm not saying we shouldn't create detectors (we should), but I think we also need to be aware that this is a cat and mouse game and we need to have true social conversations about this. Though we still fail to have these conversations when it comes to computer security so maybe we're doomed (or maybe this will be the catalyst). It's also important to note that a lot of damage can be done even with images and text that are easy to detect as fake. I've seen plenty of fake Twitter and Linkedin accounts that use StyleGAN profile pictures (with all the telltale signs).
I'm curious what kind of model you're using? How interpretable is it?
If we go back to the fear of AIs supplanting humans, Accelerando style, and the desire to keep humans relevant. Then assume that we accept the premise that humans-augmented-by-AI work is an acceptable outcome for that future, in an attempt to keep humans relevant.
Then is it a problem if a human slightly modified an AI's work? Isn't that the desirable outcome? And if the work itself is so useless that it would be worthy of a zero grade, then perhaps another way of measuring usefulness should be used.
In a way, I feel like we can't impose old-world grading techniques to new-world content synthesis technologies.
I avoided any and all classes that required writing papers and essays. I was able to satisfy the university liberal arts requirement by taking foreign language and accounting.
I have no regrets about that.
I wanted to learn math, science, and engineering. I didn't want a degree, I wanted to learn the stuff. What was the point of cheating? I've derived a lifetime of pleasure from knowing that stuff.
I don't understand wanting a liberal arts degree and then cheating to get it.
Then you're facing a career of faking it because you didn't learn the job skills. Also looking forward to complaining a lot about job screening tests, and having your job outsourced.
Employers know about the cheaters, fakers, frauds, and charlatans. The job screening tests are an attempt to screen them out. Nobody wants to hire cheaters.
Right — and yet, there’s a positive expected value.
I get you don’t like that, but that’s the reality of US labor market, where useless credentialism has crept in:
Many people get degrees they don’t want for jobs that don’t need them to check a box and get hired - because as a society, we’ve built that into our hiring.
I had a college student complain to me recently that many of his fellow students bragged about cheating. I suggested that was a good thing - those were the students you wouldn't want to set up shop with after graduation. 'Cause sooner or later they'll cheat you. It was nice of them to warn you that they're cheaters.
Oddly enough, I was the opposite case. By my reckoning, I was the first student at my small liberal arts college to submit a word processed paper. It was before most of the professors had started using word processing themselves. The ability to bypass a lot of the mechanical effort of writing a paper made the process easy for me, and doing it a lot probably made me a better (or at least quicker) writer. For instance it also improved my handwritten "blue book" exams.
Writing was easier than studying for me.
For this reason, outside of my math & science majors, I chose exclusively courses where the grade was based on written work.
But likewise, I wasn't in college to get an empty credential. I already had a marketable skill -- programming. Learning was fun, but I was also conscious of the fact that I didn't want to waste my parents' money.
The writing has affected my career: My work involves not necessarily writing papers, but certainly communicating ideas and knowledge to other people, including non-specialists.
Did you not want to learn to write, did you hold the writing courses had nothing to teach you, or are were you holding that courses requiring writing had bad time ROI (return on time investment)?
As co-founder and CTO of any number of tech companies, I've observed engineers (programmers) with broad humanities (history and language related, mostly) education tend to grok business narratives then architect and write better software: coherent systems expressed concisely. Now I look for history major with comp sci minor, or English and linear algebra, or other self-selected humanities + STEM pairings.
I'd hazard there's some relationship in how the mind is applied in common between domain driven design in service of business narrative and the sweeping cause and effect of history, or between software architecture and well organized long-form writing.
Consider that the people I know who learned math, science and engineering without going to university hovers around zero. I've also known degree'd engineers who were so hopelessly inept I suspected that they cheated their way through. Their careers didn't go well.
I once knew an MIT graduate who thought a coordinate system conversion matrix was black magic (I wrote it to convert screen graphics to the printer's idea of a page). I was floored. Howinell did he get through MIT?
> The college senior isn't alone in the race to rein in AI plagiarism and forgery. OpenAI, the developer of ChatGPT, has signaled a commitment to preventing AI plagiarism and other nefarious applications. Last month, Scott Aaronson, a researcher currently focusing on AI safety at OpenAI, revealed that the company has been working on a way to "watermark" GPT-generated text with an "unnoticeable secret signal" to identify its source.
Ah, so magnanimous of them to sell both the plague and the cure.
I say that tongue-in-cheek of course, but it seems that's where this is going, whether they make and sell both or let others do it. That is, both sides of the production and detection of this spam are about to profit handsomely from the overwhelming piles of it we're about to be drowning in.
160 comments
[ 5.2 ms ] story [ 233 ms ] thread"Your GPTZero score corresponds to the likelihood of the text being AI generated: 338.59305536714294 Your text is likely human generated!"
ChatGPT wrote that, not me. How do we know we aren't simply reading ChatGPT arguing with itself?
ChatGPT wrote that. But I'll say that this is a lot easier said than done. This will always be an increasingly difficult cat and mouse game.
A good idea superficially, but ultimately solves the wrong problem.
EDIT: On second thought, one can just put more emphasis on exams taken in-premises with supervision.
It's about maximizing your college admissions chances, which then leads to better jobs.
Id rather be stupid and rich than "educated" and poor.
The greeks had an ideal. It's been so corrupted by the modern system that even the ideal is dead.
I don’t think so? Maybe in some rare edge cases. If you were born to a rich family even with midling grades you will remain rich. If you were born to a poor family even with excelent grades you likely won’t raise above middle class, and that if you are lucky. If you were born into a middle class family and you are lucky enough to avoid catastrophies you will likely stay okay, but not lobster termidor every day rich.
Honestly it is much more likely that if they give you great grades and put the idea into your head that you should be a poet or a philosopher or similar that they send you down the ramen path.
I don’t know where you got this notion.
seems like a pipe dream to me.
The answer is no. A life of material richness vs being poor and highly educated. Ask a postdoc this question and I bet the majority would take lifetime wealth to having to slave away in a lab.
If someone takes the approach that school is required for learning, then they'll stop learning after finishing university (as most people do) and within a decade they'll be stupid compared to someone who believed learning is a personal thing and kept learning after their formal schooling had ended.
First you're projecting your personal experience in your (I guess crap) school in a crap educational system to the norm. It's not.
Second, jobs having higher educational standards is a reflection of the booming of educated youngs, not the other way around.
Third, skilled uneducated professionals (e.g. plumbers, cooks, electricians, mechanics) make much more than graduates on average, it's a night and day difference, comes also debt free and you can start compounding much earlier. If money is your goal an education is actually not necessarily needed if not dangerous.
And no, the stats say that tradespeople are massively underpaid on avg compared to BS holders. Ofc, no one believes this here since HNs demographics are highly clustered in the bay area where trades people basically just tell you how hard you're going to bend over for them...
I'm European, know our school systems in various countries and it has no resemblance to the ideas you have about education and we don't really have such gatekeeping at college at all.
Third, about trades, I've never lived in SF, but I lived in italy/poland/switzerland and even 6 months at OSU in columbus and I've seen that all the time: skilled tradesmen in all these places make MUCH more than average graduate. There's no comparison. I'm 100% sure the average decent carpenter or electrician makes more than a stem graduate in Italy or Switzerland.
Don’t pretend, we will see you ;)
exams will show what you’re really capable of.
lol
And with all of that extra time, I'll study for your "exams".
Just replace chatgpt with Wikipedia, did this for 4 years of uni and another 2 at an ivy n school. No one was the wiser and now I have a cushy MBB job to show for it.
I get my question, research it, provide chatgpt with an outline and ask it to fluf it out in a nice style and maybe copy it down once. Even if I only remember 80% of what I wrote then that is still a win.
I too think that's exactly what's wrong with the education system nowadays - too much homework and not enough time for drugs, Call of Duty, $120 apparel from LuluLemon, browsing TikTok all day in bed, and sending nudes over SnapChat.
Childhood. Solved.
As of my 30s, I can now mostly read warning labels and some of the Spanish manual if I happen to grab the wrong one out of the box. Conversations with a native speaker? No way.
If that wasn’t a complete waste of my time, I don’t know what is.
Presumably, from your point of view, all of the time invested into it was a waste of time, not just the homework part. That said, it doesn't follow from this that any homework is a waste of time. It doesn't even follow that Spanish homework is generally a waste of time.
One cannot generalize from your example at all because you might be an outlier. You'd have to show that learning Spanish provided no value to the majority of kids who studied it, or that class time is sufficient to learn Spanish so homework isn't really needed, which would be a very hard thing to argue for a human language.
For what it's worth, I took Bulgarian folk music - I'm not Bulgarian, nor do I live in Bulgaria, nor can I play any musical instruments. Was it a waste of time? Acquiring knowledge is never a waste of time for me. Typing comments on HN might be though.
but if i were an english teacher assigning research-like papers, i'd definitely incorporate chatgpt responses into my assignments. make them write a draft with chatgpt3 then revise it themselves in class with pen and paper. then write a version in their own words without plagiarizing the chatgpt version, but they can only use google docs or some format that can track changes, so the teacher can see the process. or something like that — it would need some iterating and real-world feedback.
now THAT is a creative thing to do. I like that. I suppose the calculator version would be to do something like: "here is a simple graph plot, create an equation that produces a similar plot"
Wish there were more teachers like you embracing this stuff as a tool rather than doing your best to design a curriculum to 'defeat' it.
Similarly, OpenAI's tools can be used to create all sorts of content marketing and other webspam. They they'll create a monetizable search engine-killer that knows how to ignore all this fluff (which will probably constitute 99% of the internet within a few years).
I wonder if we'll look back one day and say "remember back before 2020, when the internet was filled with junk, but at least it was human-written junk?".
I doubt a reliable technique for this is even ever possible.
While these detectors are of course interesting from the hacker/academic points of view, I don't think they will really solve much in practice.
On the one hand, our (teachers') adversary is not ChatGPT. It's ChatGPT + a human, which will be aware of these techniques and can tweak the text to avoid detection.
On the other hand, even if we could detect generated texts with high accuracy, are we going to fail a student just because a black-box AI model says so? Doesn't seem very ethical to me. It's funny how some people worried about ChatGPT's ethical implications seem to be OK with this, just because it goes against their personal demon.
I don't think there is any realistic alternative to just assuming that these systems will be used wherever text is required and you can't physically watch the writer, and living with it.
My guess is that students just have to spend 3-5 hours on a Saturday writing essays, doing programming assignments etc…
Do us a favor and get the COVID vaccines. They're safe and effective!
I like ChatGPT and spend quite a few hours a week working with it; it's a lot more interesting than just a text generator. That said, people using it to do schoolwork for them are short-changing themselves, because they're opting out of developing core cognitive skills.
We already have that, today.
With fantastic tools such as TurnItIn[0], that regularly flags my own work as plagiarizing my own work. And we see with those tools, already, there is a presumption of guilt. The professor will believe the tool over the student, often without the slightest glance to see why said tool is saying the student is guilty.
[0] https://www.turnitin.com/
Why are we making kids write entire persuasive essays and research papers when in the real world these are almost always the result of collaborative effort?
If you think of ChatGPT as a fancy calculator, you can just apply the same techniques we've been using for a while: tell students not to use them on homework, make homework only a small % of the grade, and rely on in-class, paper only exams for most the grade.
I think we make kids hate school via "sucking it up" when we really don't have to. Learning to write in the right context is much more effective than learning in the wrong context. But our educational system is so standardize that we have to fit most kids into the same interest peg.
These days, that would be considered insane! But in 1989...
[1] https://en.wikipedia.org/wiki/Itzhak_Perlman#/media/File:Itz...
If we habitually do hard things, then we generally have no problem, stepping up to the plate, for when it’s showtime.
Same reason that musicians practice scales, all the time.
I write code, every day. I walk 5K, every day. I solve problems, every day (detect a pattern, here?). It makes showtime a lot less stressful, and folks seem happy with my work.
First thing I do, is a 5K brisk walk. Gets the exercise quota satisfied. I tried running, but kept injuring myself, so walking, it is. I can do that, seven days a week. Takes a bit less than an hour (I walk about 10min/Km).
I usually "triage" my day, on these walks. I mentally sort through the tasks ahead, and scope out my plans. Sometimes, it's straightforward, other times, I have to develop a plan.
When I get home, I hit the shower, and generally start coding, right away. I code throughout the day, taking breaks as I need/want. I have no issue, stopping work for a couple of hours, and watching some TV or even taking a short nap.
I'll code at any time. I don't have a "set" workday.
I have found that I am most productive in the morning, and most creative at night, but I sometimes find myself losing focus, towards the end of the day. When that happens, I generally wrap things up, as I do more damage than good. I'll often fix a bug in my head during my walk, that was confounding me the night before.
I believe in F^3 - Finite, Focused, and Finished.
- Finite: I know what "done" looks like. Tasks have a discrete beginning and end, though the schedule may be fuzzy.
- Focused: No distractions. I stay on beam, and devote full attention to the task at hand.
- Finished: I don't stop, until I have reached a point that I can wrap things. I usually punctuate this with a final tagged commit.
It's also important to scope things reasonably. Stretch, but not too much.
I write about that stuff here: https://littlegreenviper.com/miscellany/thats-not-what-ships...
I probably have a life a lot of folks around here would not find attractive. I find it WFM. YMMV.
Similarly, this is why we students have been required to cite sources instead of simply rendering an opinion on a topic. Most of these AI bots are just fancy plagiarism machines.
An acquaintance who works in the NLP space says this is done in post processing, where GPT probably hallucinates a reference but they run a web query or something to try to fix it up.
> find appropriate sources and cite them later
Edit: by Escape Room, I mean like the fun kind you would go to with friends or colleagues. I'm not saying we should torture children... Google if you haven't heard of one.
This might be useful to filter spam from places that initially got bombarded, like stack overflow. I agree with other comments, it's hard to judge the value of this tool without false positive/negative stats. The output isn't completely deterministic though so I wonder how "certain" this tool can be.
There will be paid ChatGPT services for cheating students, the ideal solution would be to rethink and rehaul how student performance is evaluated, but universities will instead pay something like Turnitin because throwing money at a problem is always easier than thinking.
>they hear nothing about it because its no longer their business
how are you so certain that
>students certainly get expelled
?
ChatGPT is not an author, it's a computer program. We don't expect bibliographies to include Microsoft Word, Grammerly, or iOS Keyboard Predictive Text in the list of references. If I use a piece of software to create an essay, the essay that's created belongs to me.
So a student that takes said output and hands it in is plagiarizing by claiming credit for something written by a different "author".
It'll probably take some time for language to evolve around non-human agents. There's no intrinsic reason to prefer separating (or combining) that terminology by humanness of agent, afaik. At least, for common usage. One will win and only time will tell.
Also, There's a commonly understood threshold for plagiarism where a single word won't qualify and a phrase probably won't qualify, but a whole paragraph does. That is why Grammarly isn't plagarism, not because it is software.
So in an academic context, for grading your understanding, it becomes meaningless to grade something written by a machine instead of by you. In other contexts, it's totally fine.
Where is the line between ChatGPT and Grammerly? Grammerly won't produce entire paragraphs, but it absolutely will rewrite sentences.
If I ask a human to write a 500 word essay on the invention of the computer, but give them the prompt, that is absolutely plagiarism.
If I ask Grammerly to write a 500 word essay on the invention of the computer, but give them the prompt, that is still absolutely plagiarism.
No, because someone else was the chef in that case. But if I bought a conveyer belt machine that made cheeseburgers automatically at the press of a button, and I used that machine to cook dinner—maybe!
If we lived in a world where machines like that were common, the meaning of "cooking" would change...
Maybe a better example is popcorn. I can make it in a microwave, or I can make it in a pot on the stove from scratch. It takes about the same amount of time, and the quality is about the same. But doing it by hand involves many more small decisions, like the amount of oil, corn kernels, heat, cook time, knowing when to stop, what flavor ingredients to add and how much etc. Now I can do it on autopilot and never have a single burned or unpopped kernel, but it took maybe 100 less-than-perfect efforts to get to that.
Students need to learn how to produce the best writing they can. If their best writing comes from feeding a prompt into ChatGPT, they should use ChatGPT. Presumably this student will also use ChatGPT in their writing post-graduation.
Of course, I would expect such a student to get marked down for factual errors, logical inconsistencies, and a lack of citations!
Some students, however, may take a hybrid approach, in which ChatGPT writes certain passages which the student manually cleans up. IMO, this is completely acceptable, and probably a good skill to practice! Students should learn how to use technology to produce the best work they can!
There are absolutely exceptions, when you'd want a student to practice writing independently. But writing the whole thing off as "plagiarism" is going much too far!
The same goes for composing an essay. It's not just about the ability to produce a cogent-looking essay, in the same way it isn't good enough that a kid can punch 17*13 into a calculator.
But, yes, calculators shouldn't always be allowed, and neither should ChatGPT. But, well, "plagiarism" is never allowed, so IMO the distinction matters!
By your logic could I not claim that I drew that ensuing picture because I "wrote the prompt"?
Let's take it a step further, what happens if I take my prompt and hire an artist on fiverr? We go back-and-forth via email as I ask them to make various changes and then wind up with a final drawing.
Am I still allowed to make the claim that I drew that picture?
I feel like "drawing" is a very specific action involving a writing utensil. I don't think I'd consider an image made in Adobe Illustrator to be "drawn", unless you were using a Wacom or similar.
But, yes, I'd say you could claim you "made" or "designed" the ensuing picture.
> Let's take it a step further, what happens if I take my prompt and hire an artist on fiverr? We go back-and-forth via email as I ask them to make various changes and then wind up with a final drawing. Am I still allowed to make the claim that I drew that picture?
No, because the artist on fiverr drew the picture.
Let's tweak this analogy a bit. Let's say you had a photograph of two people holding hands on a crowded beach. You want to retouch the photograph to remove the crowd, so the beach appears empty.
If you're lucky, you may be able to use Photoshop's Content Aware Fill to remove the crowd in less than a minute. In this case, you could absolutely claim that you retouched the photograph, even though Photoshop's algorithm did all of the work. A computer program can't be the creator of anything—the person using the program can.
If you're unlucky, Photoshop's Content Aware Fill will produce bad results, and you'll have to painstakingly go through the photo by hand. If you're short on time, you could hire a retoucher on Fiverr. In this scenario, however, you wouldn't be able to rightly claim that you retouched the photograph, because another human did that.
---
In every other case I can think of, the person using a tool gets to take credit for what they've made with the tool. Use an electronic screwdriver to fix a car? The screwdriver didn't fix the car, you did! Use a calculator to add two large numbers? The calculator didn't find the sum, you did.
None of this should change just because we have better tools. Schools may place additional limitations on students for certain assignments—math classes may restrict the use of calculators, art classes may restrict the use of drawing—but these limitations (1) should not always be in place for every assignment and thus (2) should be stated explicitly up-front.
Although I thought it would be trivial to work around this heuristic by specifying "use some short sentences and some long sentences", that didn't really work for me. When I said "use only short sentences", the average sentence length was a bit shorter. But I was surprised that this type of meta instruction didn't work better (at least in this version).
I'm betting on schools using expanded remote proctoring software to monitor and record students' homework process and flag potential irregularities for teachers to check. Whether it's positive or negative for students' education, it will be less disruptive to schools than eliminating homework or significantly altering the curriculum.
But you can fool it by first asking chatGPT if it can write in the style of some known author, letting it output an example, then asking it to write about a given topic in the same style.
Here's generic output to "write a one paragraph description of Genghis Khan". The Hugging Face output detector says it's 99.98% likely to be fake (from GPT):
> Genghis Khan was a Mongolian warrior and ruler who founded the Mongol Empire, one of the largest empires in history. He is credited with uniting the nomadic Mongol tribes and establishing a powerful and efficient military force that conquered much of Asia and Europe in the 13th century. Genghis Khan is known for his strict code of law and his innovative use of cavalry and military strategy. He is also remembered for his brutal treatment of conquered peoples, as he and his armies were responsible for the deaths of millions of people during their conquests.
But this response, in the style of Charles Bukowski, gets a 98.66% real (not GPT) rating by the checker:
> Genghis Khan was a man of fire and fury, a conqueror who swept across the land like a storm. His armies were vast and unstoppable, crushing all who opposed him beneath their hooves. He was a man of vision and determination, who united the scattered tribes of the steppes and forged them into a mighty nation. But he was also a man of blood and death, who left a trail of destruction in his wake. Villages were burned and pillaged, their inhabitants slaughtered or taken as slaves. The world trembled at the mention of his name, and even the greatest kings and emperors bowed before him in fear. Such was the power of Genghis Khan, the greatest warrior the world had ever known.
But wow, the second response was pretty exciting. I wonder if chatgpt can get present-day authors to up their game?
By the way, here's the response to "can you write in the style of Charles Bukowski?" (the first prompt in the session):
> Sure! Here's a little something I came up with: "I woke up this morning with a hangover and a broken heart. The bottle was empty, just like my bank account. Love had come and gone, leaving nothing but pain and regret in its wake. I stumbled out of bed and into the shower, hoping the water would wash away the memories of her. But as I stood there, staring at the peeling paint and the mildew on the walls, I knew that the only thing that would cure me was time and more whisky. So I dressed myself in my finest rags and headed out into the city, searching for the next fix, the next distraction, the next moment of fleeting pleasure that would take me away from the pain. Because in the end, that's all we really have: the present moment and the hope that tomorrow will be better, or at least no worse than today."
Within the context, this is a ambiguous in a way that would be unrealistic, like a bit of a micro-Chekhov's gun.
As I'm sure many of us did, I tested out the app here on GPT-3 output. Unmodified, it detected it was GPT-3. Great! However, I added about 10 additional words to the output provided by GPT-3, and it shot up my "human" score by like 60 points, and determined it was human-generated.
This is going to be the problem underscoring _any_ model that is trying to identify "AI" generated text. A human can modify it slightly, or subtract words, and it throws the entire thing off. There are other paths that we need to explore to this problem.
In all seriousness though, this is really a cat and mouse game. Working in CV I might be biased to think that images are going to be easier to detect than text, but we are going to get better. I'm not saying we shouldn't create detectors (we should), but I think we also need to be aware that this is a cat and mouse game and we need to have true social conversations about this. Though we still fail to have these conversations when it comes to computer security so maybe we're doomed (or maybe this will be the catalyst). It's also important to note that a lot of damage can be done even with images and text that are easy to detect as fake. I've seen plenty of fake Twitter and Linkedin accounts that use StyleGAN profile pictures (with all the telltale signs).
I'm curious what kind of model you're using? How interpretable is it?
Then is it a problem if a human slightly modified an AI's work? Isn't that the desirable outcome? And if the work itself is so useless that it would be worthy of a zero grade, then perhaps another way of measuring usefulness should be used.
In a way, I feel like we can't impose old-world grading techniques to new-world content synthesis technologies.
I have no regrets about that.
I wanted to learn math, science, and engineering. I didn't want a degree, I wanted to learn the stuff. What was the point of cheating? I've derived a lifetime of pleasure from knowing that stuff.
I don't understand wanting a liberal arts degree and then cheating to get it.
Employers know about the cheaters, fakers, frauds, and charlatans. The job screening tests are an attempt to screen them out. Nobody wants to hire cheaters.
But the expected value of that choice is positive (for the person making it), as evidenced by the people making it routinely.
I get you don’t like that, but that’s the reality of US labor market, where useless credentialism has crept in:
Many people get degrees they don’t want for jobs that don’t need them to check a box and get hired - because as a society, we’ve built that into our hiring.
Particularly liberal arts degrees.
Heck, as scandals have revealed, they frequently become corporate CEOs or get elected to high office.
Writing was easier than studying for me.
For this reason, outside of my math & science majors, I chose exclusively courses where the grade was based on written work.
But likewise, I wasn't in college to get an empty credential. I already had a marketable skill -- programming. Learning was fun, but I was also conscious of the fact that I didn't want to waste my parents' money.
The writing has affected my career: My work involves not necessarily writing papers, but certainly communicating ideas and knowledge to other people, including non-specialists.
As co-founder and CTO of any number of tech companies, I've observed engineers (programmers) with broad humanities (history and language related, mostly) education tend to grok business narratives then architect and write better software: coherent systems expressed concisely. Now I look for history major with comp sci minor, or English and linear algebra, or other self-selected humanities + STEM pairings.
I'd hazard there's some relationship in how the mind is applied in common between domain driven design in service of business narrative and the sweeping cause and effect of history, or between software architecture and well organized long-form writing.
https://www.digitalmars.com/articles/index.html
Consider that the people I know who learned math, science and engineering without going to university hovers around zero. I've also known degree'd engineers who were so hopelessly inept I suspected that they cheated their way through. Their careers didn't go well.
I once knew an MIT graduate who thought a coordinate system conversion matrix was black magic (I wrote it to convert screen graphics to the printer's idea of a page). I was floored. Howinell did he get through MIT?
Ah, so magnanimous of them to sell both the plague and the cure.
I say that tongue-in-cheek of course, but it seems that's where this is going, whether they make and sell both or let others do it. That is, both sides of the production and detection of this spam are about to profit handsomely from the overwhelming piles of it we're about to be drowning in.