A.I. and the Future of Cheating | Hacker News Reader

A.I. and the Future of Cheating

onezero.medium.com

43 pointsbenbreen45 comments

> What happens when universities can’t tell whether an essay is written by a human or an algorithm?

Then in that scenario, that means that essays are worthless at evaluating students, and if AI essays really are as good as human essays, then teaching the writing of essays at the college level is a waste of time.

However, looking at the generated essay, it seems to me that the real problem is that universities have reduced essay grading to an algorithm which values form over the content of the essay. The fact that AI can generate content with the required form should come as no surprise to anyone. Perhaps we should be teaching critical thinking and reading skills to evaluate the actual content of any essays.

Computers have been able to do arithmetic faster, better, and cheaper than humans for decades. But we still teach children to do arithmetic, and grade them on how well they do it (almost never as well as a computer). Why? Because doing more useful and creative math is a lot easier if you can do a little arithmetic.

The purpose of college essays has never been the quality of the product. The purpose is to have students practice so that, when they really do have to convey something in writing, they can convey it well. If an idea exists only in my head, it will take a very long time for an essay-writing bot to write it for me.

> The purpose is to have students practice so that, when they really do have to convey something in writing, they can convey it well.

Then have a human grade it. Have personalized feedback instead of a cookie-cutter grading that's just another data collection scheme + some basic keyword matching to catch stupid mistakes.

Speaking from experience, the person that writes best code (in terms of both raw quality and readability) also seem to write the best docs.

Maybe teaching essays are sideways to efficient communication?

Although we have machines that now cut and bind books, the best book binding (and most aesthetic) are done by hand by enthusiasts on, e.g., Reddit. Moreover, such book binders don't feel threatened by machines, I would think, like how people who do DIY do it because they like it, rather than because they have to do it.

Even if an AI can write an essay, that would not stop me from writing. Essays have another important dimension, which is that of creativity and the message. I think the only real relevance is in the context of school work, outside of that I don't think the AI's ability is that important. People are intelligent already, and competition between writers or the general tendency to influence each other sometimes is even a good thing.

I think you're right - an essay should communicate a key thought, broken down in logical steps, coherently tied together. This is important for a whole host of skills.

Is it possible you have cause and effect reversed? I received this as "learning to code well can also teach good communication skills"

Students could just be required to write the essays in class rather than as take-home assignments.

Also, if there's a great discrepancy in style, skill, or content of in-class essays vs take-home essays, then that could be considered evidence of cheating.

This doesn’t address the first point of the parent, which is that if AI can pass the essay Turing test, then perhaps college students shouldn’t be paying thousands of dollars and spending hundreds of hours learning to produce essays? If essays are a proxy for some other process, perhaps we should find a better proxy.

Am a university professor who frequently assigns essays, so this is something I've been thinking a lot about. The problem is not that AI can successfully write a college-level essay (maybe someday, but certainly GPT-2 isn't even close to that level yet). It's that something along the lines of GPT-2 can emulate the superficial appearance of an essay. The actual facts, quotes, and references it generates are generally nonsensical, but the structure, grammar and language seem authentic at first glance. Unfortunately, a lot of essays in university lecture courses are graded by underpaid TAs who are skimming quickly, and who might be inclined to give an AI-written essay a D or a C- for effort rather than recognizing it as fraud.

The issue isn't so much that AI has replaced essay writing. It's that it has the potential to create convincing (but fundamentally nonsensical/illogical) replicas of academic essays, which have the potential to make an already difficult job of assessment even harder (somewhat analogous to the problem of deepfakes).

By the way, I'm currently thinking about writing something on this topic; if anyone has any thoughts please get in touch. Contact info in my profile.

>>> Unfortunately, a lot of essays in university lecture courses are graded by underpaid TAs who are skimming quickly, and who might be inclined to give an AI-written essay a D or a C- for effort rather than recognizing it as fraud.

This raises an interesting point, which is that there seems to be a relationship between essay-writing and essay-reading. The students are able to cheat on essay-writing because the school is already cheating on essay-reading.

If we reach a point where essay-writing is unnecessary because it can be done by machine, it's a safe bet that essay-reading will also be done by machine. What will be the point of writing me an essay, if it's just a function of a prompt? Why not just send me the prompt?

Then professors will start to teach and assign the writing of prompts for essay-writing machines. "Write a prompt in less than 25 words that persuades analog31 to approve your project." And someone will come up with a prompt-writing machine, and a prompt-reading machine.

>This raises an interesting point, which is that there seems to be a relationship between essay-writing and essay-reading. The students are able to cheat on essay-writing because the school is already cheating on essay-reading.

When they pay so little why would a person spend hours reading and analyzing one essay. The bigger issue if the lack of resources devoted to grading and not that the people on the ground are cheating.

edit: I think the commentator's emphasis should be on the lack of resources and not a supposed negligence on the grader's part.

Indeed, I would say the system is cheating.

Imagine a model trained just on college essays, and you feed it an outline. A sentence or two intro, some topic sentences, and citations.

Probably if someone had a dataset of essays large enough, you could do this right now and cause a lot of trouble.

I frequently wonder why the cost of college is rising while the teaching / grading is relegated to teacher assistants, and now, AI.

Writing essays teaches people how to think and express ideas in a manner that allows the reader or listener to be engaged and also think and consider what’s being delivered. People who “wing it” are noticeable and don’t get the same engagement.

Doesn’t like /every/ good job require something along the lines of “Excellent Oral and Written communication”? Essays aren’t a proxy for anything, they are an end of itself. Just because this test essay can be written by an algorithm doesn’t mean that an algorithm can create actual original content you’d be expected to put out at a job.

The essays can be a learning or training exercise for the student. Required, but not graded. We had plenty in high school like this.

If the student gets the AI to do it, so be it. You'll get useless feedback and then you'll have to sit in a room and do it for real in the final exam - maybe you didn't need the practice, or maybe you'll wish you didn't waste your chances.

Exactly, it would be like colleges grading on the ability to add up long columns of numbers and complaining about spreadsheets.

At one time, being able to add up long columns of number accurately using only a pencil and paper was a valuable business skill, now not so much.

Stylometry is a hard task to be performed by humans and by computers... I've studied it for a while (master degree and PhD), and while we could have some success (>80% in verifying the authorship of a document against 100 authors), I think that it would be hard to make judgements solely based on a percentage below 99.9999%.

OTOH, perhaps I should begin to study this matter again :-)

We'll have to develop AIs that are trained to detect if the essay was written by a human or by an AI.

This could turn into a pretty interesting arms race with surprising privacy considerations. For example, the AI might require writing samples from the student under consideration from K-12 in order to get sufficient accuracy..

In a GANN, they are already trained together.

> Then in that scenario, that means that essays are worthless at evaluating students, and if AI essays really are as good as human essays, then teaching the writing of essays at the college level is a waste of time.

Case in point-- no one plays chess anymore.

There is a difference between hobbies/sports and work.

We still have horse riding, but FedEx doesn't evaluate horse riding ability before hiring delivery drivers.

I'm not sure where you got that impression. The lowered attendance rate of in-person chess clubs is more than made up by the enormous amount of chess played online, high-level professional chess is doing better than ever, and it's seen by many parents as one of the most generalizable skills for a child to learn.

(It's possible I just took a joke too seriously, but "chess is dead" is a common and incorrect take.)

I had a teacher who did pass-fail essays. His rubric only had one box "Did I feel anything while reading this essay?" I think the technology still needs to develop much more before it can fool Mr. Thayer's grading system.

Then t

The article quotes Miles Brundage:

“If it is not being used in an illicit fashion, A.I. can be a boon for generating larger, more ambitious, more creative text documents — similar to how calculators have been used to help develop the skills we want to be teaching.”

and it continues:

Calculators have spared maths and science students the laborious pain of long calculations and allowed for more advanced syllabuses, designed with the knowledge that students will have access to calculators when answering them.

The issue I have with this is that the role of the essay in education is not comparable to that of the sort of tasks that calculators automate. If there is an analog of calculators in essay-writing, it would be things like grammar-correctors and style checkers.

If AI could be used in the manner envisioned in the quote, we could end up with people writing essays that they do not understand, and presenting arguments for positions that they could not defend. I am not sure that this would be progress. It might be a way for the machines to take over.

Although I love to write, I can imagine a world in which the construction of sufficiently detailed outlines, along with some sort of style examples, could be automatically converted into a full-fledged essay that elaborates on the connections entered by a student or individual.

> If a university has no way of determining whether an assignment was written by a human or an algorithm, existing grading systems lose any semblance of meritocracy or fairness.

Just because it was hard to tell if it was written by an AI doesn't mean the paper succeeds at whatever the assignment was. All GPT-2 does is predict the next word. I don't see how this can be used to write a research paper.

XLNet and Transformer-XL can maintain a memory of the last thousand or so tokens. The results are pretty remarkable. Check out the hallucinated Napoleonic-history wikipedia article at the bottom of https://arxiv.org/pdf/1901.02860.pdf

not that i believe (or not) that this will eventually be feasible, but the dream is that if you feed an introductory paragraph (which includes a thesis) to the algo, it will be able to generate the next 4 paragraphs, which completes a 5 paragraph essay.

What I would be more interested in is feeding it a topic, it goes out and summarizes sources, then arranges those summaries to make a point. I'd actually really like this driving around. I could just have ai give me the tldr on topics I'm interested in.

The future of cheating is large corporations and billionaires using automation to accumulate more society points without giving back to the hard working people who make it all possible in the first place.

* Note: this is also the present of cheating.

* * Note: I hope this cynical statement of mine doesn't come true.

The future of cheating is no one having to work so hard, because machines are doing all the boring work.

I like your idea better

Note that this was posted September 13th; before the full-size GPT-2 was released yesterday after concerns about GPT-2's danger were unproven.

The practical use case for GPT-2 in writing is writing assistance/autocomplete, such as with Write With Transformer (https://transformer.huggingface.co). Text that partially AI-written would pass a Turing test so-to-speak.

Highly recommend testing this new cheating tool out by entering a custom prompt at https://talktotransformer.com/ (which uses GPT-2) and reading what the AI generates for your custom essay snippet.

A single sentence with a euphemism should suffice...

"I" was able to create some hilarious franken-stories. Anyone care to share their greatest hits?

It seems to have a habit of responding to writing prompts by continuing to write the prompt. College essay fodder like "Write about a situation in which you overcame adversity." produces output that continues with "This doesn't have to be a life-changing event, but you have to be honest." before diverging directly into bullets 3 through 5 of a numbered list of other prompts like "A story about how you did something right or how you have a great life now, and then you go out and do it over and over again.".

On the other hand, I get interesting responses if I instead provide it with the first two sentences of an answer to a prompt. I tried "I was only eleven years old when he died. I started playing soccer and it made me who I am today." and got:

"He died from complications of chronic renal failure. It's just a really, really bad way to die.

You think about that. You're talking to a kid who just died from a medical issue that was preventable, if not treatable. And you think about the kids who die on the soccer field, all over this country, and you think about how that's not OK."

It's not going to fool anyone, but it's a big step up from anything I've seen before.

That was interesting, but it is no Crashtest Dummies.

Prompt: Once there was this kid who got into an accident and couldn't come to school.

Essay: The police told me they would call me if they found out if I would come. And I couldn't wait.

I knew how to talk to police because I had done that all of my life. But this guy was different. I talked to the cops and told them that I wasn't going to tell on him and they were cool with that. They let me off easy.

I wasn't going to come on to him, because I knew if he found out that I'd call him the next day I would probably be fired, so it was just a waste of time.

The next morning the police called me, and I had to sit there for five hours. I was really nervous and the whole time I was thinking about how I'd never really seen this kid in my life. And then after they told me that I was on the payroll I was really excited and I kept thinking that if he called me I would have to tell him that I'd never been fired before.

Then when I was going over the names in his file

At first glance I was Wow. But then it went downhill very fast.

Prompt Is Capitalism going to win at the end? Answer I don't think so. I think that, in the end, the people have won this battle, the people have won this fight. The people have done everything they could to take back their lives. They are tired, but they still want to have a future. The fact that capitalism has not won is because the people are still fighting.

You can read more on the struggle and the victories here: www.sabr.org

Prompt Is Communism going to win at the end? Answer So, if you've listened to this podcast for a while, you'll know that I'm a huge fan of the book "Against Our Better Judgement". It's a very well-known book which argues that the current financial crisis is due to bad investment decisions made in the last decades by Western nations. These investments were largely based on faulty economic reasoning which has resulted in high levels of unemployment and under-investment.

In this podcast, I'm going to take a look at the book from a slightly different angle, which is why it was written. What I will argue is that it is also very relevant in today's world because the same flawed reasoning is still in place, which is causing the current situation in the first place.

My arguments are going to be based

Prompt Is gravity going to win at the end? Answer We won't know the answer for sure, but we do have a few clues about how the final battle might unfold.

Here's what we know:

Gravity will win at the end: The most obvious thing to do is just to say gravity is going to win. But what does that really mean?

If we assume the universe is a static two-dimensional space and all matter is a singularity-in-a-bottle, then the amount of mass in the universe is a constant.

What will it take for the universe to collapse? One way to put this is that the amount of mass in the universe will double every second, so the universe will expand by about 1.5 billion kilometers in the time it takes you

------- Notice some really weird text like Advertisement, random web site in the reproduced text etc. This makes it look like copying data from various web sites.

So I've been playing around with GPT-2 and wow. It will write about anything. It is like I am talking to everyone on reddit, all at the same time. And I mean E V E R Y O N E on reddit.

GPT-2 is an abomination.

Imagine this use-case:

1) Copy some text to be plagiarized.

2) Feed text to AI so it's rephrased.

3) Present result as own work.

Interestingly enough, I fed those lines into an AI[1] and it spit out:

1) Start copying some text from another source.

2) Start rephrasing text.

3) Present result as own work.

You are now a master plagiarist.

-----

So I think you're onto something. The AI is a far faster rewriter of text than human beings.

[1] https://talktotransformer.com/

I can see that item 3 is the same. Still very impressive.

Reply on news.ycombinator.com