Every undergrad study is, as of now, possible to pass with AI models. Some of them not only pass, but pass with flying colors.
The harsh reality is that academia as a whole needs to be revamped. The easy solution would be to revert back to paper only exams, and physical attendance - but that would also exclude a ton students. A huge number of modern students are online students, or similar programs where you don't need to show up physically. Moreover, I don't think universities / colleges themselves want to revert back, as it would mean hiring more people, spending more on buildings, etc.
So you gave the easy solution. What's the hard solution?
Honestly, the pervasiveness of LLMs looks to really erode the critical thinking of entire future generations. Whatever the solution, we need to be taking these existential threats a lot more seriously than how we treated social media (the plague before this current plague).
Ask students to solve harder problems, assuming they will use AI to learn more effectively.
Invert the examination process to include teaching others, which you can’t fake. Or rework it to bring the viva voce into evaluation earlier than PhD.
There are plenty of ideas. The problem is, a generation of teachers likely need to be cycled through for this to really work. Much harder for tenured professors.
Every technical revolution “threatened to erode the critical thinking of a generation”, and sure, the printing press meant that fewer texts were memorized rote… not to say there are no risks this time, but rather that it’s hard to predict in advance. I can easily imagine access to personalized tutors making education much better for those who want/need to learn something.
I’m more worried about post-truth civilization than post-college writing civilization for sure.
I think the hard solution is to massively increase expectations. Think Star Trek where the grade schoolers are learning quantum mechanics. If everyone has access to the oracle of all human knowledge, then you should teach and test to the maximum of what a student could do with all that power. Find the frontier where the AI fails and the human adds value and teach there.
> Honestly, the pervasiveness of LLMs looks to really erode the critical thinking of entire future generations.
Yes and no.
Upper middle class parents as a group will still instill critical thinking skills in their kids.
But the above comment reveals more about SES (socioeconomic status) and education in general rather than something specific to critical thinking or LLMs. The current education environment in the US heavily favors kids from higher SES families for a number of reason. LLMs won’t change this.
The challenge for the education system, imho, is to find a way for lower SES kids to thrive in an LLM environment. Pre-LLM, this was already a challenge, but was possible. Post-LLM, the LLM crutch may be too easy for some lower SES folks to lean on such that they don’t develop the skills they need to develop higher order skills.
I suspect this is the true fermi paradox. Once a civilization reaches a certain point, automation becomes harmful to the point that no one knows how do anything on their own. Societal collapse may be back to bronze age, if not more regressed.
I hold remote interviews and I can tell when candidates use AI to answer questions, real time with camera on. They repeat my technical question, pause for a few seconds, their voice drops to a monotone and they quickly recite a bulleted list of low level technical details that sounds like a wikipedia page. I worry that candidates will learn to act more subtle, maybe configure their LLM to return with an anecdote around the tech in question, and practice "selling" their vocal communication.
Most of this stuff isn't that hard to learn and apply. If they put that much effort into setting up an LLM cheating tool they could have just built something with whatever tool you're talking about.
One of the things I would say I saw coming from a long ways away is that a lot of effort was put into detecting the "default" voice of LLMs, but people would eventually figure out how to kick them out of it, since it is after literally just a matter of asking. It took some months but word does seem to be finally getting out, I'm seeing the idea emerge in more and more places.
I gave it the prompt "Suppose you are in a job interview for a front-end web position and someone asks you about how you use the React library and the hardest problem you even had to solve with it. How might you react, along with a somewhat amusing anecdote?"[1] and it did pretty well. I think I'd play with it a bit to see if I can still suppress some of the LLM-isms that came out, but a human could edit them out in real-time with just a bit of practice too... it's not like you can just read it to your interviewer, you will need to Drama Class 101 this up a bit anyhow. It'll be easier to improv a bit over this than a bare Wikipedia list.
In other words, as with the question the article title asks, the question isn't about what happens "when" this starts being possible... the capability has run ahead of all but the most fervent AI user's understanding and it is already here. It's just a matter of the word-of-mouth getting around as to how to prompt the AIs to be less obvious. I also anticipate that in the next couple of years, the AI companies will be getting tired of people complaining about the "default LLM voice" and it'll shift to be something less obvious than it is now. Both remote interviews and college writing are really already destroyed, the news just hasn't gotten around to everybody yet.
(In fact I suspect that "default LLM voice" will eventually become a sort of cultural touchstone of 2024-2026 and be deliberately used in future cultural references to set stories in this time period. It's a transient quality of current-day LLMs, easy to get them out of even today, and I expect future LLMs to have much different "default voices".)
[1]: And in keeping with my own philosophy of "there's not a lot of value of just pasting in LLM responses" if you want to see what comes out you are welcome to play with it yourself. No huge surprises though. It did the job.
> They repeat my technical question, pause for a few seconds, their voice drops to a monotone and they quickly recite a bulleted list of low level technical details that sounds like a wikipedia page.
It may get resolved on its own. These days people study to get good grades in order to prove to future gatekeepers (like employers, or higher rungs of academia) that they know the material well. Post AGI, however, the gatekeepers may not be so interested in humans anymore, and we might not need grades at all. Studying anything could become something done exclusively for ones own interest, and the only point of a grade would be to give one-self a goal to achieve.
Alternatively, if we still want to cling on to this ritual of measuring the performance of students, you could give each and every one of them oral examinations with AI professors.
We are not post AGI. We still need to hire people who did not cheat their way through school and interviews with chatGPT. Even post AGI, we would want to hire qualified people.
Institutions that prepare people for future jobs have an even harder time to justify what they’re doing than the people who are looking for jobs right now. It’s just inertia at this point.
Not to mention that AI can educate the people better by solving Bloom’s Two Sigma Problem.
So colleges are obsolete except as four year cruises for entertainment and networking.
Don't even need to go that far. Provide a locked-down computer and have students write essays in a dedicated space. I have personally done that and it was a reasonably good experience.
Which raises the issue of what education is for. Is it to know things and solve specific problems in a controlled environment, or is it to work with available tools and resources across a range dynamically changing contexts? Does being at a locked down computer in a dedicated space match likely work settings?
1) Written in person exams that were most of the grade (this includes "blue book" exams where you have to sit in front of the professor and write an essay on whatever topic he writes on the board that morning as well as your typical math/algorithms tests on paper.)
2) Written homework where you have to essentially have a satisfactory discussion on the topic (no word range, you get graded on creative interpretation of the course subject matter.)
Language models could maybe help you with 2 but will actually kill your ability to handle 1 if you're cheating on homework with them. If anything language models will mean the end of those retarded make-work cookie cutter graded homework assignments that got in the way of actually studying and learning.
You're greatly exaggerating the problem. Literally any requirement can exclude students. Establishment of a trusted proctor network for administering exams is how you solve this problem. If you're an online student, you'll have to show up a few days per year to prove your knowledge in person. I believe this is how many remote study programs already work, because AI is just the latest way to cheat. You could always pay people to do your work for you before AI, and the solution is the same.
Exams were never the pinnacle of what a grad can do. They were an efficient test, under severe time constraints, that correlates well with overall ability in humans.
That AI can pass these tests doesn't mean it is as smart and capable as a grad. I mean, it might be, or if not today then in a few years, but not because it can pass exams, having digested past exams and sample solutions into its bellows.
> The easy solution would be to revert back to paper only exams, and physical attendance - but that would also exclude a ton students.
Which students?
If it's just about travel-distance, maybe schools could organize themselves to offer local test-centers where students could attend exams under observation. Reusing existing facilities in this way is pretty common in my countries education-system since decades.
A renowned South African university, UNISA, has done remote learning for decades. Students had to mail in their work every month, and they would set up exam centres all over the country where students can take their tests.
This is not an unsolvable problem if handwritten work becomes a requirement.
The cheap version of learning is dead, and AI killed it.
Not that we were learning all that much to begin with. I mean, walk into any sorority and ask to see the test bank. The students and Profs were phoning it in for a while, by and large. Not all of them were though, and good on yah.
But now that the fig leaf is torn away, we're left with the Oxbridge model and not much else - small classes, under 10, likely under 5, with a grad level tutor, social pressure making sure you've done the work. The great thing about this though is that you'll have an AI listening in all the time and helping out, streamlining the busywork and allowing the group to get down to business.
But that version is very expensive. You're looking at ~$50k / student year [0] at a baseline Oxbridge model in secondary school on up - ~$400k / student from 9th to university graduation.
Assume a 6% loan rate for 30 years (a mortgage, essentially), and you've got ~$2,300 monthly payments for all your working life, ~$46k/year down the drain. How in the hell are you going to manage student loans like that and then try to live a life without a really good job? How the hell is a nation going to be expected to pay for that per kid if you make school free for them?
Cheap learning wasn't good, but it sufficed. The new models of education must answer to the fundamental question of education: How much does it cost?
[0] 2 hours 3x a week per class; 4 classes per tutor per week. Assume $100k/tutor and 5 students/tutor. So $5k / student / tutor. 4 classes / student. So $20k / student in just raw tutors. At least double that for overhead if not triple.
AI isn't destroying anything. Don't blame the technology for what humans do with it.
AI should allow every student to have personalized instruction and tutoring. It should be a massive win.
If everyone instead of taking advantage of that refuses to do any work and decided to lie and pass the AIs output off as their own, that is not something the AI did. The students did that.
> AI should allow every student to have personalized instruction and tutoring.
I admire your optimism.
Funny how everyone has their own dream of the miracles that “AI” should perform. It's just the perfect silver screen for everyone to project their wishes on.
We turned higher ed into a qualification producing factory subsidized by the government at the expense of the kid's financial future. We overemphasized passing over learning, as the education is about the title, not the knowledge. It's not the student's fault that we created this incentive structure. The students that want to learn can still learn, those that come to higher education with a transactional mindset now can just pay for their degree. The truth is we are at the point where the logic of the commodification of our higher education system is being taken to its logical conclusion which is its own undoing.
Let's throw away the potential of society because young adults are lazy and AI must be empowered. Or, we could realize the realities of human behavior and INTELLIGENTLY integrate AI. But nah, fuck society/fuck young adults for having the typical young adult mentality.
Second this. Sick of seeing posts like this because correlation =/= causation, proved time and time again. It's just too easy to 'relate' these two things and leads to lazy writing / persisting this narrative which has in no way been proven to be true yet.
IMO the underlying cause has much more to do with a hiring cycle issue: the boom of the low-interest / free money / I-don't-need-to-pay-for-an-office covid years is now leading to the relative hiring "bust" (even though it's not really a bust, unemployment is at 4.2%, certainly nothing out of the ordinary for the US)
Even though I majored in CompSci, I still remember my college essay class and learning about primary and secondary sources and their relative quality, how to craft an argument, and how to articulate your argument to be persuasive. Outside of just writing, those skills have been useful in other scenarios too (like when subconsciously evaluating someone else's argument)
Of course, I still treated it like a lazy college student: I did it in 2.1 or 2.2 line spacing to hit the page requirements, and flipped my thesis because it was easier to research (I started out arguing against the US invading Iraq, but found it way easier to find sources that supported an invasion... well, we all know how reliable those sources were).
The leverage has been flipped. We all had awful college classes teaching next to nothing, and now that you can get good grades without attending, what's left? "We lost critical thinking!" No, we were barely getting that in the first place. Now, classes need to be more valuable.
This is exactly it. Are we surprised that civil engineering students forced to take a humanities class satisfied by psych 101 and having to pay thousands of dollars for the 3 credit hours are cheating on their term paper?
It's not surprising, as there are plenty of technical-minded people who believe they should never have to study anything related to "soft sciences", and will do anything to get out of it. But I don't think that people doing so with AI justifies the idea that civil engineering students should not be taught any humanities.
A college-friend from the Unviersity of Oxford, where students write one or two essays a week, got the top first (best mark) in his history degree. Initially impressed, one day I asked him his exam method - where each student must produce 3 essays in 3 hours (or did then) across about 5 or 6 papers. My friend’s approach was to thoroughly research 12 essay questions and pre-write 16 page essays for each paper, which he would then learn verbatim and trot out word-for-word the best fit to each exam question.
This compared to my method of reading widely, learning quotes and ideas and then writing each essay fresh in the exam hall - and I would typically manage about 3-4 pages per essay. (Reader, I did not get a top first).
I relate this anecdote as I don’t really see my friend’s method as being much better than using AI. Although I do acknowledge his 16 page essays must have been reasonably good.
Your friend's approach doesn't sound like cheating, after all the wrote the original essays.
It's more similar to spending hours preparing small exam cheat sheets, and then realizing that you didn't need them during the exam, as you had learnt the material.
It definitely wasn’t cheating. But I felt it was not in the spirit of the exam system which I believed - maybe wrongly - was designed to test one’s ability to write a fresh essay from scratch under timed conditions.
What would you say about someone getting AI to write high standard essays and simply spend learn those word-for-word?
It’s also not cheating but not in the spirit of the thing I think.
> friend’s method as being much better than using AI
Why not? He wrote all the essays himself, after all, and in a setting that's much more relevant to real life vs. the artificial constraints of a shorter exam. With AI he would've written/learned nothing himself.
It’s a fair point, but as a thought experiment, how would you feel about AI writing the essays and simply regurgitating those? Legal but not in the spirit of things I think.
What if education became research? If, in the hypothetical future, the AI can answer any question about any book or scientific theory, perhaps the educational system could focus on teaching people how to come up with good ideas to research, and how to do that research effectively? Rather than making the questions about historical information more difficult, or answering them in person or writing them in bluebooks, make the process of learning about how to create new knowledge? Educators would become people who teach you how to learn, how to design questions, and how to research those questions to produce factual answers. We've known lectures have been the worst way to teach for decades. Why maintain that failed system? If the reductionist goal of the college system is a degree that certifies you as an expert in historical knowledge, maybe we can just throw that away since the AI can handle that part now, and instead certify that people know how to ask the right questions of the AI, and how to interpret their answers to create new knowledge for humanity?
Readit News
The harsh reality is that academia as a whole needs to be revamped. The easy solution would be to revert back to paper only exams, and physical attendance - but that would also exclude a ton students. A huge number of modern students are online students, or similar programs where you don't need to show up physically. Moreover, I don't think universities / colleges themselves want to revert back, as it would mean hiring more people, spending more on buildings, etc.
Honestly, the pervasiveness of LLMs looks to really erode the critical thinking of entire future generations. Whatever the solution, we need to be taking these existential threats a lot more seriously than how we treated social media (the plague before this current plague).
Ask students to solve harder problems, assuming they will use AI to learn more effectively.
Invert the examination process to include teaching others, which you can’t fake. Or rework it to bring the viva voce into evaluation earlier than PhD.
There are plenty of ideas. The problem is, a generation of teachers likely need to be cycled through for this to really work. Much harder for tenured professors.
Every technical revolution “threatened to erode the critical thinking of a generation”, and sure, the printing press meant that fewer texts were memorized rote… not to say there are no risks this time, but rather that it’s hard to predict in advance. I can easily imagine access to personalized tutors making education much better for those who want/need to learn something.
I’m more worried about post-truth civilization than post-college writing civilization for sure.
Yes and no.
Upper middle class parents as a group will still instill critical thinking skills in their kids.
But the above comment reveals more about SES (socioeconomic status) and education in general rather than something specific to critical thinking or LLMs. The current education environment in the US heavily favors kids from higher SES families for a number of reason. LLMs won’t change this.
The challenge for the education system, imho, is to find a way for lower SES kids to thrive in an LLM environment. Pre-LLM, this was already a challenge, but was possible. Post-LLM, the LLM crutch may be too easy for some lower SES folks to lean on such that they don’t develop the skills they need to develop higher order skills.
I suspect this is the true fermi paradox. Once a civilization reaches a certain point, automation becomes harmful to the point that no one knows how do anything on their own. Societal collapse may be back to bronze age, if not more regressed.
Deleted Comment
Exactly the threat of AI. With regards to jobs, we'll have a shock but we will adapt as with any other wave of automation.
I gave it the prompt "Suppose you are in a job interview for a front-end web position and someone asks you about how you use the React library and the hardest problem you even had to solve with it. How might you react, along with a somewhat amusing anecdote?"[1] and it did pretty well. I think I'd play with it a bit to see if I can still suppress some of the LLM-isms that came out, but a human could edit them out in real-time with just a bit of practice too... it's not like you can just read it to your interviewer, you will need to Drama Class 101 this up a bit anyhow. It'll be easier to improv a bit over this than a bare Wikipedia list.
In other words, as with the question the article title asks, the question isn't about what happens "when" this starts being possible... the capability has run ahead of all but the most fervent AI user's understanding and it is already here. It's just a matter of the word-of-mouth getting around as to how to prompt the AIs to be less obvious. I also anticipate that in the next couple of years, the AI companies will be getting tired of people complaining about the "default LLM voice" and it'll shift to be something less obvious than it is now. Both remote interviews and college writing are really already destroyed, the news just hasn't gotten around to everybody yet.
(In fact I suspect that "default LLM voice" will eventually become a sort of cultural touchstone of 2024-2026 and be deliberately used in future cultural references to set stories in this time period. It's a transient quality of current-day LLMs, easy to get them out of even today, and I expect future LLMs to have much different "default voices".)
[1]: And in keeping with my own philosophy of "there's not a lot of value of just pasting in LLM responses" if you want to see what comes out you are welcome to play with it yourself. No huge surprises though. It did the job.
Can you provide an example?
Alternatively, if we still want to cling on to this ritual of measuring the performance of students, you could give each and every one of them oral examinations with AI professors.
Institutions that prepare people for future jobs have an even harder time to justify what they’re doing than the people who are looking for jobs right now. It’s just inertia at this point.
Not to mention that AI can educate the people better by solving Bloom’s Two Sigma Problem.
So colleges are obsolete except as four year cruises for entertainment and networking.
Deleted Comment
1) Written in person exams that were most of the grade (this includes "blue book" exams where you have to sit in front of the professor and write an essay on whatever topic he writes on the board that morning as well as your typical math/algorithms tests on paper.)
2) Written homework where you have to essentially have a satisfactory discussion on the topic (no word range, you get graded on creative interpretation of the course subject matter.)
Language models could maybe help you with 2 but will actually kill your ability to handle 1 if you're cheating on homework with them. If anything language models will mean the end of those retarded make-work cookie cutter graded homework assignments that got in the way of actually studying and learning.
That AI can pass these tests doesn't mean it is as smart and capable as a grad. I mean, it might be, or if not today then in a few years, but not because it can pass exams, having digested past exams and sample solutions into its bellows.
Which students?
If it's just about travel-distance, maybe schools could organize themselves to offer local test-centers where students could attend exams under observation. Reusing existing facilities in this way is pretty common in my countries education-system since decades.
This is not an unsolvable problem if handwritten work becomes a requirement.
Not that we were learning all that much to begin with. I mean, walk into any sorority and ask to see the test bank. The students and Profs were phoning it in for a while, by and large. Not all of them were though, and good on yah.
But now that the fig leaf is torn away, we're left with the Oxbridge model and not much else - small classes, under 10, likely under 5, with a grad level tutor, social pressure making sure you've done the work. The great thing about this though is that you'll have an AI listening in all the time and helping out, streamlining the busywork and allowing the group to get down to business.
But that version is very expensive. You're looking at ~$50k / student year [0] at a baseline Oxbridge model in secondary school on up - ~$400k / student from 9th to university graduation.
Assume a 6% loan rate for 30 years (a mortgage, essentially), and you've got ~$2,300 monthly payments for all your working life, ~$46k/year down the drain. How in the hell are you going to manage student loans like that and then try to live a life without a really good job? How the hell is a nation going to be expected to pay for that per kid if you make school free for them?
Cheap learning wasn't good, but it sufficed. The new models of education must answer to the fundamental question of education: How much does it cost?
[0] 2 hours 3x a week per class; 4 classes per tutor per week. Assume $100k/tutor and 5 students/tutor. So $5k / student / tutor. 4 classes / student. So $20k / student in just raw tutors. At least double that for overhead if not triple.
AI should allow every student to have personalized instruction and tutoring. It should be a massive win.
If everyone instead of taking advantage of that refuses to do any work and decided to lie and pass the AIs output off as their own, that is not something the AI did. The students did that.
I admire your optimism.
Funny how everyone has their own dream of the miracles that “AI” should perform. It's just the perfect silver screen for everyone to project their wishes on.
But wouldn't, so we only have the loss of cheating replacing learning.
IMO the underlying cause has much more to do with a hiring cycle issue: the boom of the low-interest / free money / I-don't-need-to-pay-for-an-office covid years is now leading to the relative hiring "bust" (even though it's not really a bust, unemployment is at 4.2%, certainly nothing out of the ordinary for the US)
Of course, I still treated it like a lazy college student: I did it in 2.1 or 2.2 line spacing to hit the page requirements, and flipped my thesis because it was easier to research (I started out arguing against the US invading Iraq, but found it way easier to find sources that supported an invasion... well, we all know how reliable those sources were).
This compared to my method of reading widely, learning quotes and ideas and then writing each essay fresh in the exam hall - and I would typically manage about 3-4 pages per essay. (Reader, I did not get a top first).
I relate this anecdote as I don’t really see my friend’s method as being much better than using AI. Although I do acknowledge his 16 page essays must have been reasonably good.
It's more similar to spending hours preparing small exam cheat sheets, and then realizing that you didn't need them during the exam, as you had learnt the material.
What would you say about someone getting AI to write high standard essays and simply spend learn those word-for-word?
It’s also not cheating but not in the spirit of the thing I think.
Why not? He wrote all the essays himself, after all, and in a setting that's much more relevant to real life vs. the artificial constraints of a shorter exam. With AI he would've written/learned nothing himself.