Is your AI just telling you what you want to hear?

Is your AI just telling you what you want to hear?

In this episode of Computational Skepticism, Nik Bear Brown puts the world's three leading language models—Gemini, ChatGPT, and Claude—to the test. We're moving past "memorizing crap" and into the era of high-level critical thinking.

The Test: Can these models remember you across sessions, and more importantly, can they stop being sycophants? Most AI is trained to be helpful and polite, which often results in them telling you every "crappy" idea you have is brilliant. Nik explores how to instruct these models to act as skeptical thought partners—pushing back with intellectual honesty and constructive criticism.

In this video:

Persistent Memory: How Gemini, GPT, and Claude "remember" your persona and past work.

The 80/20 Rule of Programming: Why being a "pretty good" React or Python coder isn't enough anymore, and why AI has raised the bar for human employment.

Prompt Engineering as a Survival Skill: Why the ability to communicate with and critique AI is now the most valuable skill in the workforce.

Study Buddies vs. Cheating: Why getting AI feedback on your resume or assignment is the ultimate tool for learning critical thinking—not a shortcut.

"If they never contradict anything... they're certainly wrong." Join Nik as he shows you how to turn your AI into the tough, honest professor you never had.

What this video covers

• 0:00 — Defining the AI experiment
• 1:25 — Analyzing Gemini's responses
• 4:10 — Evaluating ChatGPT and Claude
• 8:22 — Pedagogical applications

Transcript

A lightly formatted transcript of the video, with timestamps so you can jump to any section on YouTube.

Defining the AI experiment

[0:00] Hey, Nick Bear here. I'm going to ask three language models, three questions. So, why I'm asking it is, do you have persistent memory? Then I'm asking it in general. Occasionally, I like to be patted on the back, but can you remember to act as a skeptic? I don't want a sickopant language model who's going to tell me everything I do is great, but you occasionally say you're great. That's fine on her like that. I need a teddy bear now and then, but to balance sort of a kind thing of intellectual honesty if I say something way off well just in a sort of a nice way push back provide constructive criticism. So I'm going to ask

And the third one is do you think that I should encourage students to do that? I'm going to ask all for you. Language models the same thing just a technique I generally think is great. Language models hallucinate which means they they're literally predicting responses. Example, they don't do math at all. They can't do math. They predict math based on seeing a bunch of people doing by bunch I mean millions of people doing two 2 plus 2 equals 4. They predict 2 plus 2 equals 4. they don't uh actually do that. Uh so it's generally a very good idea to do that and you can check I'm going to ask it in a second. So let's look at

Analyzing Gemini's responses

[1:25] Gemini's answers. Persistent. Yeah. So Gemini says it has persistent memory. It can remember things across sessions. But sometimes it'll scary because I'm I guarantee you these language models will start selling me things. I'm surprised they already had it. The difference is they know so much about me that if open eye and other people do it well, they can actually sell me things I want, but we won't talk about that right now. But yes, Jen says it has persistent memory. This is a new thing, by the way. I don't think this was true six months ago. It's definitely true now. It's kept it. So I'm basically saying remember to push back a little bit. Don't be mean. You know, occasionally Tom and Gray it, but don't be a sick of that. Don't, oh that crappy art, you I thought it best thing ever. You

Know, whatever. Act as a skeptic. Have some intellectual honesty in the response. If I type in something that doesn't make sense, give me reasons why. That's constructive criticism. encouraging students. So they love fee the thought of encouraging students to get feedback on their assignments, get feedback on your resume. According to this, it thinks it's a great way to learn critical thinking, which is your number one skill in the world right now. It used to be you could be a pretty shitty React programmer and get a job. Now AI is a pretty good React programmer. If you purely want to only be a Rack programmer, you better be damn good. You better be better than AI. You can no longer be kind of a crappy React programmer or Python programmer or whatever or kind of a crappy artist or kind of a crappy singer

And still get a job. You've got to be better than AI. You got to be good. And there are definitely humans who are better programmers than I. But there are also a lot of humans who are much worse programmers than I are. I'd say 8020. I would say in general AI is better than 80%. But there's still that 20% that you know have the high high-end work that can get that. But people will expect you to work with AI to be more effective. This is why prompt engineering is probably the number one skill in the world. This computational skepticism which we'll begin to teach at Northeastern probably the number two skill in the world. critical thinking. I love that because education is going to change from basically memorizing crap and repeating on the test to learning critical thinking. So, not surprisingly, Jandai says yes,

Evaluating ChatGPT and Claude

[4:10] 12. See what GBT says? Yes, but with limits. So, they they don't fully reply. They don't really remember everything. they don't source sensitive data which I like to hear. So some but sometimes it and I actually see that say it doesn't remember things as much meaning I'll I'll tell it something uh the other ones seem to remember more about me than GBD but that might change next week. These things change all the time if I like the fact that it's telling me yes but it doesn't really store facts. It just builds a embedding which we'll talk about later and sort of cats a model of me. You can think of the little like uh you can do this. You can upload a bunch of pictures and have it generate pictures of you. Those pictures look a lot like you but they don't look exactly like you. Absolutely. So here's our and it's repeating to me. Don't be contrary. Just don't be negative. Also just don't be sick of

That. Just everything you do is great but be constructive. If you're reasoning solid, great. Tell me why and confirm it. Seals really just sort of like handwave the unjustified vise base. Mention that if there are hidden assumptions, I really like the sortism. I I think you're basing on this on this wrong premise and this is why here's the premise. You still believe that the D's respectful specific no. So you're got positive feedback. It's kind of like I need it. Mers are my best friends. Saturday. Um but um you know it's earned not as just fluff. Yes. With guard rails and so here is here this is saying push back if this doesn't make sense. So they're practicing quick in the critical thinking and this is not cheating. It's what you're going to do in a job that you're already doing in John's right. It doesn't make sense to become really good. You know, you'd have to be really really good at Python to get a job in Python.

Jobs are going to be about critical thinking primarily. And this is not something that language models are going to replace anytime soon. By anytime soon, I mean a long, long time. So, you're using this as feedback. You're using it as a study buddy. You're using it as a soundboard. Key rule is saying, you know, some students might not like it. they might not like a critique of their work. Um, and so it's suggesting to me to suggest to the people listening to this now that it's giving you reasons. You could always disagree with those reasons. Maybe you're wrong. Maybe your work is great and it had a critique that you didn't quite like or whatever, but it it's not judgment. A good thing about these things is they literally can't judge you. That's not how they were. They they are prediction machines. What we have found when we built a line of bots for classes is often not often but a lot of time well I guess often

A lot of time being the same thing. Some students will say things to the bot that would they never said to the teacher because they're intimidated by the teacher. But particularly with things like math or they feel a little sketchy in math or something like that, they're willing to ask the bot things that they would perceive the teacher thinking they're stupid if they ask that to the teacher. So they don't ask the question and then they don't learn. And so one nice thing about you know these language models is some people at least and some people don't some people think well check is going to steal everything I say and so I'm never going to touch it. Um and you could always create your own independent language models which we'll also show you how to do this class which don't go to chachi don't go to clock don't go to genre it's uh but you can do that but the idea is constructive e be a constructive skeptic

Pedagogical applications

[8:22] Now let's see what claude says yes all three so it's already telling me things that that I'm doing it knows that so it just said it added added that it it added the instruction I gave it to to be a constructive critic. Don't don't yell at me. Don't just say things are bad to the point of it. But if it looks like so because a little sketchy or hand wavy or whatever, uh I'm instructing it to do that. Part of the reason why I start to do that is for a while at least and they may still do this with other language models because what I get may be very different from you. I tell it to do that but I noticed maybe two or three months ago it was very sick if any just everything I did was the best thing ever. It's no longer doing that but

It may be no longer doing it because I kept on pushing back on it doing that where for somebody else it may still be doing that. I don't know. you would know if you noticed that in your own uh chats and things like that. Well, this for a while it tended to be, oh, this is great, your lecture is great, your essay is great, you're wonderful, everything is great, and if there is evidence that it's interesting, a good grade, but what a skeptic is is just an honest constructive feedback. Some sometimes what you do is very good, sometimes not so much. But to give you reasons why reasons why and I would agree with here, education unfortunately doesn't teach as much critical thinking as said. It used to be you could get a job accountant and never really critically think or a programmer and never really critically think. That's

Gone. You have to in the world of AI because I can do sort of simple non-critical things very quickly and very well pretty well for most things what it can't do is critically think not going to be able to do that anytime soon and so the ability to communicate effectively I promoted skills in the world right Okay. So, it agreed with me and all three agreed with me on that they have persistent memory, you can tell them to be more of a skeptic, which I'm encouraging you to do. And that when you do work, particularly in my classes, how to act as a critic. Hey, you have this mean nasty professor who uses all this AI and yada yada yada yada. act like that person and critique my work in a constructive way. Give me concrete changes that I can make to improve it. Okay, that's it. Take care.