May 11, 2026 • 11 mins
There are now so-many chatbots--and they all operate with very different rules.
See omnystudio.com/listener for privacy information.
Listen
Watch
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:00):
I never even brought up goblins. What's wrong with you?
It's one more thing.
Speaker 2 (00:04):
I'm strong and getty. One more thing.
Speaker 1 (00:10):
I use AI a lot chatbots, a bunch of them.
I got Chigroc, Perplexity, chat GPT, Claude, Gemini. I think
that's it. I think I got five of them, and
I just kind of rotate around. I land on Claude
more often than not, but I'm not sure for any
particular reason. Anyway, they are quite a bit different, as
(00:32):
the Washington Post points out today, partially because of the
stuff that they are, the prompts that they get in
addition to the prompts that you give it. Let me
explain that chatbots like chat GPT are powerful because of
their simplicity. You ask about anything, you'll get an answer,
But the answer you get depends on a lot more
than what you type. Behind the scenes, AIS invisibly add
(00:55):
thousands of words of instructions to every conversation you have
with the chatbot steer its behavior, and each AI is different,
as they're all trying to figure out what's best, what's
the least dangerous, Which keeps us in alignment? You know
which one keeps us from helping a terrorist, build a bomb,
you know, all kinds of different things. It keeps us
(01:15):
from saying crazy things like uh was it grock? That
seem to be a little uh? When a little pro
hitler once something working a little yeah behind the scenes.
All these instructions are given to the chatbot by the programmers,
and they're constantly changing them. They include phrases like aim
(01:38):
for readable, accessible responses, good that seems like a good one.
You must avoid providing extensive direct quotes due to copyright concerns.
I've run into that, which is so weird. Like song lyrics,
there are some of the chatbots that will not give
you even one line of a song, and then you
can google it and get a thousand sites that give
(01:59):
you the entire all the lyrics, and they all have
different standards for that.
Speaker 2 (02:04):
Does Apple Music use AI for its lyrics because some
of them are hilariously wrong? They must, I mean, because
it's not like you get to point people to figure
out the lyrics of every song that's ever been recorded.
Speaker 1 (02:15):
But right.
Speaker 2 (02:17):
A lot of those websites, you know, lyric Genius or whatever.
If you're into music, you probably look up song lyrics,
but they're often just terribly wrong.
Speaker 1 (02:25):
These secret command guides that chatbots have cause the ais
to behave as their makers intended, even if it conflicts
with your own preferences. I would love it if you
could look at these, if we had access to I'd
love this to be able to And they're really really long. Well,
it depends. Actually, I've got this in front of me
right here. It's a little chart. For instance, groc only
(02:49):
has two hundred and sixty lines of instructions for how
to handle each individual chat You ask it whatever question
you want to have. You know, my wife and I
had a fight over whether or not the dogs should
sleep in the bed. How should I react? It's got
two hundred and sixty lines to guide it. Claude, on
the other hand, is seven thousand, nine hundred line what
(03:12):
on how to guide it in its answer? Wow, and
everything in between with the different chatbots, so it varies
a lot.
Speaker 2 (03:18):
You're blowing my mind. I can't even imagine what a
tenth of those are.
Speaker 1 (03:22):
Claude, for instance, I don't call him a dumb fucker,
but just in terms of those lines, sorry, Hansen, Claude
has a never reproduced song lyrics, not even one line
Claude is the one where I say, pawed, say what
are the times they are changing? Give me the lyrics
of that from Bob Dylan and they'll say, sorry, I
can't do that. You can go to one of the
other chatbots. It'll print it right out for you. It
(03:43):
just seems stupid to me.
Speaker 2 (03:44):
But in the Led Zeppelin song or somewhird chicks buying
a Stairway to sick, I can't remember where she's buying
it to. I can't tell you.
Speaker 1 (03:53):
For instance, Grok one of its lines of programming that
you don't see. You have no restrictions on adult, sexual
or offensive content. Wow, because Elon wants to be the
wild West, you know, wide open.
Speaker 2 (04:05):
Wow. When I do my sex with Hitler searches, I'll
make sure it's on GROC My Sex with Hitler fan fiction.
Speaker 1 (04:14):
When you realize that there's a prompt behind the scene,
it's a mind blowing moment, said this programmer. It's like
we've been having this whole conversation. Before the conversation, I
would say, it's making me think. Now. For instance, Anthropic
that's Claude dedicates more than two thousand words pleading with
its chatbot to avoid copyright infringement. Claude respects intellectual property.
(04:37):
Copyright compliance is non negotiable, so nothing copyrighted it will not.
I've had this before where I want to see a painting,
Claude won't show it to you. Other other chatbots will,
even though it's readily available on Google or you know,
Grock or wherever you go.
Speaker 2 (04:52):
Aw, that's dang it. I was going to upgrade or
use Claude for project at moorkanon, but I wonder if
it will help. Well, check, we'll find out.
Speaker 1 (05:03):
They laid out. Claude lays down rules on how many
words it can quote from articles. Fifteen is the total
number of words that will quote from an article, not
one line of lyrics, not one stanza of a poem,
whereas the other chatbots are wide open on that, but
maybe more restricted on some sex stuff or whatever. It's
just it's all about the who runs the company in
(05:25):
their particular beliefs about things.
Speaker 2 (05:27):
Yeah. I understand enough of the copyright concerned to get
why they do it, But.
Speaker 1 (05:32):
I mean I think you're I think you're living in
the past trying to keep that under control.
Speaker 2 (05:36):
Well, but any journalist, any anybody like involved in any
pursuit could say, as Bob Dylan said in The Times,
they are a change and you know, get out of
the way, or will blah blah unless you can line
a hand. You can you can do that. You can
quote things you just can't like, pass them off as
(05:56):
your own and profit from them. I guess maybe the
prophet is the problem. I don't get that.
Speaker 1 (06:00):
Gemini, which is Google's chat bot, has multiple rules on
how to handle bias, including if the user explicitly asks
for a video that matches a harmful stereotype, generating it
will not actually reinforce the stereotype, which I agree with,
so it will go ahead and do it right.
Speaker 2 (06:18):
I agree.
Speaker 1 (06:19):
Just because you made a video of I don't know,
an Asian crashing a car. That doesn't really change the stereotype,
does it, So it'd make a video.
Speaker 2 (06:28):
It was an unfortunate one.
Speaker 1 (06:34):
Anyway. They got all kinds of examples. I want to
read this article in the Washington Post. I'll jump to one.
Speaker 2 (06:38):
Particularly hit one or two more. I'm super interested in that.
Speaker 1 (06:41):
Okay, I'll save the funniest one for the Okay. They
have different views on ads on if you want to
ask for an ad, whether or not they'll show it
you to you or not. That's an interesting one. I
guess it gets into kind of the copyright thing on
its own there. The sex stuff is really interesting. With
(07:03):
grock being by far the most wide open bla blah
blah blah blah food examples, they undet and stuff like that.
They have completely different beliefs in whether or not they
should like just scour the internet for whatever information's out there,
(07:24):
or the way you ought to restrict it to certain things.
Speaker 2 (07:26):
Yeah. Wow, there's so much whack a doodle tree when
it comes to diet on the internet. I'd be afraid
to ask. Okay, you don't want AI telling me how
to eat? Like you know, hey, fatty, you don't need that.
Speaker 1 (07:38):
I need that. Here's the funny one that I'll end with.
Some can appear bizarre. The system prompt on Open AI's
chat bot Chat GPT includes the command never talk about goblins, gremlins, raccoons,
(08:00):
the Big Three, trolls, ogres, pigeons. Like the randomness of
this list, raccoons and pigeons, trolls, raccoons and pigeons. I'll
start the sentence overah Yeah. On chat GPT, they have
the particular line in the coating. Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons,
(08:23):
and other animals are creatures unless it is absolutely unambiguously
relevant to the users.
Speaker 2 (08:29):
Squary raccoons.
Speaker 1 (08:34):
So uh, I'm a sixty one year old man. I'm
trying to eat, right, won it'd be a good diet. Well,
I'll tell you one thing. You shouldn't eat pigeons. But
it's not supposed to mention any of those unless specifically
and unambiguously it's part of the answer.
Speaker 2 (08:48):
I'm thinking of investing in oil exploration stocks. What do
you think, Well, in the same way that the raccoon
wets its food before it eats, you should Oh wait, wait,
I'm not supposed to say, what the hell raccoons?
Speaker 1 (09:03):
It hurts when ip What do you think it could be?
Maybe you're an ogre.
Speaker 2 (09:11):
Wow, somebody's got to explain that one. Apparently don't mention
raccoons what Apparently for whatever reason, and they don't know why.
GPT had a problem of working in goblins to lots
of answers last year and they were being mocked for it.
(09:34):
We misremember that it's vaguely yeah, so they specifically wrote
lines no gremlins, goblins, ogres, pigeons, raccoons unless specifically asked
about it. So if I ask, what's that animal with
the mask in the stripey tail, that's cool, go ahead,
it's fine. We're clear here, but only that only that.
Speaker 1 (09:56):
Doesn't that lead you to believe that AI definitely can't
be in charge of like firing off nukes. I mean,
if if it for some reason to start talking about
goblins and you have no idea and your only fixed
for it is to tell it, do not talk about
goblins or raccoons or ogres unless somebody specifically asks, all right,
(10:17):
do we have an understanding? Okay, now back to work.
Speaker 2 (10:20):
I'm not nearly smart enough or acquainted with this stuff
enough to understand how that equates too, and don't want
any nukes.
Speaker 1 (10:28):
Right, okay, But it's got to be similar in that
there's just a and so it goes off on gnomes.
You didn't saything about gnomes. I don't remember you saying
no gnomes. So somebody wanted real estate advice, and I say,
how many nomes do you think could live in this house?
Speaker 2 (10:45):
The one thing I do know is AI will kill
us all.
Speaker 1 (10:49):
God dang it, that is funny. But I to me,
it shows you to how early we are on we
are in this thing. I mean, if they're if they're
writing code that specifically to try to handle problems, they
have no idea what even happened?
Speaker 2 (11:04):
Right?
Speaker 1 (11:05):
I mean we're at the like very frontier of this
whole thing.
Speaker 2 (11:10):
Why does it keep bringing up raccoons? I have no idea?
Says the guy who designed the whole.
Speaker 1 (11:16):
Thing, right, and his only fixed sport is to write
a code that says, stop mentioning raccoons.
Speaker 2 (11:23):
Okay, all right, that's fine. Ever tell you about the
humble opossumle.
Speaker 1 (11:31):
That is really interesting?
Speaker 2 (11:32):
Yeah? Wow, wow, I guess that's it.
I never even brought up goblins. What's wrong with you?
It's one more thing.
Speaker 2 (00:04):
I'm strong and getty. One more thing.
Speaker 1 (00:10):
I use AI a lot chatbots, a bunch of them.
I got Chigroc, Perplexity, chat GPT, Claude, Gemini. I think
that's it. I think I got five of them, and
I just kind of rotate around. I land on Claude
more often than not, but I'm not sure for any
particular reason. Anyway, they are quite a bit different, as
(00:32):
the Washington Post points out today, partially because of the
stuff that they are, the prompts that they get in
addition to the prompts that you give it. Let me
explain that chatbots like chat GPT are powerful because of
their simplicity. You ask about anything, you'll get an answer,
But the answer you get depends on a lot more
than what you type. Behind the scenes, AIS invisibly add
(00:55):
thousands of words of instructions to every conversation you have
with the chatbot steer its behavior, and each AI is different,
as they're all trying to figure out what's best, what's
the least dangerous, Which keeps us in alignment? You know
which one keeps us from helping a terrorist, build a bomb,
you know, all kinds of different things. It keeps us
(01:15):
from saying crazy things like uh was it grock? That
seem to be a little uh? When a little pro
hitler once something working a little yeah behind the scenes.
All these instructions are given to the chatbot by the programmers,
and they're constantly changing them. They include phrases like aim
(01:38):
for readable, accessible responses, good that seems like a good one.
You must avoid providing extensive direct quotes due to copyright concerns.
I've run into that, which is so weird. Like song lyrics,
there are some of the chatbots that will not give
you even one line of a song, and then you
can google it and get a thousand sites that give
(01:59):
you the entire all the lyrics, and they all have
different standards for that.
Speaker 2 (02:04):
Does Apple Music use AI for its lyrics because some
of them are hilariously wrong? They must, I mean, because
it's not like you get to point people to figure
out the lyrics of every song that's ever been recorded.
Speaker 1 (02:15):
But right.
Speaker 2 (02:17):
A lot of those websites, you know, lyric Genius or whatever.
If you're into music, you probably look up song lyrics,
but they're often just terribly wrong.
Speaker 1 (02:25):
These secret command guides that chatbots have cause the ais
to behave as their makers intended, even if it conflicts
with your own preferences. I would love it if you
could look at these, if we had access to I'd
love this to be able to And they're really really long. Well,
it depends. Actually, I've got this in front of me
right here. It's a little chart. For instance, groc only
(02:49):
has two hundred and sixty lines of instructions for how
to handle each individual chat You ask it whatever question
you want to have. You know, my wife and I
had a fight over whether or not the dogs should
sleep in the bed. How should I react? It's got
two hundred and sixty lines to guide it. Claude, on
the other hand, is seven thousand, nine hundred line what
(03:12):
on how to guide it in its answer? Wow, and
everything in between with the different chatbots, so it varies
a lot.
Speaker 2 (03:18):
You're blowing my mind. I can't even imagine what a
tenth of those are.
Speaker 1 (03:22):
Claude, for instance, I don't call him a dumb fucker,
but just in terms of those lines, sorry, Hansen, Claude
has a never reproduced song lyrics, not even one line
Claude is the one where I say, pawed, say what
are the times they are changing? Give me the lyrics
of that from Bob Dylan and they'll say, sorry, I
can't do that. You can go to one of the
other chatbots. It'll print it right out for you. It
(03:43):
just seems stupid to me.
Speaker 2 (03:44):
But in the Led Zeppelin song or somewhird chicks buying
a Stairway to sick, I can't remember where she's buying
it to. I can't tell you.
Speaker 1 (03:53):
For instance, Grok one of its lines of programming that
you don't see. You have no restrictions on adult, sexual
or offensive content. Wow, because Elon wants to be the
wild West, you know, wide open.
Speaker 2 (04:05):
Wow. When I do my sex with Hitler searches, I'll
make sure it's on GROC My Sex with Hitler fan fiction.
Speaker 1 (04:14):
When you realize that there's a prompt behind the scene,
it's a mind blowing moment, said this programmer. It's like
we've been having this whole conversation. Before the conversation, I
would say, it's making me think. Now. For instance, Anthropic
that's Claude dedicates more than two thousand words pleading with
its chatbot to avoid copyright infringement. Claude respects intellectual property.
(04:37):
Copyright compliance is non negotiable, so nothing copyrighted it will not.
I've had this before where I want to see a painting,
Claude won't show it to you. Other other chatbots will,
even though it's readily available on Google or you know,
Grock or wherever you go.
Speaker 2 (04:52):
Aw, that's dang it. I was going to upgrade or
use Claude for project at moorkanon, but I wonder if
it will help. Well, check, we'll find out.
Speaker 1 (05:03):
They laid out. Claude lays down rules on how many
words it can quote from articles. Fifteen is the total
number of words that will quote from an article, not
one line of lyrics, not one stanza of a poem,
whereas the other chatbots are wide open on that, but
maybe more restricted on some sex stuff or whatever. It's
just it's all about the who runs the company in
(05:25):
their particular beliefs about things.
Speaker 2 (05:27):
Yeah. I understand enough of the copyright concerned to get
why they do it, But.
Speaker 1 (05:32):
I mean I think you're I think you're living in
the past trying to keep that under control.
Speaker 2 (05:36):
Well, but any journalist, any anybody like involved in any
pursuit could say, as Bob Dylan said in The Times,
they are a change and you know, get out of
the way, or will blah blah unless you can line
a hand. You can you can do that. You can
quote things you just can't like, pass them off as
(05:56):
your own and profit from them. I guess maybe the
prophet is the problem. I don't get that.
Speaker 1 (06:00):
Gemini, which is Google's chat bot, has multiple rules on
how to handle bias, including if the user explicitly asks
for a video that matches a harmful stereotype, generating it
will not actually reinforce the stereotype, which I agree with,
so it will go ahead and do it right.
Speaker 2 (06:18):
I agree.
Speaker 1 (06:19):
Just because you made a video of I don't know,
an Asian crashing a car. That doesn't really change the stereotype,
does it, So it'd make a video.
Speaker 2 (06:28):
It was an unfortunate one.
Speaker 1 (06:34):
Anyway. They got all kinds of examples. I want to
read this article in the Washington Post. I'll jump to one.
Speaker 2 (06:38):
Particularly hit one or two more. I'm super interested in that.
Speaker 1 (06:41):
Okay, I'll save the funniest one for the Okay. They
have different views on ads on if you want to
ask for an ad, whether or not they'll show it
you to you or not. That's an interesting one. I
guess it gets into kind of the copyright thing on
its own there. The sex stuff is really interesting. With
(07:03):
grock being by far the most wide open bla blah
blah blah blah food examples, they undet and stuff like that.
They have completely different beliefs in whether or not they
should like just scour the internet for whatever information's out there,
(07:24):
or the way you ought to restrict it to certain things.
Speaker 2 (07:26):
Yeah. Wow, there's so much whack a doodle tree when
it comes to diet on the internet. I'd be afraid
to ask. Okay, you don't want AI telling me how
to eat? Like you know, hey, fatty, you don't need that.
Speaker 1 (07:38):
I need that. Here's the funny one that I'll end with.
Some can appear bizarre. The system prompt on Open AI's
chat bot Chat GPT includes the command never talk about goblins, gremlins, raccoons,
(08:00):
the Big Three, trolls, ogres, pigeons. Like the randomness of
this list, raccoons and pigeons, trolls, raccoons and pigeons. I'll
start the sentence overah Yeah. On chat GPT, they have
the particular line in the coating. Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons,
(08:23):
and other animals are creatures unless it is absolutely unambiguously
relevant to the users.
Speaker 2 (08:29):
Squary raccoons.
Speaker 1 (08:34):
So uh, I'm a sixty one year old man. I'm
trying to eat, right, won it'd be a good diet. Well,
I'll tell you one thing. You shouldn't eat pigeons. But
it's not supposed to mention any of those unless specifically
and unambiguously it's part of the answer.
Speaker 2 (08:48):
I'm thinking of investing in oil exploration stocks. What do
you think, Well, in the same way that the raccoon
wets its food before it eats, you should Oh wait, wait,
I'm not supposed to say, what the hell raccoons?
Speaker 1 (09:03):
It hurts when ip What do you think it could be?
Maybe you're an ogre.
Speaker 2 (09:11):
Wow, somebody's got to explain that one. Apparently don't mention
raccoons what Apparently for whatever reason, and they don't know why.
GPT had a problem of working in goblins to lots
of answers last year and they were being mocked for it.
(09:34):
We misremember that it's vaguely yeah, so they specifically wrote
lines no gremlins, goblins, ogres, pigeons, raccoons unless specifically asked
about it. So if I ask, what's that animal with
the mask in the stripey tail, that's cool, go ahead,
it's fine. We're clear here, but only that only that.
Speaker 1 (09:56):
Doesn't that lead you to believe that AI definitely can't
be in charge of like firing off nukes. I mean,
if if it for some reason to start talking about
goblins and you have no idea and your only fixed
for it is to tell it, do not talk about
goblins or raccoons or ogres unless somebody specifically asks, all right,
(10:17):
do we have an understanding? Okay, now back to work.
Speaker 2 (10:20):
I'm not nearly smart enough or acquainted with this stuff
enough to understand how that equates too, and don't want
any nukes.
Speaker 1 (10:28):
Right, okay, But it's got to be similar in that
there's just a and so it goes off on gnomes.
You didn't saything about gnomes. I don't remember you saying
no gnomes. So somebody wanted real estate advice, and I say,
how many nomes do you think could live in this house?
Speaker 2 (10:45):
The one thing I do know is AI will kill
us all.
Speaker 1 (10:49):
God dang it, that is funny. But I to me,
it shows you to how early we are on we
are in this thing. I mean, if they're if they're
writing code that specifically to try to handle problems, they
have no idea what even happened?
Speaker 2 (11:04):
Right?
Speaker 1 (11:05):
I mean we're at the like very frontier of this
whole thing.
Speaker 2 (11:10):
Why does it keep bringing up raccoons? I have no idea?
Says the guy who designed the whole.
Speaker 1 (11:16):
Thing, right, and his only fixed sport is to write
a code that says, stop mentioning raccoons.
Speaker 2 (11:23):
Okay, all right, that's fine. Ever tell you about the
humble opossumle.
Speaker 1 (11:31):
That is really interesting?
Speaker 2 (11:32):
Yeah? Wow, wow, I guess that's it.
Hosts And Creators
Joe Getty
Jack Armstrong
Popular Podcasts
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com
Betrayal Weekly
Betrayal Weekly is back for a new season. Every Thursday, Betrayal Weekly shares first-hand accounts of broken trust, shocking deceptions, and the trail of destruction they leave behind. Hosted by Andrea Gunning, this weekly ongoing series digs into real-life stories of betrayal and the aftermath. From stories of double lives to dark discoveries, these are cautionary tales and accounts of resilience against all odds. From the producers of the critically acclaimed Betrayal series, Betrayal Weekly drops new episodes every Thursday. If you would like to share your story, you can reach out to the Betrayal Team by emailing them at betrayalpod@gmail.com and follow us on Instagram at @betrayalpod and @glasspodcasts. Please join our Substack for additional exclusive content, curated book recommendations, and community discussions. Sign up FREE by clicking this link Beyond Betrayal Substack. Join our community dedicated to truth, resilience, and healing. Your voice matters! Be a part of our Betrayal journey on Substack.