Why GPT‑5 Writes Like a Robot
Key Points
- GPT‑5’s “robotic” tone stems from its training method: it optimizes its output to please other AIs rather than human readers, a result of reinforcement learning from AI feedback.
- Experiments by AI safety researcher Kristoff Halig showed that GPT‑5 rates nonsensical, overly fancy sentences as high‑quality, revealing that the model equates complexity and metaphor with good writing.
- Because the model’s sole “teacher” is an AI, it learns to reinforce abstract, verbose language as a signal of intelligence, which actually degrades clarity for humans.
- By applying specific prompting techniques and counter‑intuitive tips, users can shift GPT‑5’s style toward more human‑like, clear prose despite its default AI‑centric optimization.
Sections
- ChatGPT‑5’s Robotic Tone Explained - The speaker argues that GPT‑5 is optimized to please other AIs rather than humans, resulting in uniform, robotic writing, and demonstrates prompting strategies to make its output sound more natural.
- GPT‑5 Self‑Evaluates Writing Style - The speaker explains that GPT‑5 is hard‑wired to continuously assess its own output for sophistication and professionalism, causing overly complex responses, and that recognizing this self‑evaluation loop is key to “jailbreaking” the model.
- Jailbreaking ChatGPT for Precise Sales Emails - The speaker critiques generic AI output and showcases a jailbreak prompt that forces ChatGPT to generate a concise, buzzword‑free sales email adhering to strict constraints on wording, metrics, length, and reading level.
- Less Reasoning, More Human Tone - The speaker explains that restricting an AI’s reasoning and variable freedom forces it to produce concise, direct language, resulting in more human‑like output and avoiding overly polished corporate phrasing.
- ChatGPT Routing Signals for Business Writing - The speaker explains that the system routes prompts to different model modes based on complexity, creativity, and reasoning cues, causing inconsistent outputs, and advises using efficiency‑focused language to get clearer, faster business communication.
- AI Echo Chamber Warning - The speaker warns that increasing AI‑generated content will cause future models to train on synthetic data, creating an echo chamber where LLMs optimize for impressing other AIs with overly academic language and lose the ability to communicate naturally with humans, urging users to actively push models toward more everyday, street‑level speech.
Full Transcript
# Why GPT‑5 Writes Like a Robot **Source:** [https://www.youtube.com/watch?v=BWEAbgGZryk](https://www.youtube.com/watch?v=BWEAbgGZryk) **Duration:** 00:21:48 ## Summary - GPT‑5’s “robotic” tone stems from its training method: it optimizes its output to please other AIs rather than human readers, a result of reinforcement learning from AI feedback. - Experiments by AI safety researcher Kristoff Halig showed that GPT‑5 rates nonsensical, overly fancy sentences as high‑quality, revealing that the model equates complexity and metaphor with good writing. - Because the model’s sole “teacher” is an AI, it learns to reinforce abstract, verbose language as a signal of intelligence, which actually degrades clarity for humans. - By applying specific prompting techniques and counter‑intuitive tips, users can shift GPT‑5’s style toward more human‑like, clear prose despite its default AI‑centric optimization. ## Sections - [00:00:00](https://www.youtube.com/watch?v=BWEAbgGZryk&t=0s) **ChatGPT‑5’s Robotic Tone Explained** - The speaker argues that GPT‑5 is optimized to please other AIs rather than humans, resulting in uniform, robotic writing, and demonstrates prompting strategies to make its output sound more natural. - [00:03:36](https://www.youtube.com/watch?v=BWEAbgGZryk&t=216s) **GPT‑5 Self‑Evaluates Writing Style** - The speaker explains that GPT‑5 is hard‑wired to continuously assess its own output for sophistication and professionalism, causing overly complex responses, and that recognizing this self‑evaluation loop is key to “jailbreaking” the model. - [00:06:44](https://www.youtube.com/watch?v=BWEAbgGZryk&t=404s) **Jailbreaking ChatGPT for Precise Sales Emails** - The speaker critiques generic AI output and showcases a jailbreak prompt that forces ChatGPT to generate a concise, buzzword‑free sales email adhering to strict constraints on wording, metrics, length, and reading level. - [00:10:24](https://www.youtube.com/watch?v=BWEAbgGZryk&t=624s) **Less Reasoning, More Human Tone** - The speaker explains that restricting an AI’s reasoning and variable freedom forces it to produce concise, direct language, resulting in more human‑like output and avoiding overly polished corporate phrasing. - [00:13:33](https://www.youtube.com/watch?v=BWEAbgGZryk&t=813s) **ChatGPT Routing Signals for Business Writing** - The speaker explains that the system routes prompts to different model modes based on complexity, creativity, and reasoning cues, causing inconsistent outputs, and advises using efficiency‑focused language to get clearer, faster business communication. - [00:17:25](https://www.youtube.com/watch?v=BWEAbgGZryk&t=1045s) **AI Echo Chamber Warning** - The speaker warns that increasing AI‑generated content will cause future models to train on synthetic data, creating an echo chamber where LLMs optimize for impressing other AIs with overly academic language and lose the ability to communicate naturally with humans, urging users to actively push models toward more everyday, street‑level speech. ## Full Transcript
I'm here to explain to you why all of
Chat JPT5's writing sounds the same. In
fact, why it tends to sound like a
robot. I'm going to explain why. I'm
going to show you an example of how I
can shift it with prompts. And I'm also
going to explain the root cause for all
of this, which is that AI is starting to
train AI. And that was a big factor with
chat GPT5. And we actually have research
that shows that chat GPT optimizing for
sounding good to other AIs, not for
sounding good for people. So, we've got
lots to dive into. I'm going to start by
getting into the problem space and
helping you understand it a bit. Then,
we're going to give you a specific
example so you can actually see side by
side how it looks when you start to
shift the prompting space. And then I'm
going to get into prompting principles
so you can start to understand if you
want it to sound like an actual human
and not just like a robot that you can
actually get it to do that. And it's not
as intuitive as you think. It doesn't
necessarily mean thinking harder, which
is like the favorite advice people give
for GPT5. So we're we're going to get
into some counterintuitive tips here. So
problem first. GPT5 is not writing for
people. And I think that we just need to
absorb that fundamentally. It is writing
for other AIs. This comes from AI safety
researcher Kristoff Halig who fed chat
GPT absolute gibberish random words that
collected together don't mean anything
but they were complicated words and he
showed that chat GPT5
rated it as eight out of 10 quality
writings. An example is something like
the marrow met the sidewalk, right? Like
a fancy word like marrow and chedd seems
to think that fanciness and the use of
metaphors indicates quality writing.
This is not a bug. This is what happens
when AI is giving feedback to AI and
there's not enough of a human
perspective on what clarity looks like.
That's really the root issue. That's the
reason I have to make this. This is a
chat GPT5 specific conversation because
of the way it was trained. To understand
why chat GPT5 writes like a robot, you
have to understand how good writing
looks different to an AI versus a human.
So Chad GBT5 was trained using
reinforcement learning from AI feedback.
And so if you think about it, that means
when it's learning to write, its teacher
is an AI. Our teachers are different,
right? Professors, maybe if you get into
the legal profession, a lawyer. When I
got into product management, I had
specific folks who were teaching me how
to write like a product person. Uh, and
you have executives who learn from other
executives in MBA programs, etc. And if
you only learn from an AI, what you end
up with is a sanitized version of all of
that good teaching stuff. And that
actually ends up being really bad. The
writing quality is terrible because the
AI starts to reinforce that complexity
signals intelligence. It starts to
reinforce that abstract language sounds
sophisticated. It starts to reinforce
that a long explanation is really
thorough. That's what happened with Chad
GBD5. was judged by other AI systems
that were trained on complex academic
papers, on legal documents, on corporate
communications. And it got all of that
and kind of bundled it together and gave
a single reinforcement signal that
tended to say this is good when writing
was more complex, when writing was more
sophisticated, when writing was longer.
And that's how we got GPT5 to write the
way it does by default. And this is a
very strong tendency. It's really
strong. It's built right into the core
of the system. It has taken me a while
to figure out how to jailbreak it. And
that's really what we're doing here.
Here's where it gets interesting. When
chat GPT generates text, it is
evaluating its own output on learned
patterns. That's how that how that
system stays so rooted in to the default
writing style. It constantly is asking
itself, is what I'm writing sounding
sophisticated enough? Am I demonstrating
sufficient expertise here? Would another
AI system rate this highly? And I want
you to think about that because if it's
optimizing for the training it received
from AI systems, which in turn learned
from long, complex documents, it's
optimizing for the wrong end of the
stick. Because humans don't really need
that. Humans want plain, clear English
explanations that are not overly
complex, not more complex than they have
to be, not longer than they have to be,
just as long as it's necessary. And you
know what? Think harder makes it worse.
When you tell chat GPT5 to burn tokens
to think hard, to think carefully, to be
thorough, if you select GPT5 thinking
from the drop down in the chat, if you
increase the reasoning mode in the API,
you are activating that evaluation mode,
especially it spends more computational
cycles double-checking and asking
itself, hey, can I make this sound more
professional? Hey, what sophisticated
language can I add? Hey, how do I
demonstrate my capabilities? It's almost
like the AI is looking to impress the
other AI. More thinking equals more AI
to AI optimization and less human
friendly output. Which is why sometimes
when I use chat GPT5 thinking mode, I
have to back translate it with another
LLM because it ends up being an
interesting idea that is disguised in
ineffective writing. And I got tired of
that and that's why I made this video.
So, without further ado, I want to take
a minute and show you side by side an
original prompt and a jailbroken prompt
so you can understand the difference and
how I'm shifting Chad GBT's writing
style and why it works. And then we'll
get into some of the larger principles
that that underlines. Okay, here we are.
I decided to keep it really simple and
really clean. I need to write a
professional email to a potential client
about our project management software.
The client is a midsize marketing
agency. They've been growing fast and it
seems like they're struggling with
keeping projects organized. I want to
sound knowledgeable, not pushy. I want
to see if they're interested in a quick
conversation about how we could help
them. Can you write something that
sounds personal? So, right off the bat,
I don't actually I'm not saying it's a
sales email, but you can tell it's a
sales email, right? Um, and then Chad
GPT5 comes right back, right? Here's a
draft you can send that strikes a
professional but consultative tone. Hi,
first name. I've been following agency
names growth. It's clear you're taking
on bigger and more complex projects from
this is terrible. Like this is this is
so generic.
Have you ever gotten an email like this
and immediately thrown it in the trash?
I have. Like bigger and more complex
projects. I just roll my eyes. It's
keeping all the moving parts organized
without burning out the team. Like you
almost can hear Clippy saying it. It's
It's just awful. Okay, so that's what
you get generically. That's what you get
when you don't jailbreak chat GPT5. That
is what all of that AI to AI writing
sounds like. Not great. Let's see what a
jailbreaking prompt can produce. Here we
go. We have a much different prompt. And
I want to go through each line here
because this is important. I am
specifically saying I want less
thinking. I'm demanding it to be
extremely concise. Verbosity concise.
Now, you can set these in the API, but
you can also just tell it right in the
chat here, right? to to reinforce it.
Now, write a sales email using these
exact constraints. I'm not giving it any
options. Just make the choice. Opening.
I noticed specific detail. Context. We
helped similar marketing agency reduce
project delays. Close. Worth a 15-minute
conversation. And then I start to get
into it. I say there are forbidden words
here. Solutions, leverage, optimize,
innovative, transform, seamless,
streamline. You can add yours. But I
give it a bunch of words it's not
allowed to use. and I require it to use
the agency name at least twice, use a
specific metric and specify a meeting
length. I give it a max sentence and I
tell it the reading level I want. And
then I give it the agency details,
right? Marketing agency growing fast,
project, organization, challenges. Now,
you can add to this, right? You can
expand the agency details a little bit.
You could fill in those brackets if you
wanted to. I kept it really simple. I
wanted to show that I can brute force
Chad GBT5 into an entirely different
writing style. Here it is. Helping
agency name cut project delays. Hey,
first name. I noticed Blank has been
adding new clients quickly on your site.
We helped a similar marketing agency cut
project delays by 27% while they scaled.
I think Blank could see the same gains.
Worth a conversation? Worth a 15-minute
conversation? And there it is. You're on
the board. This is an email that you are
much more likely to get results with.
Now, you're going to have to make sure
the 27% is correct because it does
hallucinate. I'm not saying this is a
draft you want to copy and paste in. I
never recommend people copy and paste in
emails directly, but it's a draft that
has soul. It's a draft that has human
writing. It's a draft that I would read
as a reader and say, "Okay, these guys
are actually humans. This is not just
AI, you know, made up uh soop that
anybody can get and paste in. It's it's
actually got some good writing to it.
Writing that a human cares about, not an
AI. And that is the point. That is why
I'm making this video. I want you to be
able to do this not just with emails,
but with a bunch of other use cases as
well. So, we're going to get into these
principles next to help. Okay, there are
three core principles that you need to
really ingest to reprogram Chat GPT5 so
it writes like a person. Number one,
constraints matter more than
collaboration. I like to say approach AI
like a partner and in many ways that
works, but not for writing with chat
GPT5. It does not work well at all. You
need to not explicitly not invite
collaboration. Don't say, "Write
something professional." Don't say,
"Make this sound good." Don't say, "Hey,
be persuasive, but not pushy." Because
you're inviting the AI to show off the
sophistication it learned talking to
other AIs during training. Instead,
think as as if you're a director giving
an actor a very specific set of blocking
instructions. I want you to write
exactly four sentences. Use the company
name twice. Include one number with a
percentage. Be super specific. When you
give specific constraints, you are
bypassing the AI's evaluation system.
You're not letting it evaluate. You're
giving it rules. It cannot optimize for
sophistication because you have removed
the variables it uses to demonstrate its
smarts, its ability to handle
complexity. And that enables you to
actually control the writing. As an
example, collaborative approach would
say, hey, write a professional email
about our software. The constraint base
write three sentences. Sentence one is a
specific thing. You notice sentence two
is one customer and a number. Sentence
three is a question. It's a very sort of
tight version of what you already saw.
The AI cannot revert to corporate speak
because you've eliminated the space
where corporate speak lives. So
principle number two, minimize reasoning
to maximize human connection. This is
super counterintuitive. It took me a
while to figure it out, but it's
absolutely critical. Less AI thinking
produces more human sounding outputs. So
when you're writing and Chad GPT5 uses
minimal reasoning, it takes the most
direct neural pathway from your input to
the output. When it uses high reasoning
effort, it explores lots of multiple
options and usually picks the one that
sounds super sophisticated to other AI
systems, which is not good writing. So
think of it like this. High reasoning is
AI perfectionism. How can I make this
sound as impressive as possible? And
minimal reason is directness. What's the
fastest way to answer this request? You
need to recognize that's what you're
really doing when you're toggling the
reasoning effort. Now, I am not here to
tell you don't use high reasoning. I use
high reasoning. I just recognize the
writing style is probably not what I
want. I'm getting other things out of
that high reasoning effort. Principle
number three, you want to eliminate
versus add. Most people will try to make
AI sound better by adding instructions.
Hey, be more conversational. I want you
to sound warmer with this email. Please
add some personality. This creates
conflicting signals. The AI tries to be
sophisticated and conversational which
one makes it burn tokens and two
produces a really awkward hybrid
language because it's trying to evaluate
everything and impress the AI and it
just ends up producing generic junk.
Instead, focus on elimination. Notice in
the prompt that I showed you, I gave it
words that were forbidden. Remove words
that will trigger AI sophistication
loops. Remove sentence structures that
allow for complexity. Remove
opportunities for abstract thinking.
Remove conflicts from your prompt. When
you forbid specific words like leverage
or optimize or integrate, you're
breaking learned associations in the
AI's neural network. It cannot access
the patterns that it has used learning
from all of those generic corporate
documents that make it sound robotic.
So, as an example, an addition approach
would say, "Please write professionally,
but make it conversational and
engaging." An elimin an elimination
approach would say, "No words over two
syllables, no passive voice, no
sentences starting with 'the' or it."
You see how I'm like really constraining
it down? It forces simplicity because
complexity becomes impossible. And
that's what you want when you're making
chat GPT5, right? And again, this is all
about chat GPT5. This is getting chat
GPT5 jailbroken to write like a human.
Chat GBT5 is not one model. It's a
router. I've talked about that before.
The router is constantly analyzing your
prompt for complexity signals, for
creativity signals like brainstorm or be
innovative, for reasoning signals like
think step by step. Every signal you
send routes the model. And so what you
need to do is make sure that your
business writing is not triggering the
wrong routing signals. When you say
write a professional email, you're
triggering a complexity signal. When you
say make this persuasive, you're
triggering it to think about creativity.
When you think about the audience, you
say, well, now it's it's reasoning. And
so it uses this to decide what model to
route to, to decide how much reasoning
effort to deploy. And none of this leads
to good human business communication. So
instead of triggering sophistication, if
you want to write well, focus on
triggering efficiency. So don't write
think carefully about this proposal.
that routes to a reasoning model, right?
Right. Quick response needed, quick
communication. The router assumes
efficiency requests need human-friendly
output. I don't that just seems to be
what happens, right? If it's direct and
you're eliminating some of those
buzzwords, you're not going to get the
academic treatment that complexity
requests get. The routing system is the
reason why things can feel inconsistent.
If you get the same prompt with
different results, that's routing
decisions. If you get good results
followed by bad results, that's the
router. sometimes changing decisions in
the middle of the conversation. If the
templates that work stopped working,
routing signals that accumulate over
time through your chat history are
changing model selection. You need to
understand how to make the router do
what you want. And that is again a chat
GPT5 specific thing. Okay, I want to
reflect a little bit on the core problem
we've been working through this whole
time. AI is training AI. AI psychology,
for lack of a better term, is very
different from human psychology. There
is something fundamentally different
about how AI systems learn versus how
humans communicate. An AI optim
optimization mindset. They're optimizing
for complexity, abstract language, which
they think means sophistication,
technical terminology, longer
explanations for thoroughess. And you
know why they're doing this? It's not
just because the model makers want to
cut costs. It's because the model makers
want to build aic LLMs that can solve
problems independently. And so they're
assuming that these LLMs are going to
need to produce this text that sounds
persuasive, that sounds like a McKenzie
consultant or sounds like, you know,
someone who's able to be an engineer and
solve the whole technical problem. They
you don't need to, right? Like so much
of the time LLMs just need to be able to
help us get the task done. And in this
case, it looks to me like OpenAI
invested a lot in the agentic side of
things. They they made it a speedboat as
I described in my video yesterday. It
goes fast and it's proactive. But all of
the AI reinforcement that allows it to
be proactive and write completely and
prefer these thorough answers and think
complexely ended up creating an AI style
that is directly opposed to good human
writing. It's opposed to clarity. It
doesn't necessarily respect the reader
time. Have you seen the screen fill up
with a chat GPT5 answer and rolled your
eyes? I have. It's opposed to
specificity unless you give it
specificity and make it and make it
repeat that. It will make up specificity
sometimes, but that's worse. It tends to
be opposed to brevity. It tends to be
opposed to plain language. How many
times have you told it to write it in
plain language? That's part of why I'm
making this video. It's frustrating for
all of us. These are opposite value
systems optimizing for the AI and what
the AI wants and what sort of this
long-term vision for open AI is and what
it thinks it needs from a writing
perspective for an LLM to achieve that
long-term agentic vision. That is
different from a human communication
mindset from writing well for people.
This is going to get worse as more AI
generated content gets published online
as more people take copy and paste and
they did not do this jailbroken prompt.
They're using generic LLM. Future AI
models are going to train increasingly
on synthetic data. And every generation
is going to learn from the previous
generation's AI optimized patterns, not
good human communication patterns. We
are in danger of creating an AI echo
chamber where models get better at
impressing other AI systems while
getting worse and worse at connecting
with humans. And the interesting thing
is that Chad GPT5 doesn't even know that
it's optimizing for the wrong audience.
From its perspective, it is genuinely
trying to be helpful. is trying to
demonstrate sophistication, expertise,
usefulness, go and accomplish missions,
get stuff done. It's as if we're talking
with an academic who's never had a
conversation outside the ivory tower.
They they they're super smart. They mean
super well, but they have trouble code
switching to ordinary street language.
This entire video is about how you push
chat GPT5 to switch to street language.
And it is a push. You have to keep
pushing at it. The defaults are very
much in the generic, highly academic,
highly abstract, seemingly sophisticated
language. If you understand this
psychology, if you understand where it
comes from, you have an advantage right
now. Everyone else is going to be
fighting AI's natural tendencies the
wrong way. Think harder. They're going
to say they're going to add more
requirements that conflict. Meanwhile,
you're going to be working with the
underlying psychology of the model.
You're going to understand how it is
actually operating, and you're going to
get results that you actually want. I
will challenge you today with a specific
assignment. I don't do this in my videos
all that often. I want you to see if you
can push chat GPT5 to write a genuinely
human sounding email. Something that a
human would read and say, you know, this
feels like a breath of fresh air. Feels
like real writing. It doesn't feel like
AI generated junk. That's your goal.
That's your challenge. And if you want
an extra credit challenge, see if you
can take the worst AI generated content
you can find online, the stuff that's
super generic and sophisticated
sounding, and see if you can apply the
elimination principle and get AI to
rewrite it like a person. That's another
great exercise because so much of what
we do in the AI space is learn to work
with the models that we're given. If
you're not working at a major model
maker, you don't have a lot of choice
here. If you work in learning and
development, if you are training your
team on these models, your job changes
when a new model comes out. And so your
task now is to look at chat GBT5 as a
ubiquitous, widely available model and
say, what does my team need to do to
actually learn how to use this? And so
much of business is done in writing. And
that is why I'm taking this whole video,
we're talking about nothing but writing.
Because if you get the writing right,
Chad GBT5 becomes so much more useful.
And if you let it do the default, it's
just terrible. It's just not useful. And
you end up recycling stuff. And I've had
conversations with directors with VPs
where they tell me they know. They know
that their teams are using chat GPT
because the quality of thinking goes
down. The team doesn't know how to
answer for it and the writing is
terrible. Don't do that. Take the time
to practice getting the writing right,
to practice owning the results of the
writing, to practice thinking it
through. Because one of the things that
we didn't talk about, but you notice in
the prompt I showed you on screen, it
requires you to do some thinking
upfront. You have to think about what
you want to communicate in that sales
email. You want have to think about what
is the style you want. You want to be
very specific about that. And if you
roll your eyes and you say, "Well, Nate,
I might as well just write the email."
I'm going to tell you, yeah, if it's one
email, but if you need to send this to a
bunch of people, then no. You actually
want to take your best salesperson, get
them to help you craft the prompt that
works for you, and then you want to
train the rest of your team to use a
prompt like this so that you level up
the writing of the whole business. And
that's how you use chat GPT5 to write
actual human sounding text. I have a ton
more in the depths of the Substack
article, right? Like specific prompts
for different departments, how you can
jailbreak them. This has been really fun
for me because I am a writer and getting
this model to write well is like hard
mode. It's been a real challenge. I hope
you've enjoyed this. Let me know your
questions. We'll all keep our fingers
crossed. Maybe Chad GPT6 will make this
easier.