#6: AI Manners: The Etiquette of Artificial Intelligence Explained
Welcome back to another episode of The Junction.
We're going to be talking about AI manners today.
So the etiquette of artificial intelligence sorry, excuse
me, I just I had to cough.
Well, for starters, etiquette cover
your mouth when you cough. Yes.
And don't interrupt when someone
else is speaking thing.
I'm covering my mouth right now.
All right, I think we're adequately teed up.
Let's jump into it.
So there's a lot of how do I use it?
But some people aren't really thinking about the best
use or especially in the context of work. Right.
So it's one thing to kind of go out on
chat GPT and type in make me a meal plan.
But if you are in a role where
you're interacting with or touching customer data right?
Like you don't want to be going and pumping in an
export from your CRM database out into a chat GPT window.
Right, right.
It's probably not the I mean, I just don't see
a ton of use cases where you're typing in or
pasting in people's Social Security numbers right now.
You might do it on accident.
That's unintentional.
Anything that can be used for good can be used for bad.
Right.
So there's mean you start to wonder of what
what are those areas where I should be know?
And it's probably mostly aligned with a lot of
the current day nomenclature right around what Europe's doing
as far as privacy rules and things like know,
if those are things you would be concerned about
without AI involved, then that doesn't necessarily change the
way you should interact with AI. That's a great point.
Yeah.
We don't go out handing out all
of the information about our customers, so
I probably wouldn't do that with AI.
One of the things that comes to mind for me
is, well, there's two bias and then the hallucination, right.
So I think we should talk about those because that's
where I think some of these companies that are coming
out with AI ethics or principles, if you will, AI,
here's our use policy at the company.
Those are some of the things that
are top of mind for me.
You first have to think about, of course,
what kind of data do you have?
Is it something you should even be worried about?
And I think for the most part, if
you're in business and you're not worried about
liability, then maybe you're in the wrong business
or you need like a chief compliance officer. Right?
But as far as privacy and policies and things you
need to be thinking about, those are things that, again,
take out the AI piece and look at the data.
Determine what should we be concerned about.
Maybe this even goes for your CRM, right?
Like, if you need to cover up sensitive info that
maybe your internal people shouldn't be seeing, then you should
really start taking a look into and being concerned and
generating policies that ensure that your staff are aware of
what they should and should not do.
You see a lot of this in the
CRM world and in the ERP space, right.
People want to keep track of Social Security numbers,
credit card numbers, all these things that effectively identify
somebody in some form or fashion is something that
maybe you should include in those policies. Sure.
So that's utilizing your own database, your own data.
But again, let's go out to Chachi BT and type
in some prompts and then I'm going to go use
what the answer that it gives me and I'm going
to go use that for doing my day job, right?
Aren't there some guidelines that
we should be exercising around?
Hey, always have some editorial.
You go type up an email and you shoot it out
to a customer without naturally doing any kind of review.
Right.
You could be putting your company in a bad spot.
Oh, for sure.
You do, like you mentioned,
have to worry about hallucinations.
You do have to worry about when you are
relying on it to provide you factual statements. Right.
I'm thinking about that guy that pulled up cases,
case law, and he's referring to these cases or
the professor down south in Houston or it was
in the Texas A M University of right.
He failed his entire class because he relied on Chat
GBT, saying that all of the students work was know.
When you put that level of discernment on the
AI, you're naturally going to run into issues.
So I would first start off by saying the
data or the things that you're typing into these
large language models or you're expecting them to do.
You should almost act like, and I think I
said this on a previous episode, it's an intern,
they're going to do some pretty weird, stupid, potentially
unethical things, but they're not necessarily going to do
it because they're designed to do it like that. Right.
The hallucinations are simply things where it's just predicting what
it thinks is the next best thing to say.
Well, I can try to predict the lottery numbers, but I'm
still going to be wrong in terms of what people are
typing in and things that you should be worried about.
It's more about necessarily like copying pasting data in
that's one area you should be worrying about and
then relying on what it spits out.
It can hallucinate answers to math questions.
So if you can land somewhere in between right.
And not necessarily wholly rely on it for factual
statements, but potentially just have it write an email
that you revise multiple times, that's going to be
a sweet spot for ethical concerns.
From a business standpoint, can you do some of
the tools, let's say the paid version of the
tools, does that help mitigate some of the risk,
at least as it relates to OpenAI?
3.5 does tend to hallucinate more than 4.0.
And these are all things that they've said in
their blogs and their press releases and things that
I've seen on my end when I'm typing the
same exact question between models, it does respond differently.
And in the back end, if you're doing any
kind of developer work, you do have the ability
to kind of try to limit the hallucinations.
And there are actually methods in
the way that you ask questions
programmatically to reduce that hallucination rate.
And let's remind for those listening hallucination,
what's your quick one sentence definition?
They don't know.
Well, I'm going to go back to the lottery statement
like mel, the numbers for Powerball tomorrow are the five
numbers or six numbers or whatever it is. Sure.
Well, nobody can predict what the
numbers are going to be.
So that would be something that
it would be called a hallucination. Right.
It's telling you that the five numbers and
you're going to win tomorrow are these.
Because it's designed to give you an answer.
It wants to give you an answer with high confidence.
Unless it says, yeah, well, I'm
a bot, I can't answer that.
I've seen that a few times.
The hallucination, it's simply mimicking what
we do in real life.
And we talked about this on another episode. Right.
It's confidently answering your question,
predicting what it thinks.
The answer is not necessarily knowing that it's wrong.
It's just thinking like, well,
the first number is eleven. Right.
Has no idea.
It's just predicting and it's seeing in
the things that it's learned from that.
Hey, well, in a couple of
places here, the Powerball results.
I'm sure it picked that up through
gathering all the data on the internet.
It sees that number eleven pops
up in some of the results.
Well, I'm going to use number eleven. Mel.
The first ball is eleven. Right? Right.
And it confidently says that.
But if you took the time to look at
that and know the context of the question that
you're asking, you would know that it's wrong.
So as with all of these things, there's a spectrum.
There's harmless and very harmful.
And one harmless example of hallucination that I've
seen in the last week using OpenAI.
So I've been pumping in some transcripts, trying
to get some language to put on the
front end for some recruiting emails.
And if anyone knows anything about vent
technology, we have a beloved Yeti mascot.
If you don't know, go
out to our website, ventechnology.com.
He's adorable.
So Bjorn is our mascot.
At some point in these various transcripts that I've been
uploading some blogs again, trying to kind of put together
some language based on our existing data in three different
spots within the same like, I'm working in the OpenAI
playground and I'm asking it to tell me like summarize
this thing in three different spots.
OpenAI said that Bjorn was a different person or thing.
So in the first version, Bjorn was
the chief executive officer of Ven Technology. Nice.
Okay.
In the second version, Bjorn was the office dog.
The office pet. Okay. Yeah.
And in the third version, I think he may
have actually been another like a team member contributor.
But it was just interesting how we're
and I didn't correct it ever.
I wasn't like, hey, Bjorn's not a dog. Right.
Bjorn's our mascot.
Actually, I do think by the fourth iteration I
did, it was slightly entertaining just to see what
it would come back with each time.
But why from version to version?
Even though I didn't ask it to change Bjorn, I was
asking it to, hey, make it more conversational or witty.
Why did it all of a sudden decide that Bjorn
was no longer the CEO and he was the office?
Like that's a harmless example of what it can do.
So again, if you're not, you start to
think about some of the more personal identifiable
information or customer information, things that actually could
potentially cost the business loss and revenues.
That's where my head goes.
Yeah, you go back to this intern idea, right?
The intern just stepped in today and he doesn't
know, or she doesn't know who Bjorn is.
And if you said, I need a
statement right now, give it to me.
Well, that person's going to say,
well, I don't know, Bjorn.
Sounds important, let's go with you.
You are forcing me to type something right now.
I think where it tries to iterate or just
give you a different answer every time is dependent
on the prompt that you are giving it.
And in some cases I think it is playfully just
switching out what it doesn't know with something else.
And that goes back to the idea of this kind of
one shot mentality where within one prompt you ask it a
question and you expect it to return the correct answer.
What tends to work a whole lot better is where
you do this iterative approach and you start to weed
out some of the I'll call ethical concerns. Right.
You can ask, hey, where did
you get that information from?
Bard does this pretty well and it actually will
refer to the website where it found the information.
Chat GBT doesn't do that yet.
There are some plugins that it can
quote, unquote, connect to the internet.
But you do have to be careful, right?
Like, if you don't know the answer and you're expecting the
AI to give you the answer and you don't have any
way to check it, well, you've got a problem, right?
Right.
So you probably shouldn't ask know
like theoretical physics questions, right.
Or how quantum computing works.
You probably should ask it about things that you
already know or you're already an expert in.
So you can be the one that verifies them or
maybe you have a team that can verify them.
It's like the example of the CFO that used
it to verify some gap compliance or something, right.
That's his domain. Right.
But you start thinking about how many of
those rules and number of compliance things that
you have to keep up with in that?
Um, it's kind of like that gut know?
I could go find it. Yeah.
Well, I mean, here's a great example as a
leader in our business, Mel, if you won the
lottery, we'd be poop up a creek.
And what I would first do is GBT, what
is a great marketing plan for an ERP integration.
And it's going to spit something out, right?
And I'm not the best marketer, but I'm going to
take it at face value and be like, sounds like
you've done this before, maybe we should do that.
You have a backup plan for me?
Just waiting because you're so awesome when this
podcast blows up, but then we'll both go.
But those are some of the things that you
probably should be thinking about from an ethical perspective.
I'm asking this new tool, I'm asking this intern
to provide me questions that I'm potentially going to
make really big decisions off of that's like literally
the intern walking in the door and being like,
hey, here's the reins to my business.
What should we do?
So if you can avoid that, that would probably be good.
You don't want your business to run into the ground.
Let's move on to headlines.
So Samsung bans staff's AI use
after spotting chat GPT data leak.
Let's talk about that high
level summary, what happened?
Yeah, these guys are using I mean,
this is probably not just Samsung, right?
People across the globe are utilizing chat GPT to paste
in data, to collect insights, to ask it questions.
And so it's super easy just to paste something in.
Well, you could paste a CSV file, right?
Grab some data out of salesforce or from NetSuite or
Intact or, you know, whatever platform and boom, you don't
even know it, but you just paste it in some
Social Security numbers, right, or onto some external server or
database somewhere and boom, it's out there and you can't
delete it and you can't delete it's already gone.
The moment that you press
Enter, there's no deleting, right.
So you have to be careful about these things
because this is what ended up happening, right?
Samsung staff, they went in, they pasted
something in that was not good.
I think I've seen a couple of
cases where people were pasting in proprietary
code and now that's been transferred over.
It's like an instantaneous, hey, here's
the keys to the kingdom, right?
And the moment that you press Enter, it's over.
So they ended up banning chat GBT and
now supposedly Samsung staff can't access it anymore.
Do you agree or know?
We have spent some time with our lawyers to determine,
based on the policy that it's posted all over the
website, what are they doing with the data?
And we don't necessarily know exactly what they're doing, but
they do specifically say that we will take that data
and we will train our models off of it.
So just the fact that maybe I paste in some
Social Security numbers, well, OpenAI hopefully has some safeguards, right,
to potentially strip that out at some point.
But now the large language model, let's remove
any safety measures that they may have.
Now, the large language model is training
off of Chase and Mel's Social Security. Okay?
So next version comes out and Randall says,
hey, what is Chase's Social Security number?
Well, it knows it, right?
And it types it and it puts it in there.
And now anybody that asks, hey, do you have any
Social Security numbers it's going to paste or it's going
to type in Mel and Chase's Social Security numbers because
it trained off of that, it knows that information.
Thankfully, OpenAI has a number
of safeguards and rules.
And you've probably seen these answers that pop up.
It says, I'm sorry, but as an AI, I
don't have access to well, I for the record,
I don't know about you, I haven't been asking
it for a bunch of Social Security numbers.
Can't speak for Chase, but just give me
all the Social Security numbers so I can
open a bunch of credits again.
Anything that can be used for good can be used for bad.
Oh, absolutely. Yeah.
I'm wondering too.
I'd like to just kind of open
this up around the free version, right?
You go back to if you're not
paying for it, you're the product.
Somebody somewhere said that. Oh, absolutely. Long line.
So would some of your recommendation
be for anyone out there listening?
Would an additional safeguard be you know what,
I'm going to look into maybe what a
paid version of this would cost.
If I'm going to embrace it and I'm going
to put some policies at my company around it,
do you think it's better to go ahead and
just pony up, pay the subscription fee?
Here's the direct answer to that.
We paid our own money to go figure
out the answer before actually paying for anything.
The lawyer on our end said that based on
the way that OpenAI's policies are for paid accounts.
Now.
This isn't for chat, GPT.
This isn't for Anthropic or any of the other organizations
out there specific to OpenAI, at least as of today,
that if you use the API to access these models
that that data is yours and only yours.
Anything you put through there is not stored by OpenAI.
And we don't have to worry to an extent,
right, that we're pasting in Social Security numbers.
Now, you probably still shouldn't do that,
but based on the way the policy
is written, they are not storing that.
So if you're interested in utilizing these things to
potentially paste in potentially proprietary info or trade secrets
or things like that, or really just in general
having a good sense of not worry, right, but
comfortable, some peace of mind. Peace of mind. Yeah.
Right.
Do it through the API.
And if you don't know how to access the
API well, I happen to know some people.
You should send us an email. Yeah.
So before we get off of that topic, if
you are interested in okay, that sounds great.
What's that going to cost me?
I was actually surprised to learn how minimal,
at least in our initial stages right.
We're still just testing. Yeah.
I asked Chase today.
Hey, I've been blowing up open
API with all my transcripts.
How much money did I hit?
How much usage? Big bucks.
Mel, we're going to have to talk
about that in the budget meeting.
Was it like a dollar and thought it
was a dollar 92 in one month?
Mel, we're really going to have to talk about that.
And I used many, many multiple, probably
20 or more transcripts and generated the
equivalent of three blogs and a dollar.
And think you spent more than
that just driving to work today.
Oh, 1000%. Yeah.
That's what's great about these things, though.
And that's where some of the
ethical concerns come in, right, from.
Well, they're paying me to write content. Right.
And I just did it on Chat GPT and it
cost me a dollar, but they paid me $1,000. Right.
That's where I feel like some of those ethical
concerns come in that we were talking about.
But from a cost perspective, using the
API is, in theory, relatively inexpensive, depending
on what you want to do.
If you want to pipe in tons of data into it.
It does work on this idea of a
token model where basically take a four letter
word and that ends up being three tokens.
It's a little more complex than that. Right.
Math was not always my forte, but
the math doesn't seem to add up.
There no four letters, three tokens, and
they charge you based off of tokens
depending on the model that you're using.
The more complex or the better the
model, the more expensive it is. Right.
So when we move some of our stuff over to
the 4.0 model, things will get naturally more expensive.
If you use the cheaper models, the
faster models, things will remain relatively cheap.
Where you have to be concerned is where
you're typing in an incredible amount of context. Right.
Like the summaries of your calls.
If they're 15 minutes, probably a little bit cheaper.
If they're eight hour training sessions, that's probably
going to be a lot of money.
And then you do it 100 times in a month.
Well, now you're starting to spend some big bucks. Sure.
But you look at, again, kind of the cost
for maybe a full time resource to develop.
And again, we are never talking in the context
of ultimately eliminating the staff person or right.
As you're saying, well, no, I pay you to write blogs.
Right.
And you just went and did it in however many minutes.
What are you doing with your time.
Well, I think that's where a lot of some
of the other headlines we've been looking at, right.
That's where there's a lot of concern of like, well,
this thing can do it for a dollar, but this
person is asking me for $25 an hour.
The average business person is like same
ish level of content or quality potentially
depending on who we're talking about. Right.
There's a lot of levers there but I can
pay a dollar or I can pay $25.
It's like my wife and I went to the
websites for two different grocery stores and we put
in the same stuff for delivery and Tom Thumb
was literally $12 more for the same exact stuff.
I mean, it's pretty obvious the
direction that we're going to go.
And for this AI versus human deal, depending on the
scenario that you're talking about, AI is always going to
be and I know I might get thrown under the
bus here, but AI, depending on the scenario, is going
to be way cheaper because the AI never sleeps. Right.
It just does exactly what you want it to do.
Whereas human resources, we want to
eat, we want to drink, right.
We need to go home and sleep.
We want raises, we want promotions.
And there is a whole wealth of other individuals
that are influencing us potentially to leave or check
out or go to the next thing.
On the flip side, you have someone now
that can do more with the time given.
So you're not saying, well, come work for me for
4 hours and I'll let the AI do the rest.
I do think that we're all under the I want
to do more, faster, better, and delivering on tight timelines.
And sometimes you've got people out there
who are working on very lean teams.
I do think that you can make the argument that you're
making now around, well, now I can go do it cheaper
with a tool, but the time that it took me to
do that, I'm now able to do more of it. Yeah.
And I think the hope and the direction that
most people are going are thinking at least this
is what I'm thinking is, well, I want to
work 40 hours a week, maybe lower than that,
let's call it 30, whatever, I'll save 10 hours. Right.
But I want to continue to work.
I don't think anybody's out there is like I just
want to sit around and literally do nothing all day.
Maybe there are some folks right, but what I think
we're going to end up seeing is people are going
to come in in that 30 to 40 hours mark
and they're just going to naturally increase their productivity.
I don't see these tools coming in and just
making your job so easy that you can do
40 hours of work within 1 hour. Right.
I think what's going to end up happening is
you're going to work 30 or 40 hours.
And this tool is going to help you make
it seem like you did 80 hours of work.
It's more of like, think on
the happy side of it, right.
Rather than somebody trying to overutilize this to make
them look like they're doing 40 hours of work,
but they're sitting at home drinking some beer. Sure.
All right. Hot take.
So do you think businesses should bear legal
liability for the actions of AI systems, or
is the responsibility with the developers, the manufacturers?
This goes to the idea that Tesla or
any of the other car manufacturers that are
developing these self driving vehicles, right.
Are they the ones liable or is the
person that is driving the car liable? I don't know.
You tell me, Mr. Chase?
Who owns a Tesla?
That's a really tough question, right.
Because it's like, well, let's say it did something wrong and
all I was doing was sitting in the seat, right.
And it decided to run into a wall.
Well, I didn't drive it into
the know in this perfect scenario.
And in that I think the people
that wrote the software are still liable.
They're the ones that programmed it unintentionally
to run it into a wall.
If I was the one that put it in the direction right.
Or I steered it into the wall, well, now
I am responsible, which at this point, there are
audit trails or audit logs of those things. Yeah.
Oh, absolutely. Right.
And it's the same for the AI piece, right?
It's like, well, are you responsible, Mel?
Are you responsible for what people
do that are underneath you? Right.
If you're a business owner, are you responsible
for your employees and what they do? Well, absolutely.
That's why we have liability insurance. Right.
And I don't think this is any different where you
go to the self driving versus the person driving.
I think I just don't see a ton of
issues where this large language model is going to
connect to the Internet and spill the beans.
I mean, sure, it's possible, but if you're using
the API, your team or you are the one
that is coding it to do something very specific.
And so I think that liability still relies on
or not relies, but still lays on you and
your team for doing whatever it is they do.
The OpenAI isn't the one that's copying and pasting
Social Security numbers in that's your team doing it.
So kind of wrapping up this
topic on AI ethics and responsibility.
Some of the things that we've done here internally
is initially we had a hey until we know
how to best utilize it professionally within the work
setting, let's put a hold on use.
Since then, we have opened up the conversation.
You actually led a really awesome
lunch and learn with the team.
There was a lot of great feedback, and you
could tell that we have members of the team
who are really excited about this kind of technology.
And one of the things that we've been
talking about is actually putting together a committee.
And I think that, depending on the size of your company
and where you're at, I'm a big fan of opening it
up to people who are excited about it, who are kind
of on the bleeding edge versus just, mandating.
Well, you're a senior leader.
It's now your job.
But I would recommend open the conversation,
consider putting together a committee, and then
figuring out what is ethical today.
AI aside, in your industry, in the
best interest of your customers, your partners,
and your team members, and start there.
Start putting that stuff
together, that structure together.
Do your due diligence, right?
Like, take time to figure out you don't want to
throw your business, your process, your department, your team under
the bus simply because you didn't take enough time to
figure out, hey, how should I do this?
What is the best way to do this?
You don't go out and buy a car just
on a whim, let me spend 60 grand. Boom.
And I haven't even test driven the car, or I never got
in it, or I don't even know what it looks like.
If you do your due diligence, you're
going to be in a good spot.
All right, well, as always, we are
interested in hearing from our listeners.
What is your take?
Have you implemented an AI
regulation or policy or company?
If not, what other concerns you have?
What other questions do you have
that we didn't talk about today?
Or did you mess up? Right?
Did your team mess up?
Tell us what went wrong, and tell us
how we should be thinking about these things.
Absolutely.
Email us your take at thejunction@bentechnology.com
until then, keep it automated.
Sam.
