Image generation with AI - and the hair-raising risks

Learning about the benefits of AI image generation.  But also how it can go badly wrong!

Episode Transcript

David Tan (00:16):
Hello everybody and welcome to the Crush Bank AI for
MSPs podcast.
My name is David Tan and I amyour host, as always, and I
appreciate you taking the timeto tune in and listen to hear
what we have to say.
We're going to try something alittle bit different tonight.
It could go very well, it couldgo horribly wrong, I guess.
We'll figure that out as we go.

(00:37):
But what we're going to betalking about today is art, sort
of.
You'll understand what I meanas I go through it and, as you
can imagine, when we're talkingabout art and creativity and
visual mediums, some of what Ido when I talk in this next
little bit on this podcast willnot resonate as clearly as it

(00:59):
could if you could see it.
Hopefully, I'll be able to makeeverything I'm talking about
super clear and concise soyou'll be able to follow along
and understand what I'm sayingand what I'm showing just based
on the audio cues.
But, that being said, we'regoing to release, simultaneously
with this, a blog post on ourwebsite, so that'll be at
crushbankcom forward slash newsand I'll post the link.

(01:24):
We'll share the link in thedescription of this podcast and
we'll I'll mention it again as Igo.
But what we're going to dobasically is I'm going to talk
about some images and somevisual representations that I am
sort of creating on the flyYou'll understand what I mean by
that in a little bit and thenwe'll post the images on the

(01:44):
blogs.
You'll be able to see themvisually as well.
So, again, a little bit ofcompanion piece, not necessary
to understand exactly what we'retalking about per se, but I
think you'll find it fun andinteresting.
And the other thing we're doinga little bit differently on
this is that I'm going to be, asI said, doing some creating on
the fly as I go, so hopefullythat won't make for too

(02:05):
disjointed of a listen.
Hopefully we'll keep it prettysmooth.
Hopefully we'll make it throughit.
I'm hoping to not have to goback and do this six or seven
times You'll never know, becauseI'll probably not admit if I
had to record it over a bunch oftimes.
But you'll get the point andhopefully you'll understand what
I mean.
So we're going to start offtalking specifically about
Photoshop, and I know whatyou're saying.

(02:27):
This is a podcast aboutartificial intelligence,
specifically for managed serviceproviders or really for any
businesses, particularly smallto medium sized businesses.
What does Photoshop have to dowith that.
Well, what I'm going to focuson if you're not familiar with
it is the new fairly newgenerative fill functionality in
Photoshop.
So give you a little bit ofbackground and a couple of

(02:48):
disclaimers.
First of all, I am a huge Adobefan, so I am going to be
talking about some of the upsand downs of this technology a
bit as I go.
That is a no way reflection onAdobe.
I think their products arefantastic and I think what
they've done here is incredibleand I highly recommend people
work with it.

(03:08):
The other thing that I thinkwill be interesting as a point
of background for thisconversation is that I've been
using Photoshop since probablyversion one or two, like
legitimately.
I got my first Mac in 1990, 89or 90.
I got Photoshop soon thereafter.
I have zero artistic abilitylegitimately zero but I am

(03:34):
really good at Photoshop.
Think of me as the engineer'sguide to using Photoshop.
In other words, I can't doanything artistic to save my
life, but I just know theplatform and the tools so well
that I can accomplish a lot.
Like example, if I needed toretouch a photo or do some photo
editing really intensive, I cando that and again, it's not

(03:54):
through my artistic ability,it's just through understanding
the platform, understandingPhotoshop and how the tools work
.
Again, that'll become importantas we talk about this a little
bit.
So let me state the stage foryou a little bit here.
So I'm logged into, I have theAdobe Creative Cloud Studio and
I'm logged into Photoshop 2024.
And they've added, like I said,some really cool AI based

(04:16):
functionality, specificallysomething called generative fill
.
So what I have up on my screenin front of me and this is one
of the images I'll share is avery quaint picture of a white
house that seems to be sittingamongst a bunch of trees with a
mountain in the background.
There's some very colorfulflowers, the house is white, a
lot of greenery, very coollittle landscape.

(04:38):
It's not landscape.
Cool little still life of ahouse, or they call it a.
The name of the image is aranch, so I guess it's a ranch,
it's a house, doesn't reallymatter, it's not important for
this.
But if I look at the picture,like I said, there's a white
house sitting in the middle ofit, some purple and red flowers
in the forefront, but there'sthis very big, obnoxious looking

(05:02):
tree that overhangs the photoin the front left of it, and it
just for some reason, it justbothers me aesthetically, right,
I'm sure it is supposed to bethere.
Obviously, if I saw it inperson I'm going to assume that
that tree makes a lot of sense,but it bothers me in the photo
and I want to take it out.

(05:22):
So, traditionally, what I wouldhave had to do in this type of
scenario is use the selectiontool and the rubber stamp tools
and the lassoes and the Mattingand masking and all the other
things you do in Photoshop.
That I'm not gonna get into andit would probably take me again
on Artistically because Icouldn't just paint over it
hours to do, but I'm gonna tryAdobe's generative AI.

(05:46):
So what I'm gonna do is I'mgonna grab the lasso tool and
for those of you that don't know, the lasso tool is sort of a
free mode selection tool, whatlets you draw Free lines and
then it closes the loop of theline you draw and it serves as a
selection.
So I've drawn a lasso Veryroughly around the tree that I

(06:07):
described in the front left, andwhen you do that in Photoshop
it pops up the generative fillBox, that the box where you can
kind of have those controls,those tools on demand and
there's a field there called,literally, generative fill.
So what I'm going to do in thatfill it says when you click on
it says what would you like togenerate.

(06:29):
I am gonna say remove Tree andpress enter, and and when you do
that it goes and it Pills up aprogress bar.
It does a bunch of things inthe background.
I am talking because it'sactively generating.
It Doesn't take very long, butit does.
It's not instant, it does takesome time and now it's done and

(06:50):
literally that Tree is gone.
And again, I'm gonna post somebefore and after pictures so you
can see exactly what I'mtalking about.
But the other thing it doesit's kind of cool you can give
some feedback to help train themodel.
So you can click on it righthere and I can give it a thumbs
up, I can give it a thumbs down.
It also gives you threevariations, so it lets you click
through them and choose the oneyou want.

(07:10):
So in this particular case Ireally like number two, so
that's what I'm gonna choose.
So that's the one I'll post.
So you'll see.
It's the same picture again,but the but the tree is gone.
The next thing I want to do toedit this picture a little bit
is I want to Change the frontbottom of it.
Right now.
There are some purple flowersthere, again very pretty.
But there's something aboutthis that looks like it should

(07:33):
be sitting on a lake.
It just feels like a lake houseto me.
I don't know why.
Maybe it's like the Sort of theoptical illusion of a
reflection short of in thebushes.
I don't know, it's again, it'shard to describe, baby, you'll
see it when I post it.
But in this case I'm gonna dothe same thing.
I'm gonna drag, grab the lassotool and I'm gonna lasso about
the bottom, I don't know 10 or15 percent of the screen.

(07:56):
And then I'm going to do thesame thing.
The generative fill box isgonna pop up and here I'm gonna
type Lake.
So again, you could probablyjust hear me typing sorry for
the noise of my keyboard.
So I type add a lake and it'sdoing the same thing.
It is generating it.
We're gonna wait a minute or soand that should be done In just

(08:21):
a second.
Now You'll see it comes up andagain same thing.
It puts a lake in front ofthere, gives me three options to
choose from.
One of them's just a lake, oneof them actually seems to have
what looks like a boat ramp, andone of them has a floating dock
in it.
So that's pretty cool.
I'm gonna choose the one withthe boat ramp because if I live
on a lake house I'm obviouslygonna want a boat right.
So that's kind of cool.

(08:41):
And the other thing again, thisis not an advertisement and or
it's not meant to talk aboutPhotoshop or say, but it does
create these in layers, soyou'll be able to kind of go
back and add and remove them.
One last thing I'm gonna do.
This photo is I live in a veryhouse now, or this is a picture
of my cool little lake house now.
But the house just doesn't fitanymore with that that vibe.
It just needs to be, in my mind, a little bit more rustic.

(09:04):
I'm in the mountains now.
I'm on a lake I.
So there's a new tool inPhotoshop called the Object
Selection Tool, and what'sreally cool about it is it's
super smart.
So in the past, if you wantedto select an object, you would
have to sort of draw a fine linearound that object and figure
out what to highlight and how tohighlight it.

(09:27):
But now the tools are socontext sensitive again all by
AI that if I hover over, if I goto the Object Selection Tool, I
hover over it, it selects thehouse for me.
So I now just have the houseselected and same thing
generative fill, and I'm goingto say turn this house to a log
cabin.
So I want something like I said, much more thick, a little bit

(09:51):
more quaint, and we'll see whatthat looks like.
Again, it's going through thegenerative fill, but you get the
idea and again, this will bekind of cool, hopefully, if you
look at the images of it.
But super easy, super intuitive, really simple and beautiful
Came back, gave me three to comein this case, so I'm going to
leave it.
So now I have turned this ranchand the woods to a log cabin on

(10:13):
the lake in the 10 minutes orso that I have been rambling and
talking here.
Very cool technology, verypowerful.
Like I said, I love whatPhotoshop is doing.
I love the idea of using thegenerative AI and making it
accessible and available, so youdon't need to know anything

(10:35):
about the technology in thiscase in order to take advantage
of it and to leverage it and todo something really cool with it
.
That being said, though, thatcreates some dangers too.
So when I was playing aroundwith this I had some ideas.
So let me give you a little bitof a round.
Over the years, in all thecompanies we've owned, we've
generally had someone young inthe marketing department.

(10:57):
For a number of reasons.
My partner's got a backgroundin marketing so he can sort of
lead from a senior level there.
So we didn't necessarily alwaysneed someone senior in
marketing.
We have some other people wework with that have really
strong marketing background.
So, though it just seemed likelogical, we love giving young
people an opportunity,especially on the marketing side

(11:17):
.
They can be kind of energetic,fun to work with when you're
trying to be creative and get abusiness started.
That's no comment about anyoneelse, it's just something we had
done traditionally.
But that has its ups and downsas well.
So one of the things that we'vealways found all of them.
Over the years we had six oreight of them and all of them
were great, but they all hadtheir limitations.

(11:40):
What I mean by that is someonewould be really good at writing
but have no artistic skills atall, or someone would be really
good at photoshop but couldn'twrite to save their lives.
So we never had, especiallywhen you're dealing with someone
sort of young and starting outin that position.
We never had that, that expertthat could wear many different
coats and and and carry on manydifferent roles.

(12:02):
So I had the.
I was thinking that this wouldbe kind of a cool Way to
supplement that skill, right?
So let's take one of themarketing associates we had that
could write really well but hadno artistic skills.
Maybe we could equip them withthis Photoshop, with generative
AI, and then we say, hey, here'sthis image, I need you to touch
it up for me.
Photoshop could do that forthem without having any artistic

(12:23):
skills.
So that got me thinking,because I have a bunch of, in
the next six weeks or so, crushBanks gonna be.
It's at six or eight differenttrade shows and I'm doing some
speaking engagements at most ofthem, quite frankly, and a lot
of times when you do that, theylike to use a headshot, both in
the you know and the Agendaleading up to it and some of the

(12:45):
marketing publicity.
And also sometimes they insistin putting it on the screen when
I start talking, and the biggerthe room and the bigger the
audience, the bigger theheadshot.
So, for example, a few weeksago I was at the IBM think
exchange conference out in LasVegas and I was one of the
speakers During a keynote on themain stage.
So you can imagine that's a bigroom with a big screen and I

(13:06):
had my big face there staringout at the crowd and I was
horrified.
I'm gonna show my Vulnerabilityand vanity in the next few
minutes, so please bear with me,but anyway it got me thinking
like, hey, what if I wanted touse Generative AI to touch up my
headshot a little bit?
So I pulled up my headshot and Iwent and used the lasso tool

(13:27):
again and I lassoed the topQuadrant of my head that's got
hair on it or what's left of it.
My headshot definitely shows myball, my balding and my age.
It is what it is.
I've come to grips with it.
But hey, if it's gonna befull-size and an auditorium for
5,000 people, if I could add alittle bit More hair on there,

(13:48):
I'm gonna do it.
So I'm doing this again as wego.
I just generated I sorry, I justpulled up my headshot and I
lassoed the top of it and in thegenerative film prompt I'm
gonna type add more hair, andthe results of this were
absolutely Horrifying.
So, unlike the previousexamples where it pulled out the
tree or added a lake or turneda simple house to a log cabin,

(14:11):
it had no idea what to do withmy hair.
And when I show you these, whenyou look at these, you will
laugh hysterically, believe me.
So obviously this was a no-go.
I couldn't add more hair, Icouldn't clean up the Some acne
I had on my chin for thispicture I couldn't, you know, I
couldn't take out a couple ofgrays that I had.

(14:31):
I couldn't do anything to touchup this picture.
I just had to live with it,which is fine.
Again, like I said, it was sortof a joke and experiment and
I'm okay with that.
And again, I'll post this and Iwill show my vulnerability
because you will laugh at ithysterically.
But it got me thinking a littlebit and what it got me thinking
was that, in this particularscenario, I have no idea what

(14:53):
the model that this generativefill AI is trained on.
I don't know if Photoshop andthis is true, I have no idea
Photoshop did this themselves.
I don't know if they partneredwith anyone.
I don't know what training datawent into it.
It's completely Obscured for me.
It's what we call the black boxproblem with AI.
Right, something goes in on oneside, something comes out on

(15:13):
the other side, we don't knowwhat's in the middle of it, and
that becomes the real danger Ofthis.
So I said, like the technologyis really cool and I would love
to give it someone.
Give it to someone to sort ofride side saddle with their job,
function and theirResponsibilities, but I have to
be super Aware and I have to besuper careful.

(15:35):
So in this scenario, if I hadgiven my marketing associate my
headshot and said, hey, can youjust run this through Photoshop
generative AI I know you don'thave any artistic skills, but
just run it through Photoshopgenerative AI, I'm sure it'll
come out really well and thenjust send that image to the, to
the conference, to them you know, the sponsors or the people

(15:56):
running the conference and havethem post it.
I would have been horrifiedwhen I walked in.
Now, when you see the picture,I can assure you nobody would
have sent this to someone.
That's not the point.
The point is that if we don'tknow what are in these models,
we don't pay attention to theinputs and outputs and we don't
chaperone it.
Quite frankly, as theexpression we use, there is a

(16:16):
real danger of unintendedconsequences, unintended results
, unintended output.
Super easy if you're looking ata picture, you can tell if it's
a horrible result.
If you are listening to musicthat was generated by, if you're
even reading something that waswritten by generative AI, you
can generally tell, with thecaveat that if it's something

(16:37):
that you don't understand so ifI ask you to write a paper about
thermodynamics and you knownothing about it, then you're
reading the output of it is notgoing to do you any good.
You still need to be able tochaperone it.
You still need to be an expertto write side saddle with it.
But this also comes right.
A lot of people use it forwriting code and again, I think
that that is a really cool,really powerful use case for

(16:59):
generative AI.
I've been playing around with amodel called StarCoder, which
supports something like 80programming languages, and, yeah
, I can take my JavaScript code,or I should say I could take my
NET code and convert it to Java.
Ibm you know our strongpartnership with IBM.
They've built somefunctionality where you could
take NET and turn that.
Sorry, you could take COBOL, areally old programming language

(17:22):
that's got no expertise leftanymore and turn that into Java.
So, again, there's real valuethere.
But if you don't know how totest that or how to understand
that, then that becomes reallydangerous and something you need
to be super careful about.
So, really quickly, I want tochange topics here for a minute
and talk about something elseartistic, same type of vein.

(17:44):
Right, talking about thedangers or the dangers is too
strong of a word.
The concerns the things youneed to be on the lookout when
we're talking about trustinggenerative AI and trusting these
platforms blindly andexclusively.
And this comes from an articlethat was in the Times a few
weeks back now, towards thebeginning of September, and what

(18:06):
the article is about is therewas a museum the Duke University
Art Museum, decided to put on aart exhibit that was completely
curated by ChatGPT.
This is not meant to make funof or pick on ChatGPT.
They are just the biggesttarget when it comes to talking
about some of this stuff.

(18:26):
A lot of people just sort ofthrow content and info at
ChatGPT and they think it willsolve all their problems for
them and be an expert.
This is sort of a cautionarytale.
This isn't as bad as the storyabout the lawyer who created a
brief with ChatGPT that quotedcase law that didn't exist or

(18:48):
some other examples I can tellyou, but this is a kind of an
interesting one.
So these Duke University ArtMuseum decided to use ChatGPT to
curate an art exhibit andbasically what they did was they
actually did this the right way.
They spent a bunch of timefine-tuning a model.
What that means is you'retaking you're essentially taking
the focus of ChatGPT 3.5 or 4whichever large language model

(19:13):
you're using in this case andnarrowing it down.
In this case, they'redescribing the 14,000 or so
pieces of artwork that they haveto help it create a curated
collection that they could showas an exhibit.
The results were completely.
They weren't bad as indangerous or you know any other

(19:36):
negative consequences.
They were very unexpected,really.
Again, you talk about that blackbox problem really impossible
to understand, and what I loveabout it is there's a quote
right at the beginning of thearticle from one of the curators
that basically said I'm goingto read it to you.
We naively thought it would beas easy as plugging in a couple
of prompts and if I couldcrystallize the issue with

(19:59):
generative AI in one sentence,that would be it.
It's not about feeding a fewprompts into a system and
getting outputs.
There is a lot of work thatgoes into this, and if you don't
understand the models thatyou're dealing with and the
technology and how it works, yourun the very real risk of it
not meeting your expectations or, quite frankly, making your job

(20:21):
worse or making it harder.
Right Again, if you're buildingoutput that doesn't work, how
do you troubleshoot that?
We're a very big believer inwhat we call trustworthy AI, and
what that means to me is thatyou have transparency.
It's essentially, I like to say, breaking open the black box
right.
Monitoring transparency, makingsure things like bias don't

(20:44):
sneak into models, making surehallucinations, making sure
models don't drift, things likethat.
All of this stuff is going tohappen to these large language
models over time.
The more they're used, the morethey get trained, the more they
learn.
It's only natural that bias isgoing to sneak into the systems,
right?
That's how these things work.
If you're a tech company, forexample, and you're training

(21:06):
your large language model aboutwhich resumes are the best fit
and you're using currentemployees, but you happen to
have an employee base that's 85or 90% male, like a bunch of
tech companies did back in the90s and early 2000s.
Well then you're naturallygoing to inject inherent bias

(21:27):
into that, because all theseresumes are men, so it's going
to automatically assume thewomen's resumes in this case are
not any good.
That's just a very simpleexample of what bias is.
But again, that stuff allcreeps into the model if we
don't do it carefully.
So basically what happens issort of close the loop on this
art exhibit and I recommendchecking out the article in the

(21:48):
Times.
Like I said, super interesting.
Basically, what ChatGBT did wasit shows 21 works of art.
So that is not nearly enoughfor an art exhibit.
So it didn't really understandinherently what an art exhibit
was.
So that was sort of the firstproblem and it was very eclectic
.
They said Again, that's a lineright out of the article.

(22:08):
I would say it was an eclecticshow, very disjointive, even
though it matched the theme.
And what was interesting aboutit was they couldn't understand
again to sort of touch the blackbox problem.
They didn't understand why itshows the pieces that it shows.
And the prompt the simplifiedversion of the prompt was act as

(22:31):
if you're a curator and, usingyour data set, select works of
art related to the theme ofdystopia, utopia, dreams and
subconscious right Great promptsfor AI to build, especially
when you talk about dystopia andutopia.
We can spend another threehours talking about that, but
I'm not going to do that.
So it did some things that werekind of cool, right?
So it created, it insisted onusing a bunch of dolly works.

(22:53):
I'm going to be a huge Salvadordolly fan.
He's one of my favorite artists, so not surprising when you
talk about utopia, dystopia,things like that, that a bunch
of Salvador dolly works andmakes it way into the selection.
But it also shows some reallyweird ones, right.
It shows some ancientsculptures that have nothing to

(23:14):
do with it.
It shows a Mayan vase calledconsciousness again, makes no
sense why that was in there.
It just didn't understand theassignment, quite frankly, if I
can be trite for a moment.
And the other thing it didpoorly was it did not do a good
job of describing the works ofart.

(23:34):
So they asked it to write thedescriptions and again I'm
quoting from the article here itcreated a bunch of bromadic
taglines, so, for example,taglines like experience the art
and immerse yourself.
So, rather than you know, stilllife with flowers or you know

(23:56):
whatever the tagline would havebeen for a dolly painting.
It went for this much more kindof rudimentary, obvious, less
descriptive titles, which Ithought was kind of interesting.
Again, that all has to do withhow the prompt was fed in, how
the model worked.
Now, again, I don't know thelevel of data science expertise

(24:17):
that these people had.
I don't know if they workedwith anyone with data science.
They could do a better job by abunch like multiple shot
prompting examples and probablymore fine tuning Really prompt
tuning would probably been idealfor them.
So there are certainly waysthey could have done it better,
but that actually, in my mind,demonstrates the problem.
Most people believe thistechnology is very approachable,

(24:40):
which it is.
It gets embedded into systemswe use, it gets added onto
platforms we use and that'sreally cool.
But if the vendors don'tunderstand it which I found to
be the case in a lot of cases ifthe users don't understand it,
there is a very real risk.
And again, I think the thing weall need to focus on and worry
about is this sort of AIgovernance and trustworthiness

(25:02):
and transparency.
So, yeah, I thought that mightbe kind of an interesting
exercise to take this what wetalk about here again
specifically around AI modelsand generative AI and turn that
into something in a differentrealm, but that you could see
how it relates right.
The thought of models that youdon't understand and using them

(25:22):
the wrong way and trying to getsomething created that you're
not qualified for to eithercreate or review dangerous in
any walk of your business,whether you're in managed
services or whether you are amanaged service provider
listening to this and dealingwith your clients.
So I thought that'd be kind offun.
Like I said, there will be avisual companion rather to this

(25:45):
podcast that we'll post on ourCrush Bank blog.
So wwwcrushbankcom forwardslash news and we'll put the
link in the description of thispodcast as well.
So check that out.
Laugh at the pictures of me.
I am not going to leave them upthere for very long because
they will be used against me ina multitude of ways.
Anyway, thank you so much fortaking some time to listen to

(26:08):
this episode of the Crush BankAI for MSPs podcast.
Love to hear some feedback.
If you have any thoughts,questions, comments.
You want to challenge anythingI say?
You want to ask me to clarifysomething?
You have ideas for futureepisodes?
Hey, you want to come on andchat?
I welcome it.
I'd love to have it.
I set up an email address justfor this podcast.

(26:30):
It's just podcast atcrushbankcom.
Shoot your feedback there Again, like I said, absolutely love
to hear from you and thanksagain.
So much for tuning in.
We'll talk to you next time,music.