In this episode of In-Ear Insights, Katie and Chris tackled the thorny ethics of what you put in your AI models, based on leaked memos from Tiktok allegedly discriminating against protected classes. How do we know when an AI model is behaving badly by accident and by design? What are the steps we should take to prevent or mitigate this sort of thing in AI? Listen now to hear an in-depth explanation of AI explainability in layperson’s terms and the liability and risks you might be ignoring.
In Ear Insights is brought to you by Sponsored, a self-serve podcast marketplace that connects marketers with podcasts to reach technical audiences. Instead of hiring someone for ~$60k/year to manage and scale your podcast sponsorship channel, Sponsored does all of the heavy lifting for you to place the sponsorships, track the performance, and manage your budget. For every sponsorship booked, Sponsored donates a large portion directly to human-I-T, a non-profit organization based in the U.S. that provides low-income individuals and nonprofits with access to technology, the internet, and digital training. Go to sponsored.us/inearinsights to receive $500 off your first podcast sponsorship placement.
Watch the video here:
Can’t see anything? Watch it on YouTube here.
Listen to the audio here:
- Need help with your company’s data and analytics? Let us know!
- Join our free Slack group for marketers interested in analytics!
What follows is an AI-generated transcript. The transcript may contain errors and is not a substitute for listening to the episode.
Christopher Penn 0:02
This is In-Ear Insights, the Trust Insights podcast.
In this week’s In-Ear Insights, we’re talking about interpretability and explainability, artificial intelligence and the way it’s being misused.
So over our friends at the content marketing world Slack, brought up his whole Twitter thread from adore me, which I assume, based on the context is a lingerie company.
And they took some time on their own Twitter feed to point out that they were being discriminated against on tik tok.
And the reason for it was because of things that are essentially, you know, certainly not ethical, and possibly illegal depending on whether or not protected classes are involved, specifically Tiktok circulated internal memos saying these are the kinds of content that we want to remove from our platform, they have quality raid, like rating guidelines.
And in those guidelines, they talk about things like abnormal body shape, chubby has obvious beerbelly obese to thin, ugly facial looks, people with scars, missing teeth, facial deformities, dilapidated housing, interior backgrounds.
So it’s a very, you know, some of the stuff is definitely falls in the category of protected classes either based on race, based on disability, things like that.
And the what’s happening behind the scenes is, is being built into their algorithm in the same way that Google takes its search quality rate, like ratings, and uses that to train their AI have to say this is a good search result, this is a bad search result.
Tiktok is using the same these same quality guidelines to train their AI as to what content needs to come down.
And one of the things that dormy pointed out was that almost all their content that features someone with darker skin gets taken down.
Whereas people with lighter skin does not obviously this is a case where when you’re looking at the output of an AI model is a very serious problem.
So in the content marketing world, slack had asked, this makes me wonder about the future of AI based technology who is liable? When algorithm is based? Is it the company selling the technology, the engineering team, the brands who buy the AI based technology? So Katie, in your perspective, particularly an outsider who doesn’t use Tiktok, and things and you see, the results of an algorithm has clearly gone off the rails by design, like this, is the this is contaminated data going in? Who is responsible here? Who’s Who is responsible? And from your perspective, as an executive, as a leader? What are the solutions that you see?
Katie Robbert 2:41
Well, let me back up for a second.
And let’s just focus on that leaked internal memo.
Now, I’m looking at this on Twitter.
And, you know, a dormy is saying that this is what Tiktok is, has put out as a memo.
Now, when you look at the screenshot, there is nothing on this, to identify that that’s what Tiktok had actually put out as memo.
There’s no watermarks.
There’s no, you know, company brands, and I’m not defending Tiktok to say that they didn’t say those things.
The thing that, you know, that sort of strikes me first is the terminology used abnormal body shape, chubby, have obvious beerbelly You know, a lot of these things are completely subjective, you know, so when you say, I think this person is ugly, someone else might say, this is the most beautiful person I’ve ever seen.
And so all of these things that they’re trying to ban from their, you know, algorithm are completely subjective.
You know, it’s not a black and white thing to say, you know, we will not allow, you know, violence or, you know, you know, list a whole bunch of other things like, so that’s the first problem.
I mean, there’s so much to unpack here.
The when you’re asking who is at fault, who’s accountable, it’s whoever wrote this memo, and whoever decided, yes, this is a good idea.
You know, and so what adoring me is saying is that they’ve worked with the other major social media platforms, Facebook, Twitter, and that their content has not been a problem before.
And so you would imagine that Tiktok being one of those, they’re still newer, you know, in terms of being a social media platform, trying to gain all of those followers and sort of take take on the market, but they would start playing the same game.
In terms of Okay, what is Facebook doing? What is Twitter doing Now, granted, to say, what is Facebook doing? Let’s follow that model.
There’s a whole other dumpster fire to unpack but in terms of this particular context, it’s what you know, what are the other major social networks do That are right and or wrong that we can learn from it.
And I don’t see the other major social networks using their algorithms to discriminate in the same way.
And that’s, that’s really the crux of it.
So, you know, I don’t use Tiktok, I don’t use a lot of social media, that doesn’t change the fact that whoever came up with this model, whoever came up with this algorithm, it’s blatantly wrong.
Now, if Tiktok is going to come out and say, No, this is who we are, this is who this is what we stand for.
We don’t want these things on our platform.
Great, go ahead and say that publicly, and then people can make a decision, whether or not they want to use the platform.
But if they’re trying to pretend that it’s the algorithms fault, they never, you know, they’re like, Oh, no, we just sort of made suggestions.
It goes back to Chris, what you want to talk about is people are at the core of developing these algorithms.
So therefore, it’s not the algorithms fault.
That’s like saying, like, my soup came out too salty.
It’s the soups fault? Well, who put the salt in the soup in the first place? It was you dummy.
Christopher Penn 6:10
Oh, that’s absolutely right.
And we’re hitting on here is that at the end of the day, artificial intelligence is nothing more than software, right? It’s no different than your word processor.
It’s no different than you spreadsheet.
It is not the spreadsheets fault.
When the math comes out wrong, right.
It’s, it’s what the user put into the spreadsheet that that makes it come out wrong.
And so the the question of liability, I think it’s an important one, because yeah, if this if these things are true, and they are alleged, because this has not been a court case court finding it, then you have, you have really serious problems, it’s kind of like, it’s kind of like a car, right? If if somebody goes out and gets into an accident in the car, right, it’s not the car manufacturers fault, unless they clearly have a manufacturing defect of some kind.
But if you get drunk and go out and crash your car, and it’s your fault is the operator.
So in the case of liability, it’s not necessarily the maker of the algorithm, if it’s a third party piece of software, it’s the user.
And so in this case, it would be Tiktok is specifically around what training data they fed it because what these quality grading guidelines are, is they’re developing a training library, these are the things that we want to see more of on our platform.
And this gets to the heart of the debate between interpretability explainability, what adoring me to hear is is doing is is pointing out an explainability problem.
They see the output of the algorithm.
They say, this is what we see, and this is what’s happening here.
It’s not happening there.
And there’s a very clear difference between the two.
But what nobody has except Tiktok, and would need to come out and in a court of law is the interpretability of the Okay, open up the software, open up the model, open up the training data, and show us prove to us that the model is not discriminating that these that you know, these allegations are false.
And this is a situation where I think not enough companies are giving thought as to how they’re using their technology, and particularly how they’re using their data.
And whether it has intentional bias or use, as shown here or has accidental ones.
Katie Robbert 8:26
The other example for Tiktok is there’s a singer named Liz Oh, she’s very popular right now she plays the flute.
And she happens to be a black woman who’s on the heavier side.
But she’s someone who embraces it.
And she’s, you know, come under a lot of criticism of, well, no, you can’t possibly be in good shape.
In order to do these performances and concerts, you have to lose weight.
And she’s the first person to say, watch me.
So she’ll get on a treadmill run and sing an entire set of songs just to demonstrate.
And it’s it’s impressive to see, by the way.
But she’s also someone who’s criticized Tiktok because Tiktok continues to remove her content, have her in a bathing suit.
But yet they keep other people in bathing suits up.
And so there is definitely it’s not just a dog, it’s other people on the platform experiencing the same kind of discrimination.
And so, you know, you’re right.
I don’t think that these companies are thinking sort of that long term.
You know, what do we do with this technology? What do we want it to do? I think that there’s a lot of let me rush to be the first out of the gate in the market.
Let me be the new shiny thing that people sign up for.
Because what’s the next thing what clubhouse now like, I know that it’s audio, but there’s gonna be some, you know, algorithms and biases probably built within there too, because it’s just shoved right out the door.
People like I want to be the next big thing.
I want to be the shiny thing.
I want to be the startup that makes a bazillion dollars without that long term thinking.
And I think that bias in technology is still something that’s very not well understood, because there is an assumption that the algorithm knows what the heck it’s doing.
It knows what you want it to do.
Christopher Penn 10:24
And as customers, we don’t get access to the interpretability side, we don’t get to do code inspection.
Unless, I guess maybe for a select few vendors, if you’re a big enough spender on the platform, you could request that as part of your compliance.
But for for those of us who would be average, folks, we don’t get that.
Katie Robbert 10:41
And if Tiktok said to me, do you want to inspect my code? I probably go, sure.
I don’t know what I’m looking for.
And that’s part of the challenge as well.
So even if these companies said you can go ahead and inspect my code, most people wouldn’t even know where to start with that.
Christopher Penn 11:00
Right? So what’s up? What should a marketer be paying attention to that if there are platforms where you have allegations of, in this case, intentional bias, then there are obviously platforms where the machines themselves are making interesting and or questionable decisions like Facebook, which has gotten so big that there are no single team of engineers that knows entirely what the whole algorithm is doing? The same is true of Google, what’s a marketer do to do to reduce their liability? To avoid funding things that might be harmful to their customers? What should people be thinking about in terms of their responses?
Katie Robbert 11:43
Oh, well, I think that definitely do your homework.
So if, let’s say, for example, you know, we run paid social ads for our clients, and they say I want to run ads on tik tok, I want to run ads on Snapchat, and Facebook and all the other major players, it is our responsibility as the person representing that client first to do our homework and find out, you know, if these articles have, you know, people complaining against the software exist, you know, have people reported, yes, this is a bias platform, it’s taking things down, you know, randomly because of the way that it’s felt.
And then also, you know, it’s our responsibility to check and see if any of those, you know, developer notes exist, or any sort of version of those in a way that we can understand how the platform is built, just sort of doing our homework, in terms of what that platform not stands for, because it’s a social media platform.
It’s social media.
But really sort of like if we can get sort of any more insider information, in terms of what it does, but also, you know, what kinds of audiences are on those platforms? So am I reaching the right kind of audience? And so, you know, a really terrible example is, let’s say a client came to us and said, You know, I want to reach an audience I don’t normally reach I want to advertise on parlour, we would have to start to question, okay, why do you want to reach that audience? What about that audience Do you think makes them a good fit? And they might be a good fit for their product? Who knows? But it’s really on us to sort of do that due diligence to understand is this where you really, really, really want to be? because now you’re going to start spending money?
Christopher Penn 13:32
Is there is now an ethics question.
Is there room in the marketers toolkit to go onto this on these account platforms, set up dummy accounts, you know, which is misrepresentation, post content, to experiment to test to see like, Hey, I’m going to put up copies of pictures of from, you know, somebody’s account of different people, different shapes, sizes, languages and stuff, and see what gets taken down?
Katie Robbert 14:05
Oh, that’s a tough question.
I mean, like, you’re going into it ready for a fight, like you’re assuming the worst.
Um, you know, I do think that, you know, as a marketer, you should be having legitimate accounts, on a lot of these platforms just to understand what they’re all about, and what the interactions look like, what the engagements look like, what the algorithms tend to do.
So you should definitely have those accounts, whether or not they’re dummy accounts or real accounts, or, you know, test accounts, whatever you want to call them.
Now, if you are, I think, you know, my personal perspective, if you’re going in there, just for the sake of exploiting an algorithm and then reporting on it publicly, I think that that might start to cross that line of those ethics because instead of just sort of Like, you know, what do they call it like that kind of like screw you journalism where it’s like, I’m going to go in there and get the story, you know, and then just forget whatever the company has to say about it, you probably should be reaching out and working with the company to say, Hey, you know, I posted up these couple of things, you took them down.
Can you explain that to me why that happened? And then you can start to do that, you know, investigation explanation, but I think blindsiding, the company without having the full story is a little bit problematic, because then you’re just contributing to, you know, well, this is my opinion, and they took it down, and they just randomly took it down, you don’t know, they might have taken it down for the wrong reasons, but at least give the company an opportunity to see that this is a problem.
So I don’t think it’s a good idea to set up a test account or a dummy account, and then just start like, baiting them into doing the wrong thing.
Okay, I think that’s
Christopher Penn 16:01
what took a moment to thank our sponsor.
In-Ear Insights is brought to you by sponsored a self service podcast marketplace that connects marketers with podcasts to reach technical audiences.
Instead of hiring someone for $60,000 you to manage and scale your podcast channel, sponsored does all the heavy lifting for you to play sponsorships, track performance, and manage your budget.
For every sponsorship book sponsor donates a large portion directly to human it a nonprofit organization based in the US that provides low income individuals that profits with access to knowledge of the internet and digital training, go to sponsors.us slash In-Ear Insights to receive $500 off your first podcast sponsorship placement.
Again, that’s sponsored.us slash In-Ear Insights, and we thank them for their support of the show.
So for companies that are building their own AI, then we do have access to the interpretability.
So we can open up the code and things like that.
Do you feel like anybody’s got sort of an ethics QA process in place? Or is it all just kind of assumed that the coders know what they’re doing?
Katie Robbert 17:07
It’s a good question.
I don’t know the answer to that.
I would like to believe that there’s some sort of an ethics QA process in place.
But my guess my, you know, uninformed assumption is that it’s left up to the developers and then hand it off to the executives and the investors to say, here’s what we built.
You know, you can ask me questions, but I’m going to answer you in such a way that you don’t understand what I’m saying.
Such as the way of developers and engineers, they don’t do it on purpose, sometimes they do.
But this is something Chris, that you and I have talked about in terms of bias and artificial intelligence and some of the job opportunities that would be coming up because of AI.
So the notion of will a, I take my job, maybe sort of, but there’s other opportunities.
And one of those opportunities is that committee to make sure that your algorithm is being built in an ethical way.
And so those committees are those groups are key ways.
It can’t just be a group of your peers, it can’t be that homogenous, like, okay, I am a white woman, therefore, another white woman should be looking at this algorithm that’s problematic.
You need people from different walks of life, different voices, different backgrounds, ethnicities, experiences, skill sets, to be looking at this algorithm to be pointing out, that’s a bias, that’s a problem that works, you know, sort of check all the boxes, you also need to before you even build the thing, figure out what you want it to do.
And if you can sort of write out, you know, we want to bias against this thing.
We don’t want to bias against this thing.
Like be be clear and open upfront with what you want this thing to do.
And don’t just say, I want an algorithm that shows you know, fun puppy and kitty content.
Well, it’s going to probably grow bigger than that, or, you know, even sort of at a micro level, are you biasing against, you know, dogs of a certain color.
So a real example of that is black dogs tend to not get adopted as often or as frequently, because they either look scarier, they’re harder to photograph, you know, list a whole bunch of things that are silly, but that’s a real thing that happens to dogs.
So take that and bring that to humans.
And you can imagine what that’s like.
Christopher Penn 19:37
It’s interesting, because in Silicon Valley, in particular, with the big tech companies, there is a very binary perspective on interpretability versus explainability.
And AI, are you doing one or the other, and a lot of the big tech companies fight very hard against interpretability because it makes your code much more expensive, both to create and to maintain because you have to be able to pick two You know, stop the machine and pull it apart at any given time.
And that is a a substantial technical hurdle.
But what I hear you saying is an interesting take on it, because it’s it’s not binary in the sense of, you should have an a, an explainability, Council of non technical people to look at the output go, Yeah, that’s right.
That’s not right.
And then have a technical group to sit to look into the code and say, these are areas where there’s likely to be problems.
You know, this when this module goes into production is likely to have these issues.
And in at least in the case of Tiktok, here, you might even be able to have not interpretability on the code side, but interpretability on the data side to say like, yeah, the data you’re putting into the model is flawed.
So don’t even bother building the model yet, because you what you’re putting into the machine is wrong.
Katie Robbert 20:50
Yes, that is what I’m saying.
Because I, I feel like if you’re still trying to develop something that’s blackbox and secretive and mysterious, you know, that nobody else will ever figure out, you’re doing it wrong.
There’s so many people who now understand how programming languages and development works, that someone’s going to be able to unravel it and point out these issues that you should have caught upfront.
And you know, just because you’re being transparent about how the thing is built, doesn’t necessarily mean that you’re vulnerable, for someone else stealing the thing they can try.
But then it also sort of brings up like you need, what protections are you putting in place to make sure that, you know, your copywriting and trademarking, etc, etc.
But back to the point is that, you if you have an algorithm, if you’re building an algorithm, you should be able to explain it in a non technical way exactly how it works, so that people can understand it, buy into it.
And you can also get that constructive feedback to say, Have you thought about this particular scenario, this might be biased against it and do all of that work up front.
Companies that are building these big algorithms that aren’t doing that work, It baffles me because you’re going to spend so much money, you know, in your brand reputation, defending yourself, lawsuits, all of these things when you get it wrong, why not spend a little bit more upfront to get it? Mostly right out of the gate, but then also say, you know, what, we’re still learning help us do better.
Christopher Penn 22:36
Yeah, well, it’s same question we ask people, why didn’t you spend money on analytics and measurement upfront?
Katie Robbert 22:44
baffling, what’s, you know, because the the planning process for everyone except me is the most boring process, I think it’s fantastic.
They just want to go ahead and get the thing out the door, you know, plans and organization and timelines and, you know, discovery be damned?
Christopher Penn 23:05
Well, to sum up in the development of any kind of software model, whether it’s done with machine learning or not, bias can creep in all along the way from the people you hire, the strategy you pursue, the technology choices, you make the data you put into the machine, what the machine comes out with, and then ultimately, what you do with it.
And you need to have interpretability, which is understanding the technical, you know, bits and pieces, and explainability, explaining the real world results of what your efforts have created at each of these stages, in order to avoid creating situations where you are being lit up in the press, like a Christmas tree for creating something that is either accidentally or intentionally biased.
And if you work at a company, where they’re creating unintentional biases against protected classes, it is time to update your LinkedIn profile, because that is not a company you want to spend any time working at.
In addition to making the world a worse place, at some point, that company will be subject to a very, very large lawsuit and your job security is not assured.
So if you’ve got questions about anything we’ve talked about, in today’s episode, head on over to Trust insights.ai slash analytics for markets we can chat with over 1500 other fellow marketers about everything from basic questions like hey, how do I figure out what’s what to do on Instagram, to the most complex questions like dissecting neural models? And if you’re wherever it is, you’re watching or listening to this if you head on over to Trust insights.ai slash ti podcast, you can find out how to subscribe to make sure you never miss an episode.
Thanks for watching and listening.
We’ll talk to you soon.
want help solving your company’s data analytics and digital marketing problems? Visit Trust insights.ai today and let us know how we can help you
Need help with your marketing data and analytics?
You might also enjoy:
Get unique data, analysis, and perspectives on analytics, insights, machine learning, marketing, and AI in the weekly Trust Insights newsletter, Data in the Headlights. Subscribe now for free; new issues every Wednesday!
Want to learn more about data, analytics, and insights? Subscribe to In-Ear Insights, the Trust Insights podcast, with new 10-minute or less episodes every week.