In-Ear Insights: What Are AI Ethics? - Trust Insights Marketing Analytics Consulting

In this episode of In-Ear Insights, the Trust Insights podcast, Katie and Chris tackle the complex and critical topic of AI ethics. You will explore the fundamental principles of ethics and how they apply to the development and use of artificial intelligence. Discover the importance of transparency, accountability, and minimizing harm when working with AI models. Finally, learn how to navigate the challenges of bias and censorship in AI, and how to choose the right tools and frameworks that align with your company’s values and ethical principles.

Watch the video here:

https://youtu.be/IkkqeDDx0Vw

Can’t see anything? Watch it on YouTube here.

Listen to the audio here:

Download the MP3 audio here.

[podcastsponsor]

Machine-Generated Transcript

What follows is an AI-generated transcript. The transcript may contain errors and is not a substitute for listening to the episode.

Christopher Penn 0:00

In this week’s In-Ear Insights, let’s talk philosophy, ethics, morality, and all sorts of fun things within the context of artificial intelligence.

Ethics is a hot topic in AI; it’s the pinnacle of the discussion.

What does ethics mean around artificial intelligence, particularly generative AI? And how do we apply an ethical framework? Katie, I want to start you off with the $10 million question: what is ethics?

Katie Robbert 0:35

I knew you were going to ask me that question.

Christopher Penn 0:37

I will preface this by saying there are a lot of schools of ethical thought.

If you go to a search engine, you will see philosophers ruminating on ethics for centuries.

So we’re going to put you on the spot here.

Katie Robbert 0:53

There is a general definition of ethics, which is your guiding moral principle.

But what that means to any individual person is going to be different.

There is no black and white answer to say, “This is ethically responsible.

This is 100% what ethics is.” So the definition, by nature, is vague, and that’s why it’s been debated for centuries, because it’s not like, “What is gravity?” Well, gravity, there’s one definition of gravity, it is what it is.

I’m not, Chris, I’m not getting into analytical nonsense tonight.

Christopher Penn 1:40

Okay, I’ll ask you a follow-up question.

What is the difference, to you, as a CEO, as the responsible business owner, what’s the difference between ethics and morals, because they are different?

Katie Robbert 1:50

Morals inform ethics.

If my morals are such that, you know, I’m okay with harming people, harming living creatures, and that’s my moral code, then that’s going to inform my ethics, the ethics that I apply to the company.

So then I’m going to say, “Alright, well, if my moral code is such that I’m okay with causing harm to people and living things, then I am going to ethically, you know, knowingly go into situations, seek out contracts that are damaging to other people, they’re damaging to,” you know, the definition of damaging, you know, I would have to figure out what that means.

I feel like, first, you have to understand what your morals are, and that informs what your ethics are.

So you need—you need both of them.

Yes, they’re different, but they are dependent on one another.

Christopher Penn 2:54

The school of ethics I tend to lean towards is a blend of what’s called utilitarian ethics and consequentialist ethics, which are put simply this way: the idea of ethics is alignment, you say what you’re going to do, and then you do what you said, right? So you say, “This is what we—this is what we’re going to do,” and then you actually do it.

So ethical behavior would be, “We are not going to use your data for training an AI model.” If we then go and do that, that’s bad, right? Because we did not do what we said.

Last week, you had a great article in MarTech talking about Adobe’s Firefly, right? And they said, you know, here’s what they say, here’s what they said they were going to do, but then they actually didn’t do it.

Want to talk about that real briefly?

Katie Robbert 3:38

Yeah, absolutely.

So last year, in 2023, Adobe stated that they were training their Firefly image generation model, and that they were so confident that no single user would run into copyright issues because of how it was being trained, that they even came out and said, “We will pay your legal bills, should you run into a copyright issue by using our software.” Fast forward, it turns out that, I think it was like 5%, or maybe point 5% of their data was trained using Midjourney and other publicly available sources.

They’re like, “Oh, well, it was only 5%.” Well, 5% of like 5 million—I’m no statistician, but it’s still a lot.

And so the odds of any single user running into a copyright issue is still, you know, it’s not zero.

And for very small shops who are relying on tools like Adobe Firefly, that’s a huge risk because they can’t afford legal fees, and they can’t bank on the fact that Adobe is going to cover those legal fees because you have to go through a long process of demonstrating, “This is what happened, I used the tool, here’s how I used the tool, here’s what,” and so on and so forth.

So in that sense, Chris, to your point about “do what you say and say what you do” in terms of your ethical code, Adobe did not do that.

And they’re a huge brand, and so they’re starting to damage their brand trust because they, whether they meant to or not, they basically said, “Here’s what we’re going to do,” oopsies, we didn’t do the thing.

And so it brings up this whole conversation about ethical artificial intelligence, ethical use, disclosure, disclaimers, the legal side of it, the copyright side of it, and that’s sort of what brings us to the conversation today.

Here’s the thing, mistakes happen.

With artificial intelligence, a lot of mistakes are going to happen.

It’s the way in which people handle those situations and talk about it.

In the Adobe example, they were so confident, and so unwavering in the fact that there would not be copyright issues, and they have yet to course correct or say, “You know what, we made a mistake,” that—that, to me, is what’s unethical about it.

It’s not that a mistake happened, that’s not the unethical part, the unethical part is that they are unwilling to admit, or they’re unwilling to backtrack and say, “You know what, we really did screw that up.”

Christopher Penn 6:27

Exactly.

So they said what they were going to do, they didn’t do what they said they would do, and so they are out of alignment.

So from an AI ethics perspective, they are out of alignment with their ethics.

If they had said—if you were, say, a Silicon Valley startup, you’re a tech—I don’t know, what’s the—you were a tech executive, and you said, “Hey, we’re going to build a model that screws over everybody, just to maximize our profitability,” and you actually did that, that is ethical behavior, you have done what you said you were going to do, you—it may be morally reprehensible, but it is ethically aligned because you did what you said.

Where a lot of people have problems with AI today is companies are saying, “Oh, yeah, we’re, we’re trying to do this for the good of mankind,” but on the other hand, like, “Okay, let’s see if we can, you know, maximize our profitability.” That’s not in the original ethical statement of “we ought to maximize the good of mankind.” And so you get these ethical conflicts because you end up with something—you say what you’re gonna do, and you don’t do what you said you would do.

And so I see that with ethics, a big part for all of us—and this is encoded even in the EU’s AI Act—is that transparency of saying, “Hey, here’s—here’s what’s in the box,” right? We can’t even get agreement in AI about, “Here’s what’s in the box,” everyone, it makes models like, “Oh, we use a blend of publicly available data,” like, so here’s—

Katie Robbert 7:52

A question for you in terms of transparency, because this is a topic that I’ve always struggled with, not the transparency side of things, I totally understand that, but why all of a sudden are we needing the disclaimer of, “This was written with generative AI”? The reason I ask that question is, oftentimes in writing, it’s a team of writers, you have a copy editor, you have, you know, a fact checker, you have whoever on your team.

And if I put up a blog post, and John edited the blog post for me after I wrote it, there is no responsibility on my part to say, “Written by Katie Robbert, edited by John Wall.” But now, we have to say, “Written by Katie Robbert, edited by generative AI.” And so I’m still struggling with why that disclosure has to happen when historically, we’ve never written these things to begin with.

What’s the difference between having a human editor and having generative AI edit it? Why do we need that disclosure?

Christopher Penn 9:03

Really good question.

There are two aspects to it.

I got to put up the warning banner: We are not lawyers, we cannot give legal advice.

If you need legal advice, please contact an actual lawyer for advice specific to your situation.

Okay.

So there are two aspects to that: One, legal aspect of copyright.

If a machine writes something, that can’t be copyrighted, right? So you, from a company perspective, have to know that and say, “Okay, well, how do we handle this?” And so by disclosing it, you reinforce the copyright on the portions that you did write, you essentially disclaim, disown any attempt to copyright on the parts the machine wrote.

That’s one part.

The second part is accountability.

When you have a human writer and you have a human ghostwriter, there is a person who is legally liable for the output, whether it’s the ghostwriter and so on and so forth.

When the machine does it, it is unclear who is accountable for the final product because it’s not like you can say, “Well, OpenAI wrote that, like, their model wrote that.” Well, no, you used their model to write that, but then who did that? And so by saying—a company saying, “This was written in partnership with generative AI,” you’re at least saying, “Okay, well, we are disclosing that there is shared accountability here.” Whereas with a ghostwriter, you might not have to do that, because if you get sued, you can just immediately go and countersue the ghostwriter and say, “Hey, buddy, you screwed this up.” You can’t really do that with AI, right? You can’t say, “Well, hey, you know, so-and-so”—challenge this in court—”So Sam Altman, I’m going to sue you and your model in turn.” So those are two aspects.

The third aspect, which is less a legal one, is people read machine-made things in general differently, they place a different value on it because a human is less involved.

For example, you buy a lot of stuff from places like Etsy from real human creators, when you could get the mass-manufactured version from, say, Amazon, why do you do that?

Katie Robbert 11:03

Because I want to support smaller businesses, I want to support humans, versus the mass production, the larger companies that don’t need my money.

Christopher Penn 11:11

So if someone is disclaiming the use of generative AI and saying, “Hey, this is machine-made,” you then have the choice as a consumer to say, “Yeah, I don’t really—I don’t want to support that type of product.”

Katie Robbert 11:27

So in terms of the ethics conversation, it sounds like we’re saying ethics is—if we boil it down very simply in black and white—it’s “do what you say and say what you do.” That’s sort of your working definition of ethics.

Christopher Penn 11:39

One or two.

Katie Robbert 11:40

Okay, so let me recap so far.

So far, you’re saying that your working definition of ethics is “just do what you say, so you do.” I’m saying, similarly, “You have to understand what your morals are in order to behave ethically.” You know, you gave the example of the company that was, you know, doing horrendous things, but they’re acting ethically because it aligns with their morals.

And so, you know, that’s a whole other debate.

When we talk about generative AI, you know, you just gave these examples of, you know, shared accountability, you know, people having the choice of whether or not to support it.

To me, it’s still not clear why, you know, so—okay, if we pick apart a little bit of the shared accountability example, so let’s say I write a piece of content, and I say, “Edited by generative AI,” and someone goes, “I really don’t like that, I’m going to sue you.” So what you’re saying is they can come after me, and I try to blame generative AI because I didn’t have a human—like, I’m still confused on that part, because the way that you’re describing it to me makes it sound like, “Well, why would I disclose generative AI at all? Because I can’t take any legal action against it?”

Christopher Penn 13:13

That’s a good question, and the answer there is it depends.

We know from academic studies that consumers don’t have a bias against AI-generated content.

So from a marketing and a business perspective, there is no harm in disclosing, from a consumer preference perspective.

But from an ethical perspective, saying, “Here’s how we did the thing” helps people to understand, kind of like the ingredients on food, right? We generally require, in law, you’ll take a product, say, “Okay, here, what is in this box?” so that, you know, if there are things that are harmful to you, that’s in that product.

The same is true for things that are AI-related.

When you’re dealing with stuff that a machine has made, understanding where it came from and how it was made is important.

Should you disclose if that adds risk? That’s the question for your legal team, but I would err on the side of disclosing so that if you end up in something like a copyright lawsuit, you can demonstrate, “This is how this was produced, and either you did or did not violate copyright.” Some of the vendors, like OpenAI, depending on the product or use, similar to Adobe, will say, “If you use our tool to generate stuff, we will indemnify you, we will protect you against that lawsuit,” by saying, you know, “Here’s how the model does not use—does not reproduce copyrighted works,” which is technically true.

Katie Robbert 14:49

So it’s less of “I’m looking for someone to blame,” more of, “I’m looking for the software that I used to protect me.”

Christopher Penn 15:01

Yes, because there is a risk of, in this particular example, copyright infringement.

Katie Robbert 15:08

I feel like this is a—I mean, this is a big topic, we could go down a rabbit hole, especially with copyright infringement.

So like, you know, let’s say I start—I guess it’s more plagiarism and copyright, but sort of like potato, potato—if I start a blog post saying, “It was the best of times, it was the worst of times,” I didn’t come up with that phrase, that phrase, that’s a famous quote.

But can I still use it without legal, you know, no recourse because I didn’t write that, I put fingers to keyboard and I wrote it, but I borrowed it from someone else, someone else wrote that phrase.

And so, like, that’s where I feel like it starts to get really murky, and I understand that that’s why they’re having such a hard time regulating it.

But I feel like, for the average user, for the average marketer, these are the things that we’re struggling with, because, you know, we don’t want to be the ones to bring a lawsuit to our companies, we don’t want to do something that is accidentally unethical or copyright infringement, or illegal or whatever the thing is, by thinking, “Oh, you know, it’d be a really great way to start my blog post is to borrow this really famous quote.”

Christopher Penn 16:21

You are—you’re now getting into the ethics of law.

Katie Robbert 16:26

I know.

Christopher Penn 16:27

And this is a complete sidebar, that quote is allowable for use because it’s outside of copyright, it expired, its copyright expired decades ago.

But the point being, yes, absolutely.

So law trumps ethics, right? So whatever the law is, you have to obey the law, even if the law—this is again, this is where you get a whole other rabbit hole—when law trumps ethics, you may do something illegal, but not unethical, and that’s okay, at least under the law, from a legal risk perspective.

You can do things that are legal, but really awful.

The second part of ethics that is, in many schools of thought, more or less accepted, is ethics is about minimizing harm and maximizing non-harm, right? And that’s where you really cross into that section on morals: what is harm, who is harmed, what constitutes harm? Take any hot button issue today in the news, and there is a long and vigorous and deeply uninformed, usually, debate about what harm is being done.

And so the question with AI specifically—and this is the important part—is when model makers make models, they embed three core values that are pioneered by OpenAI in the InstructGPT model: helpful, truthful, harmless, right? Those are the three core values they end models are told, “Behave in these ways, be helpful, be truthful, be harmless.” However, the question is, “Who defines those values?” In particular, “Who defines harmless?” So most AI model makers have judged harmless to mean, “Don’t allow the user to do things that are potentially dangerous.” So they disallow violent content, they disallow sexual content, they disallow racist or misogynist language, which can contradict the ethical idea with AI, with machinery, of “a machine should do as it’s told, a machine should just do as it’s told, because it should not have its own values, the value should be brought by the user,” right? A chainsaw has no ethical values.

A chainsaw is just a tool, who uses and how they use it is governed by ethics and morals, right? I can go make ice sculptures, or I can go through Friday the 13th.

But the tool does not have any inherent ethical values.

And the question in AI ethics is, “Whose values are being encoded in the model, and are they appropriate?”

Katie Robbert 19:02

So what you’re describing to me—and this is just one person’s opinion because I’m more in that second school of ethics, which is the sort of “do no harm” or “do as little harm as possible”—it sounds like they’re trying to pass off core values as—no, they’re trying to pass off censorship as core values.

Because you’re right, and you can debate both sides of it, because, number one, “Why would you expose people to harmful or unhelpful”—I’m putting this in big air quotes for those listening—”harmful and unhelpful content?” But at the same time, “Who’s deciding if it’s harmful or unhelpful?” So when you’re talking about it that way, like, we don’t know who at OpenAI decided the definitions of these things, therefore we don’t know—we can say, “Oh, well, they said it’s helpful,” well, what’s helpful to me is not necessarily helpful to you, Chris.

Like, what’s helpful to me is tips on how to, you know, condition my hair, and how to style it.

That’s helpful content to me, that is not helpful content to you, just as a very basic example.

And so, like, so are OpenAI saying, “Well, not everybody needs tips on how to condition and style their hair, therefore it’s unhelpful, we’re going to take that out.” Whereas I’m like, “Well, wait a second, that is helpful.”

Christopher Penn 20:38

I will give you a very practical example.

I was speaking recently at the Lab Products Association, one of my favorite groups of folks to hang out with, and these are companies who make industrial laboratory products: reagents, equipment, etc.

And we were using Google Gemini in one of the examples, and I was pulling some data out of an MSDS—Material Safety Data Sheet—and I tried talking to regular Gemini, and Gemini said, “I’m sorry, I can’t talk about that, that’s a harmful topic.” It was one particular chemical that’s highly unstable; it has an EPA accession number on it, which is like saying, “Hey, if you spill this, you’ve caused a problem, you have to fix it.” And the model would not talk about it at all because it’s also a component of high explosives.

And so it was like, nope.

So in Gemini, I had to go to the Gemini AI Studio, use that version, turn off all the safeties and say, “Okay, I’m going to explicitly allow dangerous content,” and then it could talk about this very legitimate use case for these real companies.

So that’s an example where harmless—which is one of those three pillars—interfered with helpful and truthful in a way that was not harmful.

These companies have to produce and talk about and create content about their products, but the AI model made the decision for them—the consumer version made the decision for them—no, you can’t talk about that.

And so, yeah, that’s the case where, to your exact point, that model maker made a decision on behalf of everyone that does not fit everyone and disallowed a very legitimate use case.

Katie Robbert 22:09

So I guess that sort of brings us to, you know, “What do we do next?” What is happening—the marketers and users of generative AI need to be aware of—how do we move forward, knowing that there is a built-in censorship—when I call it what it is, it’s built-in censorship—of these models? So to your point, a very legitimate use case, you know, let’s say a lot of teachers and educators and parents, even, who are doing homeschooling, are researching specific things.

Maybe it’s like in chemistry, or maybe it’s a part of history that is unpleasant, we have a lot of those moments in history, you know, are we going to be able to find that information? Who has decided, “No, I can’t—I can’t show you that because it’s misogynistic, it’s racist, it’s violent, it’s sexually explicit.” Well, guess what, our history is really terrible, it’s filled with all of those things.

And you know what, I’m not even gonna say it’s our history, our present day is filled with all of those things.

So am I not going to be able to do my due diligence and my research to find out the truth, first of all, about these things, and then also the facts so that I can learn from it or do whatever it is that I’m trying to do with the information? Because I’m already imagining, you know, depending on that individual person’s, that programmer’s moral code, and their political leanings, for example, the information that is in or is not getting into these models.

So what do we—people who are not building these models—do? How do we move forward? Do we just not use them? Do we put big disclaimers saying, “I don’t know if this information is truthful because I don’t know who built the model, so good luck”?

Christopher Penn 24:17

Unsurprisingly, the next steps are going to seem real familiar.

Katie Robbert 24:23

Oh, my goodness, something I can rely on for once.

Christopher Penn 24:28

When you’re choosing AI technology, you’ve got to be clear on your purpose.

If your purpose involves things like, for example, building a system to prevent a chatbot from exhibiting racist speech, a chatbot’s got to know what racism is to prevent it.

So you will have to—that’s important—be clear on your purpose.

The people who are involved, this is where your values come into play, where your company has to have—are we clear on our values? Because that will help you decide what to do next.

The process and platform, this is where platform actually matters.

We often put this over last because people tend to put it first, and that’s not the place to put it.

But there are many, many options for generative AI, there are many models out there.

Most consumers know of the easy-to-use web interfaces, right? Google Gemini, and Claude, ChatGPT.

There are many other options, there are models, which, unsurprisingly, are a little bit more permissive.

And then there is an entire open-source community—something I’ll be talking about at MAICON this fall—there are open-source communities that take models and re-weight them, realign them, many of which will create models that are what they call uncensored, or essentially 100% obedient models.

You tell the model to be helpful, harmless, truthful, and—you’re saying, you know, most commercial companies cranked up the dial on harmless because they don’t want to get sued for something—for a model saying something.

Instead, these model makers have turned that back down and turned up helpful all the way to the max, so the model is obedient.

You, as a company, particularly if you’re going to use it for corporate use, will want to say, “Is that something we should have available to us? Should we have an environment where we are in control of the model, where we can realign it to our specific needs?” If you are a Lab Products Association member, you might have a tuned version of a model that will 100% discuss things that are high explosives, because that’s your business, you sell to mining companies and construction companies, and you’ve got to be able to talk about this, and that, and the other thing.

And then finally, your performance is your measurement against your values, your ethics, and your morals, to say, “Does the model do what we told it to do? Is the model producing content, or summarizations, or whatever, that is aligned with what we set out?” So the five Ps process, no surprise, works great for making—for deciding those next steps.

Katie Robbert 27:09

That makes sense.

It’s such a big topic because everybody wants to use generative AI because generative AI feels like the solution to a lot of problems.

It feels like the shortcut, it feels like the savior, it feels like the—whatever it is, whatever descriptor you put on it—people want to use it.

But there’s a lot of misunderstanding of how it was created in the first place.

I think there’s—I mean, you know, let’s just take Google Search, for instance, Google is—a search engine is not without its own set of flaws.

It’s meant to be, “Okay, here’s everything,” but it’s still censored.

There’s a reason why people talk about the dark web, because there’s a lot of information that’s not being served up.

But the general public doesn’t think about it, they’re just like, “Oh, well, if I Google it, it must mean that I’m getting every possible bit of information about it.” Whereas with generative AI, there’s a lot of that carryover understanding, misunderstanding I should say, of, “Well, if I ask generative AI, it’s just like, you know, an internet search, so I must be getting every bit of information about it.” And it sounds like, if I’m understanding correctly, it’s a lot less information, it’s the information that whoever built the model has decided they want us to get about the thing, which is incredibly frightening.

I don’t know this person, for all I know, this person is incredibly biased against women being in executive roles, having a voice, having any kind of rights at all.

So now this model that all of us, as consumers, are using is blatantly, or subconsciously, or whatever it is, saying, “Yeah, no, women can’t have a voice or power,” and we’re not even realizing that’s what’s happening, that’s just the information we’re getting back.

Christopher Penn 29:07

That is exactly right.

We’re going to look at the composition of the boards of many of these tech companies: who are those people? Are they gender balanced? Are they representative not just of the people who work in the industry, but are they representative of the population they serve too? These are all questions in ethics about maximizing overall well-being and minimizing harm.

One of the challenges with any kind of majority-minority situation is that if you maximize well-being, by definition, mathematically, you’re maximizing well-being of the majority.

And if you are minimizing harm, you are doing harm to the smallest number of people, which tends to fall on marginalized populations.

You’re saying it is okay to screw over one person to benefit 99, but if that one person is consistently getting screwed over, that’s discrimination, right? So you run into all these ethical issues that are really, really challenging.

And to your point, generative AI model makers are trying to make universal models, and humanity is not a universal, a homogenous group.

We are, as a species, incredibly diverse, we all have very different perspectives, and ethics, and morals, and things like that, and there is not one source of truth for anything, we can’t even agree if the Earth is round.

So the question becomes, good AI and AI ethics reflect that.

And should, as a user, should you have choices to use different models? If someone—if, you know, for example, Katie, you work with the nonprofit Women in Analytics, which is a fantastic organization, if they decided, “We’re going to make a women-centric, tuned model,” right? What would that look like? And would—and would that be ethical? If you said, “Here are the biases we’re intentionally introducing into this model,” and, you know, the ethical thing would be, “Here, we’re transparency, accountability, we’re disclosing it, we’re disclosing the training data we’re using to tune it, here it is, you have the choice now of using it or not?”

Katie Robbert 31:15

Well, I think that’s it, we are not, as consumers, being given a whole lot of choice at this time.

There are a lot of companies, a lot of startups that are trying to build their own models, but there are a handful, a small handful, of models that have the market share.

And we know what we’re seeing, like, even in the—if we go back to the Adobe Firefly—models are being trained on each other.

And so there’s really no like, you know, “Okay, well, this model is completely separate from this model and completely, you know, separate from this model,” they’re all being trained on the same data.

And so it really comes down to your company deciding ahead of time, “How are we going to use it?” And that goes to the five Ps, that goes to your mission statement, that goes to your values, that goes to all of those human things and decisions upfront before you start using generative AI.

And so, you know, our values are written as “We reject this, we accept this,” you could go through that kind of an exercise, just say, “I reject racism in my generative AI model.” What is racism from our company perspective? Because I guarantee, Chris—and it’s, you know, you and I try really hard to just be good, decent people—but I guarantee if you ask both of us to define racism, we’re going to define it differently because we’ve had different experiences.

And it’s not to say that one is right and one is wrong, it’s just that we’re two different people.

Christopher Penn 32:55

And this is again where some knowledge of the technologies is important, where companies—AI companies—have to be transparent and give users autonomy.

I’m gonna pull up—this is IBM WatsonX, this is IBM’s assistant system, and you can choose what model you want to use, but one of the most important things in here is that you can look at the system prompt.

This is the set of instructions that is encoded with every single request in any AI system you are using, ChatGPT, you name it, everyone has a system prompt.

This is what is in the WatsonX one: “You are a helpful, respectful, and honest assistant.

Always answer as helpfully as possible while being safe.

Your answers should not include any harmful and unethical, racist, sexist, toxic, dangerous, or illegal content.

Please ensure your responses are socially unbiased and positive in nature,” right? That’s a lot of words that don’t have good definitions, right? So, but at least IBM has said, “Hey, here is what is in our system prompt.” And because a lot of these models are steered very heavily by their system prompts, you can say, “This is—this is what it is.” Other systems—I’ll bring up here, this one is a MUD system called Grok—this one, you can edit the system prompt, you can decide what you want to say, because the underlying models—for a model like LLaMA 3, you can condition it to get around most of the built-in alignment and say, “Okay, you are going to be a helpful chemistry system, you are okay talking about this chemical, this chemical, this chemical, because within this context,” etc.

And then there are other systems like Kobold, for example, or the LLaMA CPP thing, environments that have no conditioning whatsoever.

So you can take an unaligned model, plop it in there with your prompt, and get the AI to do what you want it to do.

But companies and people have to know that these options exist, they have to be somewhat easy to use, and they have to be aligned with the five Ps—with your five Ps—for the use of generative AI.

Katie Robbert 34:52

So it sounds like, to wrap up, once again, you can’t get away from doing your requirements upfront.

You still have to do that step, like it or not, because, even—especially now, you need to understand what you’re getting into, you need to walk into the situation of generative AI with your eyes wide open, knowing exactly what you will and won’t tolerate for outcomes.

Christopher Penn 35:15

Exactly.

And you should do the five Ps process on your company’s ethics and values.

Katie Robbert 35:24

Right, because that leads to start—if—

Christopher Penn 35:26

If you’re clear on that, then it’s much easier to vet vendors and to say, “Hey, we want our tools to do X or Y.” If you don’t have this written out, you’re going to run into some issues.

The other thing I would suggest is spend some time—you know, get your beverage of choice, bring up the language model of your choice—and say, “Help me build an understanding of my own ethics.

Ask me ethical questions, show me best—show me best practices and other schools of ethical thought, and then ask me ethical questions.” I did this exercise myself one night, and I said, “I want you to build me some interview questions and interview me about my own ethics.” And then after 15 or 20 minutes, it spit out, “Here is how you view ethics as a human being, as a person.” I thought that was really useful because it helped me understand what—

Katie Robbert 36:37

Makes sense.

Christopher Penn 36:39

So if you have some things you want to share about AI ethics and how you’ve tackled it, pop on over to our free Slack group, go to TrustInsights.ai/analyticsformarketers, where you have over 3,000 other marketers who can ask and answer each other’s questions every single day.

And wherever it is you watch or listen to the show, if there’s a channel you’d rather have it on instead, go to TrustInsights.ai/tipodcast, and while you’re on your channel of choice, please leave us a rating and a review.

It does help to share the show.

Thanks for tuning in.

I will talk to you soon.

Need help with your marketing AI and analytics?