Introduction - Kanupriya (KP) Yakhmi
🫛 This edition of AI, Software, and Wetware features an audio interview on ethical AI with Kanupriya (KP) Yakhmi. . KP is a 🇺🇸 USA-based independent product strategy advisor to startups. She is also a women’s mentor in tech at AnitaB.org and authors KP’s Substack. We discuss:
Noticing that most people she talks to are “sitting on a stovetop right now, maybe in a pressure cooker”
Launching Stillpoint based on research into cognitive effects of pink noise, and sharing it freely with the world
Designing Stillpoint to respect people’s privacy, build trust, and avoid the usual pattern of keeping people on the site as long as possible
Avoiding AI in her mentorship work and adapting her approach to each person
Using AI tools to get oriented when doing product research across domains
Options such as synthetic data for addressing gender or age gaps in training data
Voice artists getting paid for licensing their voices for training ML models
Believing that those of us who live in the “tech bubble” are responsible for protecting people who don’t, by creating safe and responsible AI products
Incentivizing organizations to shift accountability for data to the point of aggregation, not the point of collection
Why we should ask every AI organization to state clearly: who owns what the system does to people?
and more. Check it out, and let us know what you think!
This post is part of our AI6P interview series on “AI, Software, and Wetware”. Our guests share their experiences with using AI, and how they feel about AI using their data and works.
This interview is available as an audio recording (embedded here in the post, and later in our AI6P external podcasts). This post includes the full, human-edited transcript. (If it doesn’t fit in your email client, click HERE to read the whole post online.)
Note: In this article series, “AI” means artificial intelligence and spans classical statistical methods, data analytics, machine learning, generative AI, and other non-generative AI. See this Glossary and “AI Fundamentals #01: What is Artificial Intelligence?” for reference.
Interview - Kanupriya (KP) Yakhmi
Karen: I’m delighted to welcome Kanupriya Yakhmi from the USA as my guest today on “AI, Software, and Wetware”. KP, thank you so much for joining me on this interview! Please tell us about yourself, who you are, and what you do.
KP: Yes. My name is Kanupriya, and most people call me KP. I have a very long name. I grew up in India, and I’ve been based in the US for some time now.
My background is in product, in AI systems, healthcare, platform, AV [autonomous vehicles], as well as enterprise software. I like to call it a very diverse experience, which helps me connect deeply with customer pain points as well as cross-functional teams. And across all of it, the pattern that I’ve seen emerging is that the technology is rarely the hard part of solving the problem.
The hard part is the gap between what the system is designed to do and what actually happens when it meets real people. Because in the end, we all are humans, and we can operate from biases all the time. Right now, I work with some early-stage founders building an AI space. I mentor women in tech with AnitaB.org, which is one of the largest women empowerment organizations. I’m doing a PhD in AI, and I recently launched Stillpoint, which is a free cognitive reset anti-pattern tool for people in high-pressure roles. I think the world is going through so much right now. We need compassion for ourselves and for other people, so I’m hoping that people use it, and it’s 100% free. I don’t track anything either.
Karen: That’s great that you respect privacy. I’d like to hear a little bit more about what motivated you to create Stillpoint and how you built it.
KP: Yeah. So what motivated me? I talk to a lot of people, I mentor women, and I interact with different kinds of people. What I’ve come to realize is that whether I’m talking to a client, I’m interviewing, whatever it is, most of the people are emotionally sitting on a stovetop right now, maybe in a pressure cooker. So regardless of someone working at a large or small organization, everyone faces similar emotional challenges.
And once we accomplish a threshold of a certain income, people often stare into the empty abyss which is life. We think that, “Okay, I have everything by standard metrics, but I’m not happy.” So I did some research, and I came to know that there is something called pink noise that helps to regulate our emotions, which is not currently done by any of the products. Heavily research-based. So that’s why I built it.
I used Gemini, ChatGPT, and Claude to build it. I also did market research, by which I mean that it is not just another music product app that already exists. It’s competitive research. Secondly, I introduced product design and product sense into it, which talks about architectural design, software, etc. And thirdly, I took research data from institutions that have done peer-reviewed research on pink noise frequency. And the website that I built, it gives people a reason as to what they are using and why it would help them, which is actually a segue of building trust with the customer. If a customer knows why it would help them, then it can bring in more people to use the product.
Karen: That’s very interesting that you are focused on pink noise. At one point back in my corporate career, I was working in an office. It had cubicles, and very noisy, and I had headphones. I was like, “Okay, what’s a good thing I can listen to that would help me?” And I tried some white noise, and then I did actually try pink noise, just to put it on my headphones and drown out the noises. So I kind of experimented with that.
And then when I first started writing on Substack, I was looking into AI and music, and I found there were some tools that were actually ethically developed were in what they called functional music. And it sounds like your Stillpoint fits, sort of, into that space.
KP: Yeah. I also mentor people, so there is a common pattern that emerges where people tell me that they can’t sleep, or they wake up at 2:00 AM, 3:00 AM. And one of my mentees, I recommended to them that they can listen to sleep music. At the time, they used to sleep four hours in 24 hours, and they had a job for around 10 to 20 hours a week. So if they are not sleeping, then there is something else happening. So I recommended it to them, and I received their feedback. It was a mentee not through AnitaB. They emailed me and said that, “Now I sleep like a baby.” So that is also something that I’ve seen. It helps people a lot.
Karen: That’s very nice. And it’s great that you are releasing it as a free tool, basically as a kind gesture to the world.
KP: Thank you. That’s very kind of you.
Karen: You mentioned that you have a PhD in AI, so you obviously have some professional experience in studying the technologies and using them. Can you talk a little bit about both your personal and professional uses of AI?
KP: Yes. I was doing a PhD. I changed my program to PhD in AI this January. But my experience spans both professional and personal, and they have informed each other more than I expected. For example, personally, I use AI systems across different domains where the decisions, these systems, have influenced real consequences. And what that experience gives you isn’t technical depth; it gives you judgment. It’s like developing an instinct for the difference between a system that performs well in a demo versus the one that holds up in production when it’s dealing with a lot of people. People are complex, right? And real situations. So I have a very opinionated take on different LLM products especially.
Personally, I have used browser search assistants with AI built in. And what I found as the biggest learning from personal use isn’t about technology, it’s about how humans relate to confidence, which is that the tools communicate with consistent confidence regardless of the accuracy. And we are wired to trust confident communication a lot more than if the product will tell us that “I don’t have an answer”. So the skill that matters the most when using these tools isn’t knowing how to prompt, it’s knowing how to verify, and how much to trust it as well. So probably that’s why prompt engineering is such a big thing right now in a training environment, not for the people.
Karen: Thank you for sharing that. So can you give some examples of the domains that you’ve worked with AI in, and maybe share an example of a personal task that you use AI tools to help you do?
KP: Yeah. I use AI for research and synthesis tasks. Sometimes I use it across different kinds of work regularly. For example, Stillpoint. I had to do deep research on what exists and what can exist that will have a unique value proposition. Where are there gaps in the market that I can fill in? So that is where, for example, deep research of different LLM products came into the picture. And then comparing them against each other can go a long way, instead of focusing on just one LLM product.
I think that what works genuinely well is speed of getting oriented. When you’re moving across domains quickly, the product work requires being able to get up to speed and find the shape of the problem and pressure test the product, which is genuinely valuable. Because right now pressure testing is incredibly faster than it used to be at some point, and it also can be more channelized rather than throwing darts in the dark that, “Hey, let’s try it.” Now we are trying it in a more channelized version. And what I’ve learned the hard way is that the output reads the same, whether it’s right or wrong. There is no signal. So human writing that’s uncertain usually sounds uncertain. That’s the texture to it. But AI output doesn’t have that same texture. So things have changed for me practically. I use these tools for speed and structure. Anything consequential, I get it verified independently. And it’s not like I’m criticizing these products. It’s using these products very mindfully.
Karen: You mentioned earlier that you’d done some work with healthcare platforms, autonomous systems, enterprise software. I think they’re all interesting. We all have exposure to healthcare platforms. I recently interviewed a medical researcher working out of Georgia Tech, and she’s done a lot of work on understanding from the clinical perspective. She’s an MD and has a master’s degree, and so she has one perspective on healthcare platforms. And then we all have a perspective from using MyChart and having systems that collect our data and use it. So I think healthcare is a particularly interesting area.
Autonomous systems are an area that I worked in a lot. So I’m curious if you worked with some founders in these areas, or what your thoughts are about those domains? What’s an example of a problem in those spaces that you worked on?
KP: Mostly what I worked on is, for example, evals, edge cases. With respect to edge cases, if you’re working in AV [autonomous vehicles], there is no standard database. A lot of edge cases show up when they show up. You cannot predict everything. So that is something that I worked on.
With healthcare, I worked on research products which expedite forming a diagnosis, and also adding to research of global institutions. I think we all need a lot more guardrails when it comes to healthcare because there are compliance laws, etc., which companies do follow, but there’s also a lot of private information of a person. So separating private information from the diagnosis when you’re training a ML model can go a long way. You don’t have to give a person’s personal information, but you have to build technology that solves problems at scale. So that is what I think I have observed.
Other than that, in many cases, like highly sensitive products in fintech or in legal, a lot of companies working with founders are framing as being AI-first companies, in that many companies tend to forget that in the end the audience is human, so they need to ensure that there are guardrails and they’re following compliance laws.
So I’m hoping moving forward in the future, when it comes to these highly sensitive industries, there is a consortium of people who can say what AI can or what AI cannot do legally. So I’m hoping it happens in the future.
Karen: Yeah, that’s an important perspective, especially with healthcare. In the US we have HIPAA laws, but there’s also concern about … you want to have enough data that you do something meaningful and you have enough context. A lot of the medical diagnosis tools are built around, for instance, data that came mostly from men. And women of certain ages are not often included in studies. Or in some cases, older people who metabolize medications differently and have different conditions are not always included. So there’s a lot of aspects to it. You want to have enough data that you have something that applies to the patient that you’re trying to diagnose, but at the same time, there’s a risk of using someone’s data inappropriately. Some of these algorithms have been reported to be gender-biased and race-biased and other aspects of the tools. Like you said, it’s one thing to train it and have it work in a demo, and it’s another thing to have it work in the real world with all the great variety of people that we have.
KP: Yeah. Adding to it, if there is, for example, a lot of data about just men, no data about women, companies are right now adding synthetic data. Organizations in healthcare can manufacture women’s data to train ML models so that they can form a hypothesis and solve problems at scale. Same as if they don’t have enough data about older people, they can use predictive analytics to see the journey of a human from a child. Usually doctors have data of a person when they become adult, everything, and they can use those patterns to build synthetic data for older people so that they have some database to train the models on. So something like that can also be helpful without introducing biases.
Karen: Yeah, synthetic data is really interesting. We’ll probably have to park that for another follow-up discussion. One of my colleagues, Lakshmi Veeramani, wrote an article about using AI tools to generate synthetic data for tools, and it’s an interesting area.
In some of the work that I had done in the corporate world, we used physics-based models to generate data to use for training. For instance, for equipment health, which is a different area than human health, but it’s the same in that the synthetic data is only as good as the models and the assumptions that are driving the creation of the data. So you can get into some really interesting paths there. But yeah, it’s a good example of a time when synthetic data is useful.
So you’ve talked a lot about different ways that you’ve used AI tools that had these different features, using it for research and for synthesis. I wanted to go back to that, just make sure that we’ve covered what you wanted to say there, as far as a specific story on ways that you’ve used a tool that had these features? What worked well, and maybe an example of something that didn’t work so well?
KP: So when I say what works well, I would need to deeply understand what product I’m using. For example, if I’m using ChatGPT, ChatGPT is very quick to give me information that’s absolutely needed, maybe in conversations, etc. But with Gemini, it tends to introduce its own opinions. Sometimes I tell it to not be Shakespearean and not be... It’s not a soap opera. I just need an answer! I have to explicitly tell it. But sometimes its introduction of opinions helps me a lot in deep research, etc. Now, when it comes to Claude, Claude is trained to deeply understand context a lot better than other products. So it depends on whatever I’m using the product for. That depends.
But what hasn’t worked well is, or what I haven’t used it for, is mentorship, for example. If I use AI to help people in mentorship, then it would be a moot point of me starting this journey of empowering people. The whole idea … I started this journey because everyone is different. Looking back, I really wish that I had someone like me in the beginning of my career. I did not have someone, so I became that person for other people.
Now, everyone is different, so I have to change my approach. It's also a lot of mindset, as well as confidence mentorship. One of my mentees recently started working directly with a Director of Engineering, and she’s the only one who’s hired for the company’s AI initiatives. No one else in her engineering team was hired. And she told me she could do this because I made her recite what she did confidently, so that when the Director called her on a vacation that, “Hey, I’m doing this. How are you using AI?”, she could confidently say it. She already had it under the belt, but she did not know how to own it. So that has helped her a lot. And I am also helping her build her brand. So in this case, using AI really doesn’t help me at all. It’s very human-to-human.
Karen: Yeah, that makes total sense. That human connection for the mentorship I think is an important aspect. And it’s great that you’re keeping a focus on that and individualizing it to the different people that you mentor. How about when you’re writing? Were you writing for your newsletter?
KP: When it comes to writing, I polish my writing using AI, but I don’t want to do basic stuff. A long time ago, I used to write case studies. There was a recession in 2008, so I have read about it in my MBA. I think that if we study that recession, we study why certain companies thrived and why certain companies lost their business, we can see a pattern emerging. And whatever it is, we can use some of it to learn from it. Similarly is with learning about how AI can be used in the future, whether it’s compliance, healthcare, and how usage of AI can be limited. So in that case, I prefer not to use AI to generate ideas. I use AI to just polish my writing or create images. One of my writings was an example about Gibson guitar, how the business thrived, and what we can learn from the things that happened in that case study. It was maybe a year ago.
So similarly, when I write about governance, I write about very basic challenges, foundational challenges. It’s as simple as your data taxonomy and data schema, not well to scale AI, because it’s a very foundational challenge. So that is something that has to come from within me. Also, it has to be a novel idea.
Karen: It sounds like you’ve mostly used LLMs for words. You mentioned, I think, maybe using it for some images. I’m curious if you’ve ever tried the tools for creating videos or music or any of those other modes of AI tools?
KP: I have used to create AI music. And it’s really amazing that it created great music. I’ve used it for images as well. For images I’ve used ChatGPT and Gemini, largely. I’ve noticed that Gemini follows instructions when it comes to images a lot more than ChatGPT. I try to give a comical edge to everything because we don’t want to always be serious, right? And when I tell it that the image has to be comical, It actually generates an image. With ChatGPT, the image has the text that I was anyways going to write in my post, so it’s funny that way.
Karen: Okay. Yeah, that’s interesting. So with music, you said that it created some great music. I had done a series on music a couple years ago, I guess two years ago when I first started writing on Substack about AI and music, and there were a couple of generative AI tools for music that were actually ethically trained. One of them was called Beatoven, and it was actually built around libraries of ethically sourced Indian classical music. I thought that was very interesting, that they paid artists for doing the work and then used that as the basis for the tool. There aren’t that many tools that are actually ethically developed, so I was pleasantly surprised to find that one. So I’m always curious to hear which tools people have tried.
KP: That’s amazing there are actually such products. I know a few people who are using voice-based training for ML models so that they can generate music. A lot of people are, for example, lending their voice, and the companies are paying them. They are not just using their voice, because their music, their voice anyways, is on YouTube or on different platforms, but they’re getting paid. So that is another example of ethical use of AI.
Karen: Oh, that’s great, too. So they’re voice artists and the companies are actually compensating them appropriately. That’s great!
KP: They’re actually selling their voice to train ML models, which is really good, I think.
Karen: One of the concerns that we hear a lot about is where these tools get the data that they use to train on. Sometimes it comes from systems, like you mentioned YouTube. Do you think that companies should be required to get consent, or credit, or compensate the people whose data they want to use for training? Or do you feel like it’s okay in some cases for them not to do that?
KP: I think the framework of, for example, consent, credit, and compensation is right. But I think the conversation gets stuck because it’s framed as companies versus users, and that framing tends to miss something important. Most people, for example, if there are new people in the industry, in music industry or in any industry, they would be happy to share their data if they get more gigs or if they get visibility. Because right now building a brand is significantly more important than anything else. I’m one of those people who used to tell me that “I’m not a shoe or an item of clothing, and don’t tell me to build a brand.” But over time, I’ve learned that building a brand is important. With respect to, for example, relevant recommendations, personalization, that is kind of a legitimate trade-off.
So when it comes to compensation, the compensation doesn’t have to be as large if people are using those products in their own favor as well. I think the real problem is, for example, the same data pipeline which enables useful prioritization creates exposure to bad actors. So I was thinking about this. Regardless of wherever we are, there are a lot smart people in this market who will use AI and AI products or foundational models in a bad way to game the system, so we don’t want that to happen.
So I will go back to my point that I mentioned earlier that larger companies have it in their power to set up a consortium where they’re safeguarding people who don’t understand technology, and also who don’t need to use technology as much. They’re largely consumers. For example, restaurant users. If they’re using AI products, but every time someone puts in their card and their personal information is leaked, it doesn’t have to happen like that.
So the product question would be not what data to collect, it’s what to actively prune. For example, an ML model that collects data and trains on it should be also trained to not collect any of the personal information. If I’m getting recommendations and I don’t want them, if I feel attacked, then it should be totally pruned. So I think that organizations can live through that ethical choice, because in moving forward, it can risk architecture choices for any of the large organizations.
Karen: Yeah, the collection of sensitive data is a really important consideration, like you said. A lot of times there has been an attitude of, “We don’t know yet what we can use, so let’s just collect everything and keep it.” But people don’t always realize that that creates a liability on their side. And it’s their responsibility, and data’s not actually free. It may be cheap, but it’s not free. So it’s interesting that you’re calling that out, that you need to prune it. There are data retention policies, but they vary. One of the chapters in our upcoming book is written by a lawyer, and she talks about this: about making good decisions about what to keep, how long to keep it, and having a policy about how long you keep it. And it applies to a legal practice, but it applies to, really, almost any company.
KP: Yeah. A lot of organizations actually tell you that, “We will keep your data for 72 hours, and then we will delete it.” Some people, some organizations keep it 30 days. So they disclose it to the customer, which is also building trust with the customer using that transparency.
Karen: You’ve mentioned ChatGPT and Gemini and Claude. Do you feel like they’ve been transparent with you as a user about where they got the data that they used to train their models?
KP: I’ve read from the news, not specifically with me, that they are transparent. And with a lot of LLM products that I use, they tell you things that will make sense to you. In many cases, for example, a product tells you that, “We are going to use this data, but it’s not going to be tied to you as much.” So it’s like they’re disclosing it, which is a good thing, but still it is data. And if, for example, they are training on the questions that are asked, the music that’s generated, the image that’s generated, so probably moving forward, a layperson who’s using these products, they can be paid. It could be money. It could be a pro account for some months or some days, something like that. So I think that that would be probably a good way from here on. I think it’s a very sensitive question.
Karen: It definitely is! Most of the time when I see companies talk about how long they’re keeping data, they basically say in their terms and conditions – which, you know, almost nobody reads, for good reason, because they’re not really understandable – it’ll say things like, “We keep your data as long as required by, for our business purposes and legal”. You know, they don’t really put tight limits on it. And it’s funny, I was just working on publishing our last book, AI Everywhere Volume I, as an audiobook. And so I went to sign up with an audio provider, and they wanted a copy of my government ID. You don’t need that! Were they going to use it, like you said, 72 hours to validate and then they would get rid of it? Could I trust them to get rid of it? Or were they going to keep it indefinitely and then sell it off to somebody? I just didn’t know. And I ended up deciding I wasn’t going to use that audiobook provider because I could not trust them with that personal information. I had no faith that they would treat it properly. But then you wonder how many people actually gave it to them. So I’m always happy to hear of companies that are doing it the right way, so if you have any examples?
KP: I know that Apple keeps it for some time, and then some other companies also. But I noticed that, for example, a company takes your information. Maybe it keeps it for six months, a year, something like that. But if they have third-party organizations that they partner with to solve their business problems, that is where the problem starts. Transparency through one-on-one is okay, but a third-party organization doesn’t have to disclose to you for how long they’re going to keep your data and how they’re going to use it. And a layperson doesn’t know because T&Cs are very ambiguous, and there are a lot of jargon. So people don’t understand what’s going on.
Transparency on training data can be very hard, but it’s important to dissect the problem rather than jumping to the solutions. For example, the problem isn’t that organizations are hiding information, it’s that data provenance isn’t being treated as a product requirement from the start. So it gets handled as a compliance checkbox at the end. And by then, the information needed to answer any questions is scattered across vendors, pipelines. You don’t know how many pipelines your information is going to go through.
So the solution that would actually scale is a law that can apply to any industry, any size of organization, treating data lineage the way software engineering treats version control. Every input is tracked, and every transformation is documented. So wherever your information is going, it should be auto-logged so that I as a user can see who all has my information. And every decision is logged at the point, decision as in what is done with the data, and not reconstructed later or for a regulator. This has to be built in the beginning as a standard practice. I often think that as a product person, we can learn from different industries what we can apply in software industries. If you look at ML terminologies, a lot of them are picked up from electrical engineering or hardware products, etc. So supply chains do this, and pharmaceuticals do this. And the reason that AI hasn’t is that the cost of not doing it hasn’t failed yet at an organizational level. It has failed at some level, but it yet hasn’t created a ripple effect of negative impact on companies, their revenue, etc. So I think that it’s important to build auditability in the very beginning of the product.
Karen: Definitely a really good point. We’re starting to hear a little more talk about data provenance and the importance of having it from the beginning, but I think it needs more attention, and I’m glad that you’re highlighting that.
KP: I was actually reading about it, and I was thinking that if all the companies join their hands and decide what safe AI would look like, what data provenance would look like, then they can make a lot of impact on the consumer side of this, the product users. I live in the tech bubble. A lot of people have the luxury to live in the tech bubble, but it is still our responsibility towards people who don’t live in that bubble. We still owe it to them to create safe and responsible AI products.
Karen: Absolutely. Yeah. So as consumers ourselves and members of the public, our information’s being used by a lot of AI-based tools and systems, some that we can control and some that we can’t. Do you know of any specific cases that you could share where your data’s been used?
KP: Yeah, actually, funny enough, one of my friends told me a while ago that she receives a lot of spam calls, and people have her email addresses and phone number. And I have been receiving not only emails and phone calls; I have also been receiving mails in this mailbox outside of my house. A lot of times, someone reaches out to me that “We have a lawsuit and you will get $3,000, $4,000; click on this link.” But it’s very funny to see that they have written at the end of the letter that, “Don’t call the court or the court clerk to confirm if this is a real case.” So kind of a joke that people are doing this, and they’re not doing it very seriously probably. That’s why they have written it at the end. So this is kind of a real texture of how data exposure affects ordinary people.
Now, I’m educated enough to understand that I should not go through those links, etc. But sometimes people, regardless of education or whatever, they’re going through some life-threatening issues. What if they click on the links and their identity is stolen? What if their money is taken away? So many people have limited money for their living expenses, healthcare crisis, etc. And once the money is gone, some transactions through certain companies are final. They cannot be reverted back. So if something like that happens, a layperson tends to lose a lot. I haven’t experienced this apart from the scam calls I receive. I don’t take those calls, so they found my address, probably.
Karen: Yeah, I have had my phone set for ages to reject unknown calls. I don’t even know how many I get per day, so I think my phone number got sold many years ago. This started long before AI came along.
KP: Yeah, mine too, but since AI came along, I have started receiving scam calls. Before that, I did not receive as many. Like, maybe one or two in a year used to be scam calls. Now there are five or six per day, every day.
Karen: Wow. That’s annoying.
KP: Yeah, and with such people, they know what your area code is, so they will call from the same area code or similar area codes which are commonly used in the geographical location that you live in. So they make you feel that they are calling from your area. It’s a psychological play, I think.
Karen: Yeah, and we don’t know where they collected that. But one thing that I think we have seen with AI is that the scams are getting technically better because they have more information about you, and they can put it together in better ways. The whole phenomenon of deepfakes, right? The quality is improving for the way that they generate, and that’s a problem.
KP: Yeah. I think we should be responsible for that.
Karen: I like your emphasis on us being responsible for helping all the people who don’t live in the tech bubble like we do to be able to live safely.
Do you know anyone that you gave your information to that made you aware that they might use it for training an AI system?
KP: I don’t want to name the organizations. I haven’t read T&Cs, but I think every company where you, for example, you check a box, if you receive credit card offers, or if you are signing up with an AI bot that your mobile app, banking app is using. So unless you check that terms and conditions box, then you don’t know what’s in the T&Cs. And T&Cs are, like, 17 or 18 pages long. They’re not meant for laypeople to understand. So the companies haven’t made us aware. But I think that coming to the solution part of it, wearing the product hat, the solution is not ‘we simplify the process’. It’s granularity and recurrence. Most people would genuinely engage with their data choices if those choices help them, and if those choices were made in plain language, broken into meaningful categories. So what I was thinking is that instead of people having to check a terms and condition box, that terms and condition can be generated, maybe using generative AI, in the layman’s person, and people can check what they want to say yes to and what they want to say no to, instead of saying yes to all T&Cs together.
Secondly, people maybe can get reminders in a month or two months that, “Are you still okay with your choice?” I would have to do testing, how much time is good enough for this? So, “Are you comfortable with this? Here’s what has changed”. And this is a very solvable product problem. I think the reason it hasn’t been solved is that there is a constant friction that reduces data available for monetization, and the incentives run exactly the other way.
Which means that this won’t get fixed through good intentions. It does require structural pressure from the people who can have that power.
Karen: You mentioned that you’ve been getting a lot of scam calls. Have you ever had your privacy invaded, lost income, or had any phishing scams that cause trouble for you?
KP: Yes. So I did lose some money a long time ago. But what I’ve noticed is that the most common experience of data misuse is not that dramatic. It’s very ambient, which is that it accumulates quietly in ways that are hard to trace and which are impossible to reverse. It could be a different currency type. The pattern that affects the most people is that most context is aggregation. Individual data points seem harmless in isolation: a name here, a phone number there, and an employer somewhere else, which becomes something that is more powerful when it’s combined. And that combination happens invisibly across systems that were never designed to talk to each other, but they do.
So I think that the reason that it is hard to solve isn’t technical. It’s incentive design. For example, the data has value precisely because it’s connected. The more it connects, the more valuable it becomes, because it’s easier for the companies to understand your psychographics and demographics, of course. So the natural market runs this thing. The direction runs exactly opposite to user safety, which has been natural for quite some time. The solution that I think would actually work, scalably across industries, is shifting accountability to the point of aggregation, not the point of collection. Right now, every individual data collection feels reasonable, but it’s only at the aggregation layer that the harm becomes visible. And also, regulation and product design, they both need to catch up to that reality. Regulation cannot be behind product design at this point.
Karen: Thank you for sharing that. Last question, and then we can talk about anything else that you want. We’re seeing that public distrust of these AI and tech companies has been growing, and partly it’s because we’re realizing more and more what they do with our data and how they’re sharing it and what they’re using it for. I’m wondering what your thoughts are about trust. Is there any company that you do trust? And what would you think is the one most important thing a company would need to do to earn and keep your trust, if that’s even possible?
KP: I think it’s important that, for example, accountability and trust to be created, it needs to be incentive-based. In an organization which is building a lot of AI products, even if there is incentive to them building trust with the customer and building a safe AI, it could very well be deprioritized, because right now there are AI wars going on in the market. So I think that if it is incentive-based, for example, if the product or the company says, “You can trust me with your information,” they can have a defensible moat in this market.
And what I’ve observed is that trust erodes, not because companies are careless, but because the way the system gets built, it distributes responsibility across many functions that nobody really owns the outcome end to end. So companies can have teams which make data decisions very early on, before everything starts. Before even the first iteration starts, when you’re launching a beta product, or when you’re internally launching products. So that way, if the decisions are coming from a singular entity within an organization, it can be scaled very well.
And also I want to see that every AI organization is being able to answer one question clearly, which is: who owns what the system does to people? Not the process, the compliance checklist, but actually human outcome, because we have seen people going through a lot of pain in their life because AI products recommended something to them, and it led to detrimental results. So if that answer is distributed across five teams, because every team is looking for their incentive – like, sales team has to meet a quota, an engineering team has to release the product which is engineering efficient, and product team has to listen to all the teams and coordinate what’s best for business and customer outcomes – then it is not going to be very effective.
Rather, the first decision should come from the team that maybe has two votes versus one with respect to other teams, so that they know what trust looks like moving forward.
Karen: Yeah. That’s a great summary. The question of who owns what the system does to the people – as you said, if the ownership is fragmented, then basically nobody owns it, right?
KP: Yeah. I think that some companies have this thing in them. They have teams for safety and regulation, but some teams are under a lot of pressure, so they have to make the best decision to make market entry and to make penetration to the market for their business. So when push comes to shove, these teams have to be responsible.
Karen: All right. KP, I really appreciate you making time for this interview today. Is there anything else that you’d like to share with our audience?
KP: Yes. I would love to share that I launched a product, which is Stillpoint. The design philosophy is that most tech products are optimized for time on app, or DAUs [daily active users], MAUs [monthly active users]. The longer you stay, the better the metrics look. But Stillpoint is a deliberate counter-pattern to that. As a PM, it is designed for immediate utility and rapid disengagement. You come in, you reset, and you leave restored. The goal is to get you off it as soon as possible, and it’s an intentional architectural choice, not a limitation. Full case study is on my Substack.
The Stillpoint link is stillpoint-by-kp.netlify.app. So this is the product. I would love for people to use it and leave a review and maybe send me feedback as well. I would love to scale it.
And also, I’m looking for a new product role, so if anyone is hiring or anyone who knows anyone who’s hiring, please feel free to reach out.
Karen: You mentioned your Substack. We’ll put a link in the interview, but for the people who are hearing it instead, can you tell our listening audience the name and the URL for your Substack?
KP: It’s kanupriyayakhmi.substack.com. It’s K-A-N-U-P-R-I-Y-A Y-A-K-H-M-I dot substack dot com. I tried to shorten it to KP Yakhmi, but it led to some other page, so I just changed it back.
Karen: All right. And what would be the best way for someone to contact you if they have a product role that might be of interest?
KP: They can contact me via my LinkedIn. So I’ll paste the LinkedIn in the transcript of this one. They can also email me at my first name last name at gmail.com.
Karen: The spelling is the same as your Substack, right?
KP: Yeah, it is.
Karen: Yep. So that’ll work for two purposes. Great. Well, thank you so much. I really appreciate you sharing your insights from the product perspective, because that’s where the solutions need to start: making sure we understand the problem that the product needs to solve, and keeping that perspective, keeping the user at front. And I think it’s great that you’re focused on that.
I wish you a lot of luck in your search for a new position, and with getting Stillpoint to take off. And I’m sure we’ll be staying in touch, so thank you so much.
KP: Thank you. Thank you so much, Karen, for making the time for me, and for everything that you do. Thank you.
Interview References and Links
Kanupriya (KP) Yakhmi, MBA, PhD in AI on LinkedIn
KP’s website
Stillpoint application (free)
KP Yakhmi on Substack (Stillpoint case study)
About this interview series and newsletter
This post is part of our AI6P interview series on “AI, Software, and Wetware”. It showcases how real people around the world are using their wetware (brains and human intelligence) with AI-based software tools, or are being affected by AI.
And we’re all being affected by AI nowadays in our daily lives, perhaps more than we realize. For some examples, see post “But I Don’t Use AI”:
We want to hear from a diverse pool of people worldwide in a variety of roles. (No technical experience with AI is required.) If you’re interested in being a featured interview guest, anonymous or with credit, please check out our guest FAQ and get in touch!
6 'P's in AI Pods (AI6P) is a 100% human-authored, 100% reader-supported publication. (No ads, no affiliate links, no paywalls on new posts). All new posts are FREE to read and listen to. To automatically receive new AI6P posts and support our work, consider becoming a subscriber:
Series Credits and References
Disclaimer: This content is for informational purposes only and does not and should not be considered professional advice. Information is believed to be current at the time of publication but may become outdated. Please verify details before relying on it. All works, downloads, and services provided through 6 'P's in AI Pods (AI6P) publication are subject to the Publisher Terms available here. By using this content you agree to the Publisher Terms.
Audio Sound Effect from Pixabay
Microphone photo by Michal Czyz on Unsplash (contact Michal Czyz on LinkedIn)
Credit to CIPRI (Cultural Intellectual Property Rights Initiative®) for their “3Cs' Rule: Consent. Credit. Compensation©.”
Credit to Beth Spencer for the “Created With Human Intelligence” badge we use to reflect our commitment that content in these interviews will be human-created:
If you enjoyed this interview, my guest and I would love to have your support via a heart, share, restack, or Note! (One-time tips or voluntary donations via paid subscription are always welcome and appreciated, too 😊)
