Guardians of AI: Shubham Anandani On How AI Leaders Are Keeping AI Safe, Ethical, Responsible, and True
“Again, these are my own views, but it’s very easy to fake anything nowadays. I have seen people having lots of followers while not being authentic. Platforms often take action against inauthentic accounts once they identify them. But yes, I think anything can be faked in this world, and we are all fighting against counterfeit and fake things.”
When Shubham Anandani was about five years old, he discovered that computer games were more interesting when he could change how they worked. Growing up in India in a joint household, he watched his older cousins prepare for careers in software engineering. They were learning to code, and he was watching closely. Before long, he was experimenting with BASIC and Logo, then digging into the configuration files of DOS games to modify them to his liking. It was an early introduction to a principle that would follow him throughout his career: understanding how a system works gives you the power to change it.
Today, Anandani faces a considerably more complicated version of that challenge. As an engineering leader and researcher focused on digital identity and trust, he works in a world where artificial intelligence can generate convincing code, photographs, professional profiles, and even video impersonations. The technology that makes extraordinary things possible also makes it increasingly difficult to distinguish something genuine from something manufactured. For someone who has spent his life exploring what computers can do, the question has become how to decide when their output deserves to be trusted.
His fascination with software eventually took him to Carnegie Mellon University, where he studied software engineering and investigated how people evaluate the trustworthiness of code repositories. On platforms such as GitHub, developers have traditionally relied on signals like stars and forks to judge whether a project deserves their attention. But popularity and reliability are different things. Anandani became interested in what those signals actually reveal and how similar questions might apply to professional identities.
His early career also included work at Oracle, developing software for clinical trial randomization, a field where technical systems serve particularly consequential purposes. He later joined a company whose mission of creating economic opportunities for the global workforce appealed to him. Nearly a decade later, he remains there, having moved into engineering management while continuing to explore the technical questions that originally drew him to the profession. Anandani speaks in a personal capacity; the views expressed are his own.
The shift from programmer to manager has changed the nature of his work. Anandani now leads a team of engineers, balancing their professional development with business priorities and longer-term engineering strategy. He divides his working hours among people management, planning, operational responsibilities, and increasingly, direct technical contributions supported by AI agents. Rather than abandoning hands-on engineering as his leadership responsibilities have grown, he uses these tools to remain involved in building software.
His approach to AI reveals a tension between his enthusiasm for automation and his insistence on verification. He divides its impact into two categories: “how we build” and “what we build.” The first concerns the engineering process itself, including coding, testing, and deployment. The second concerns how AI changes the products that customers actually use.
He prefers the expression “augmented intelligence” because it better captures how he believes the technology should expand human capabilities.
But Anandani is skeptical of simplistic productivity claims. An engineer might initially report becoming five or six times more productive with AI assistance. Once the additional work of checking accuracy, evaluating quality, and correcting mistakes enters the calculation, he suggests the actual improvement might be closer to 10 to 20 percent. These are illustrative estimates rather than measured industry results, but the distinction matters. Producing something quickly is different from producing something that can safely be used.
His proposed solution involves specialized AI agents working in complementary roles. One generates code, another critiques it, and another evaluates how the changes affect the larger product. Even then, he argues, highly critical systems require human oversight. The goal is to make automation more dependable without pretending that its output is automatically trustworthy.
That same concern shapes his research into professional identity. Anandani’s framework extends Frei’s Trust Triangle of authenticity, credibility, and empathy with a fourth quality, uniqueness, and argues that trust depends on all four. Verification can help establish that a person is who they claim to be. Credibility requires evidence beyond someone’s own professional assertions, including validation from employers or other trusted parties. Empathy emerges through meaningful engagement with others, while uniqueness reflects the personal qualities that distinguish one individual from another.
He calls that last quality a person’s “human superpower,” a revealing choice of words from someone whose profession increasingly involves machines that imitate human expression.
The difficulty is that nearly every conventional signal of authenticity can now be manipulated. Photographs, follower counts, professional credentials, and even video conversations are becoming less reliable as standalone evidence. Anandani points to the growing importance of liveness detection and other verification methods, while acknowledging that the challenge is far from solved.
“Anything can be faked in this world,” he observes.
There is a personal dimension to his philosophy of technology, too. Despite his enthusiasm for AI-assisted development, he resists allowing machines to replace the original thinking behind his work. “I personally like to start off with my own views and then use AI on top of it,” he explains, emphasizing the importance of preserving human originality.
It is a fitting perspective for a man whose earliest technological experiments involved modifying games to suit his own imagination. The tools have become vastly more powerful, and the consequences of manipulating digital systems are now considerably greater. Anandani remains optimistic about what artificial intelligence can accomplish, but his work increasingly returns to a question that engineering alone may never fully resolve: when almost anything can be convincingly manufactured, how do we recognize what is real?
Yitzi Weiner: Well, Shubham, it’s so nice to meet you. Before we dive in and talk about your great work, our readers would love to learn about your personal origin story. Can you share with us the story of your childhood and how you grew up, and particularly the seeds of all the amazing creative ingenuity that has come since then?
Shubham Anandani: I was born and brought up in India. I still remember I was around five years of age when I had my cousin brothers living in the same joint household setup, and they were all gearing up to be software engineers. I would see them code in their day-to-day, and I would learn basic tools at that time.
I still remember I started to get my hands on computers and started coding in BASIC or Logo as a programming language, which back then, I think none of the people in my generation right now would be able to relate with. But I really enjoyed just tinkering with the new programming languages that I was seeing them use. In turn, I would go back to all the DOS-based program games that I would play in my day-to-day, and I would try to tinker with their config files to modify them in a way as per my liking. That’s how my computer science interest grew, and I wanted to be a software engineer. That’s where I am right now.
Yitzi Weiner: Amazing. Please tell us about the amazing work that you’re doing today, and then we’ll talk about what led you to it. Tell us about your work today.
Shubham Anandani: Sure. I’m an engineering leader who builds digital identity and trust systems. I’m also secretary of the IEEE Seattle joint chapter. I started my career at Oracle on clinical trial randomization software, and I studied software engineering at Carnegie Mellon University.
These days, I’m actively researching how to build trust in the professional identity space: how, as a professional, you can prove who you are and why people should trust you. I have published with IEEE, and I’ve spoken at several IEEE conferences, SecureWorld, healthcare, and identity domain conferences. I can dive deeper if you would like and tell a little bit more about what trust truly means in professional identity.
Yitzi Weiner: Amazing. Could you tell us the story of how you got this great job and what led to it?
Shubham Anandani: Sure. I was at Carnegie Mellon, and this was my first job after graduating. I’ve been in the same company for almost 10 years now. My job-hunting process was purely driven by which company’s mission and vision I truly resonated with. At that time, I remember speaking with several employers, and this one truly stood out because, one, it’s creating economic opportunities for the global workforce, and two, I could see how I could make real-world impact through the work that I’m doing in my day-to-day.
With both of those things combined, and preparing for all the rounds of interviews every company goes through, I was lucky enough to get selected. The impact that I’ve created so far in the industry is something I’m truly proud of, and I can go out and talk about it as well.
Yitzi Weiner: Unbelievable. Could you tell us a bit about what you do? What’s your day-to-day job like? Are you doing actual coding today, or are you managing teams who are doing coding? Tell us a bit about what your job is like.
Shubham Anandani: Sure. As an engineering manager, I lead a team of engineers. My job is to make sure, one, my team is growing and successful in their own roles, making sure every individual’s growth is prioritized. The next goal, first and foremost, is to make sure I’m delivering value to the business: whatever the key priorities for the business are, working against those priorities and delivering impact that directly maps to what the business wants to prioritize.
The third, tertiary goal is obviously my own growth: how am I progressing in my career, and how am I prioritizing the things that would truly keep me on my toes and keep my learning curve steep enough? The best outcome usually comes when you align both your team’s goals and your personal goals with what the business wants. If you deliver impact there, then obviously you get higher rewards, promotions, and become more successful in your career.
As an engineering manager, my time is split across different things. I would say 20% to 30% goes into the people management aspect. Another 20% to 30% goes into shaping the strategy of my team, from what they’re doing right now to what they should be doing in the next one to five years. Another 20% to 30% goes into operations and engineering management logistics, handling the logistics of the team, what initiatives we are doing, and how we are being successful.
In the remaining time, I always have AI agents running, which are helping me be a little more hands-on, stay relevant in what’s trending in the industry, create impact passively in my current role, and land actual features that I would usually rely on my software engineers to do. Now I’m looking at optimizations that I can personally bring and scale.
Yitzi Weiner: Unbelievable. You mentioned AI agents, so I guess we might as well jump right in. Obviously, the world has changed over the past few years, and AI agents are really changing things. We’d love to hear how, as an engineering manager, you leverage AI agents and what that looks like in a big company like yours.
Shubham Anandani: I bucket them into two different pieces: the first one is how we build things here, and the second one is what we build.
“How we build” is where I see most of the productivity gains. This includes the tools that we use at each step where we want to engage AI tools, how we automate the CI/CD pipeline, and how whatever we are building gets to our users the fastest. Third, if you’re building using agents, how do we validate that before it hits production? You don’t want to keep a human in the loop always, but at the same time, you want to balance it with the right trade-off. For highly critical and visible surfaces, you want to make sure there are humans in the loop at all times. So, how do we use agents to validate and verify before anything hits production?
When it comes to “what we build,” instead of saying it’s artificial intelligence, I try to call it augmented intelligence. I resonate with that more, especially when we are using it to our benefit for scaling impact. It’s helping us see if what we are building right now is truly helpful for our users, and if we add agentic capabilities to those existing feature sets, how the user experience changes. You also want to be mindful that as you introduce these tools, obviously the cost goes up. So you want to make sure you’re balancing keeping a user-first mindset with how much you can operate within the current cost at hand. Those are the two different buckets, how we build and what we build, that are getting significantly influenced by AI.
Yitzi Weiner: Unbelievable. You said a lot of great things, and I want to learn some more. You said that you use agents to validate code or product before it goes to production. Let’s say you use Claude to make the initial code. Have you found that it’s better to use another type of model, let’s say an OpenAI model, to validate it, or is it okay to use an agent from the same family to validate it?
Shubham Anandani: Using agents from the same family works perfectly fine. It’s just that you want to have different thinking capabilities for the various agents that you are engaging. You want to make sure you create specialized agents. Imagine one agent is actually a software developer, while the other one is a strong critic of what they’re building. Then there is another agent that actually verifies how this all fits together, not just looking at a piece of code, but looking holistically at how this whole product is getting impacted by the changes that are landing.
You add all these evals in the pipeline. You add all these agents that critique, give feedback, and then communicate with the other agent to improve what they have already built and ignore all the noise that’s getting pushed in. That way, it’s holistic end-to-end development, not just purely focusing on some improvement or fixes that we started off with.
Yitzi Weiner: That’s amazing. And what’s the user interface for this? Do you use something like VS Code, a terminal, or another custom platform that has specified agents?
Shubham Anandani: It’s more a matter of developer preference. There are various tools available, and developers can choose the one that makes them the most productive. If VS Code works well for your team, then go for it. There are several other tools, like IntelliJ, or you can use the terminal itself to start with these agents and be productive. Whatever works for you is fine.
Yitzi Weiner: So, the big question that everyone struggles with, and you mentioned this, is that when you automate, you’re often giving away quality control. For example, let’s say I want to make a video. If I have different agents make the different shots, the quality is not usually the same as when I do each one myself. That’s a simple example, but I’m sure it’s similar with coding and software. How do you ensure that the automation is as good as individual, custom, bespoke code?
Shubham Anandani: Yeah, I think that’s a very valid challenge the whole industry is facing right now. We shouldn’t just be seeing these AI tools as enhancing our productivity when, in turn, there’s more work getting created just to validate, confirm, and make sure whatever we pitch to the AI agent is the same thing that’s getting generated, and with high quality.
I do also see, though, that the industry is getting very mature in terms of defining what the guardrails are, how the evals work, playing the red team in some sense, and making sure the output is of high quality. If you think about these additional investments, if you asked people earlier what the productivity gain was when using AI agents, someone might say, “Oh, I’m 5x more productive or 6x more productive.” But now, after all that validation is embedded in those estimates, you’ll realize that maybe it’s just 10% to 20% more effective, rather than increasing by several multiples.
You will find that people are investing heavily now on that side of things. I feel like once they get the hang of it, and once these frontier models become even more mature, eventually down the line you can start to see higher productivity gains as we make this side of the story more comprehensive. That’s how I envision it right now.
But I totally agree with you that you have to do a lot of hand-holding. You have to do a lot of validation yourself so that no slop gets sent down. Especially with you being on the editorial side of things, you can totally see how overwhelming content and pitches are getting with people using AI to just do it end to end. I personally like to start off with my own views and then use AI on top of it so that originality and human uniqueness stay in every piece of content. Without that, I totally agree that it’s just fluff that gets generated.
Yitzi Weiner: I’ve been trying to translate a 700-page book. I use Argon; it’s very powerful and it can translate the whole book. But now I’m wondering, because I’m going to spend so much time validating whether it’s accurate. I can have other AIs check it and they’re going to give their critique, but I can’t publish a book if I don’t really know if it’s accurate. So I’m thinking maybe it would have taken the same amount of time just to translate it myself.
Shubham Anandani: Yeah, there’s a similar parallel that I can draw with a software engineer who’s looking at a code repository with several files and folders of a codebase that they cannot parse by themselves. Obviously, they’ll have to rely on AI to summarize everything that’s going on, what the different feature sets are, and how they should be navigating that. So I can really see your pain.
Yitzi Weiner: Wow. Fascinating. Let’s talk a bit about AI transformation. Digital transformation has been a buzzword for a while, and now the new thing within that is AI transformation. Let’s say you have a doctor’s office. Maybe this is not your field, but in general, how does a company go about AI transformation, trying to transform their operations from analog and paper to AI? What does that look like?
Shubham Anandani: It’s a great question you asked in a timely manner. A lot of companies and teams are actually claiming that they are AI-native, and it’s just a buzzword unless you think of the whole process end to end.
How I think companies or teams should start looking at being AI-native is, one, they should look at what the most redundant process on their team is and see if there are any ways in which they can automate some of that to reduce their operational overhead. Second, is what they are building creating enough impact? And if it is not, how can AI augment and amplify the impact that they’re creating?
I think investing in these two dimensions using agentic capabilities can significantly help. A lot of new startups are spinning up around the first one, where you don’t have to go on call. You can use agents to let a developer sleep through the night. If there’s any issue on the website, the agents tackle it on their own: they put out the PRs, submit all the fixes, validate and ship each other’s code, and then the next morning you get to know that yes, there was downtime on your website and it was already self-handled. Investing in these two dimensions and thinking about how agents can help you in these spaces is going to make you truly AI-native in the long run, help you adopt faster, and create gains up front.
Yitzi Weiner: Let’s talk a bit about what you were talking about earlier: verifying identity and verifying that someone is human. This is something that I struggle with at Authority Magazine because we do written interviews. It used to be that when people wanted to make a fake profile for whatever nefarious reasons, they would put in an AI-generated picture, and you could kind of tell. So we said you have to put in two pictures, because it was hard to match two pictures. Nowadays, AI blows right past that. It’s almost impossible to tell; you can put in a hundred pictures of the same person, and you can send in an interview that looks perfectly human-written when it’s not. I’m sure companies like yours are struggling with what’s slop, what’s real, and what’s not. What are some basics that a company should know? How do we verify humanity?
Shubham Anandani: Yeah, it’s a very valid and timely question, especially in this age of AI when trust is no longer a given. You have to prove yourself, check yourself, and show up as authentic.
I’ll share my research work that I did around this topic. This is based on the foundation that I built at Carnegie Mellon University when I was researching on the GitHub side of things: if you’re looking at a repository of code, how do you decide whether to trust that codebase? Do you look at the star signals? Do you look at how many people forked that repository? I translated that into how you find someone trustworthy.
I divide it into four different pillars. It’s based on Frei’s Trust Triangle, which says trust can be evaluated based on authenticity, credibility, and empathy. You were talking about the authenticity piece, where verification fits right in, and I add a fourth layer of uniqueness, which is my own addition to the framework.
I’ll give you examples of how trust works in the identity space. First of all, if you want to prove that you’re authentic, you can get verified. A lot of platforms offer blue checkmarks where you can verify your workspace email or education email, and that adds some authenticity to your identity. Obviously, people used to rely on photos and having a clear name, location, and information filled out as proving authenticity.
Second is credibility. In the past, you could say, “Oh, I work at Authority Magazine, and I’ve been with them for the last eight years.” But in the age of AI, you claiming that you can do something is not enough. It’s the employer who needs to validate that yes, they worked here, they gained certain skill sets, and these are the skills. So you need second-degree endorsement, or someone vouching for you for what you’re claiming on your identity.
Third is empathy. A lot of people skip this part of trustworthiness because they think there’s always some friction in how much you want to go out and express yourself on these platforms. I always encourage people to engage with other people’s posts in a human way that doesn’t look AI-generated, sharing something resonating with the content they have posted. That demonstrates empathy.
The last one is humanness, or uniqueness, which is what I call your human superpower. This is how you come across as someone different. Let’s say I am an engineering leader working at Company Y, and there’s another engineering leader from Company Y. How do I stand out for a recruiter who is trying to hire an engineering leader for their company? People have social network profiles like TikTok where I can add those videos to my identity, or if I create YouTube content, I can use that to feature something unique about me and how I express what I speak about.
Combined holistically, it makes a trustworthy identity, and it helps someone who’s looking at your identity to meaningfully engage if they find it trustworthy at the end of the day.
Yitzi Weiner: At this point in the technology, if you see somebody that has 2,000 followers on LinkedIn and they’re all first connections, is that generally proof that the person is real, or can that be faked too?
Shubham Anandani: Again, these are my own views, but it’s very easy to fake anything nowadays. I have seen people having lots of followers while not being authentic. Platforms often take action against inauthentic accounts once they identify them. But yes, I think anything can be faked in this world, and we are all fighting against counterfeit and fake things.
Yitzi Weiner: What about doing a video call? Is that something that’s hard to fake, or is it easy?
Shubham Anandani: If you ask someone who is working on the video verification space, they would say yes, it is easy to fake. I’ve seen videos where someone uses a pre-recorded video to open someone else’s device that usually uses facial recognition. So people are looking into realness detection and liveness detection, giving challenges to people like “move this way or do that” to make sure. That’s a real challenge, and a lot of people are actually solving for that as well.
Yitzi Weiner: Wow. Actually, there was somebody I interviewed last week, and I wasn’t sure if I was talking to a real person or not. I couldn’t be sure. The background was weird, it was fuzzy, and till today I’m not sure. We’re in such a weird world now. Okay, amazing. This is our final aspirational question. Shubham, because of your amazing work, you’re really a person of enormous influence. If you could spread an idea or inspire a movement that would bring the most amount of good to the most amount of people, what would that be? Because you never know what your idea can spread.
Shubham Anandani: Oh, that’s an interesting question. In an ideal world, I would want to encourage people to act out of good spirit and with good intent while using these tools, looking at the brighter side while also knowing that these tools are a work in progress. If you think that something doesn’t work, or if you figure out a loophole, reporting it is the right next step and moving ahead with an open mind. AI creates a lot of uncertainty, but just making sure you are using things in the right way and having the right intent can go a long way for the whole society.
Yitzi Weiner: So beautiful. Shubham, how can our readers continue to follow your work? How can they support your work in any possible way?
Shubham Anandani: I have my own website, www.imshubh.com, and you can check me out on LinkedIn. Make sure you hit connect or follow me for the things I speak about. I have my own Medium page with the handle @imshubh5, so make sure to check that out as well. If you have anything in this space, or if there’s any way I can help you out, make sure to connect with me.
Yitzi Weiner: Shubham, it’s been truly a delight to meet you. I wish you continued success, good health, and blessings. I’d love to stay in touch, and I’d love to do this again next year.
Shubham Anandani: Sounds good. Thanks. It was really nice chatting with you as well. I loved our interaction; it was pretty candid and conversational. I’m looking forward to how it turns out and maybe collaborating again in the future.
Yitzi Weiner: You too. Have an amazing day, my friend. All the best.
Shubham Anandani: Thank you. Bye.
Guardians of AI: Shubham Anandani On How AI Leaders Are Keeping AI Safe, Ethical, Responsible, and… was originally published in Authority Magazine on Medium, where people are continuing the conversation by highlighting and responding to this story.