The Hidden Cost of Blocking the Wrong People: A Conversation with Kamya Batra

Every consumer platform runs on a quiet assumption: that it can tell the good actors from the bad ones. The stakes for getting it wrong keep climbing. People around the world lost $1.03 trillion to scams in a single year, which is why platforms keep adding checks and verification steps to keep bad actors out. The catch is that those same checks fall on everyone. Wrongly blocking legitimate customers, what the industry calls a false decline, already costs businesses an estimated 13 times more than the fraud those blocks prevent. Turn the safety dial too far in either direction and someone pays: users drowning in scams, or real customers turned away at the door.

Kamya Batra works at the center of that dial. A Senior Data Scientist at Meta with more than 7 years across machine learning and product strategy, she has spent her career building the systems that decide who gets through and who gets stopped on platforms used by billions, first in the global payments industry and now in messaging. She has peer-reviewed for the 5th International AIxIA Workshop on Artificial Intelligence for Healthcare, where the work is judging whether other researchers' methods hold up. Her view is that safety and growth are usually run as rival teams, and that the real craft is refusing to treat them as a choice.

We spoke with Batra about why blocking the wrong people is the most expensive mistake a platform can make, and how she designs systems that let real businesses move fast while making life hard for everyone else.

Most people think trust and safety is about catching bad actors. You talk about it as a balancing problem. Why?

Because catching bad actors is the easy half. If all you cared about was stopping fraud, you would make everything slow and suspicious, add a dozen verification steps, and block anything that looked slightly off. Fraud would drop. So would your business, because most of the people you just inconvenienced or banned were legitimate. Every control you add to stop the bad actor lands on the good one too, and the good ones are the overwhelming majority. That asymmetry is the whole game. It’s all about reducing harm but also enabling legitimate businesses to grow.

The trap is that the two failures are not felt equally. Missed fraud is loud. It shows up as complaints, chargebacks, and headlines. A legitimate customer who gets wrongly blocked just leaves, quietly, and usually never comes back, so nobody files a ticket about them. Teams keep tightening the screws because every pain they can feel points the same way. The discipline is refusing to optimize for only the failure you can see.

You work on systems that decide which businesses and people get access at scale. What does that actually involve?

A lot of it is onboarding and verification, which sounds dull until you realize it is the front door. Take business messaging. Legitimate companies want to reach customers where they already are, and that channel is growing fast: global retail spend through conversational channels is on track to reach $43 billion by 2028, up from roughly a quarter of that a few years earlier. Every one of those businesses has to be let in and verified first. If that process takes weeks and a dozen manual steps, you have strangled the growth you were trying to enable. If it takes seconds with no scrutiny, you have handed the keys to scammers running the same playbook at industrial scale.

So the work is designing a front door that opens fast for the real business and stays shut for the fake one, when both are knocking at the same time and often look identical on paper. That means figuring out which signals actually separate a legitimate company from someone impersonating one, deciding what has to be checked up front versus what can be watched after the fact, and building it so a good actor never feels the machinery quietly evaluating them. When it works, the legitimate business thinks onboarding was trivial. It was not.

When a legitimate account and a malicious one look the same on the surface, how do you tell them apart?

You stop looking at the surface. A scammer and a real business can fill out the same form with the same fields and look identical at signup. The difference shows up in behavior and context over time: who they contact, how, in what pattern, and how the people on the other end respond. A real business messaging its own customers looks nothing like an operation blasting strangers, even when their paperwork matches. So you build the decision around signals that are expensive for a bad actor to fake and cheap for a real one to produce naturally, and you weight them so a single odd data point does not sink a legitimate account.

The hard part is that you are never working with certainty. You are working with probabilities, and you have to decide how much confidence is enough to act, knowing every threshold you set trades one kind of error for the other. Set the bar to catch almost every scammer and you will wrongly stop a pile of real businesses. Loosen it so every real business sails through and you will let scammers in behind them. There is no setting that makes both errors zero. The craft is choosing the operating point deliberately, and knowing exactly what it costs on each side.

What is the hardest part of this work that people outside it never see?

The adversary adapts. That is what makes this different from most engineering. When you fix a bug, the bug stays fixed. When you close a fraud vector, the people on the other side study what you did and route around it, sometimes within days. A control that worked beautifully last quarter can quietly stop working, not because it broke, but because the behavior it keyed on changed. You are not shipping a solution. You are entering an argument that never ends.

The moment that nearly goes wrong is always the same shape. You launch a change, the bad actors adapt, and their new behavior starts to look, to your system, a little more like normal users. Now the easy way to catch them also catches real people. That is the knife edge: chasing the adaptation without dragging legitimate users into the net. I have had to take a detection approach that was catching more scammers and turn it down, because the damage to real accounts was climbing faster than the wins. The aggressive move would have looked better on a fraud dashboard and been worse for the platform.

You review and sit on the program committee for IEEE ICA 2026. What does judging other people's research change about how you work?

It makes you ruthless about what a result actually shows. When you review a paper, you are not asking whether the idea is exciting. You are asking whether the evidence supports the claim, whether the method would survive someone hostile poking at it, and whether the authors were honest about what they did not prove. You read a lot of work where the headline is strong and the foundation under it is thin, and after enough of that you start applying the same suspicion to your own systems before anyone else can.

It also keeps me current in a field that moves quickly, and it is a reminder that rigor is not something one company defines in isolation. Sitting with other researchers and arguing about what counts as solid evidence is how a field stays honest. I bring that back as a simple habit: assume a sharp skeptic in the room, or the version of me 6 months from now, is going to ask why I trusted a given signal. If I cannot answer that cleanly, I am not finished.

You argue safety and growth are the same decision. Most companies treat them as a tug of war. How do you actually change that?

You change it by refusing to let either side own a metric alone. When the safety team is measured only on harm reduction and the growth team only on signups, they will pull in opposite directions forever, and whoever has more organizational weight that quarter wins. The fix is structural and unglamorous: put both costs in front of the same people at the same time, so the conversation stops being safety against growth and becomes where exactly do we want to sit on this curve, given what each mistake costs us. Once a leader can see that over-blocking is quietly bleeding legitimate business, the framing changes on its own.

The reframe that actually lands is that trust is the product. People will not transact on a platform they think is full of scams, and legitimate businesses will not build on one that treats them like suspects. Get the balance right and safety is what makes growth possible in the first place. That is a hard commercial argument, not a soft one. It is the difference between a platform people bring their real business to and one they abandon the first time it wrongly slams the door on them.

As more of these decisions get handed to AI, what worries you, and what are you focused on?

What worries me is speed without judgment. An automated system will enforce a rule millions of times before a human notices it was wrong, and if that rule is quietly biased against a certain kind of legitimate user, you have scaled a mistake to a whole population before anyone feels it. The same automation that lets you stop scammers in seconds lets you wrongly punish real people in seconds. Handing more of this to models does not remove the trade-off I have been describing. It makes getting it right upfront matter more, because the cost of being wrong compounds faster.

So I am focused on building the guardrails while the autonomy is still being designed, when they are cheap to add. That means deciding in advance what an automated system may do on its own and where a human still has to look, and making sure a legitimate person caught in the machinery has a fast way back out. No platform is ever perfect. What I care about is one where mistakes are rare and never fall hardest on the people who did nothing wrong. Get that right and everything else, the growth and the trust, follows from it.