Trust & Safety is one of those disciplines where the stakes are high, but the work is rarely visible - unless something goes wrong, When it goes well, nothing happens - and that's the point.
Deniz Atabey Alkan, Checkstep's new VP of Trust & Safety, has spent over a decade inside some of the most complex content environments on the internet: Meta, TikTok, Snap, Bumble, and Trustpilot. She's built and run T&S teams at scale, navigated the tension between speed and accuracy, and been on the receiving end of the vendor relationships she now helps shape from the other side.
We sat down with Deniz to talk about what she's learned, what the industry's still getting wrong, and why the gap between platform-side experience and vendor-side impact is smaller than most people think.
1. You've built your career across some of the most consequential platforms in T&S - TikTok, Snap, Bumble, and most recently Trustpilot. What drew you to Trust & Safety as a discipline, and what's kept you in it?
I honestly stumbled into it by accident, similar to many others in the industry. I landed a part-time gig at Snap doing localization and media scanning for Turkey because I was bilingual. That got me noticed when I applied to Facebook a couple months later. When I joined their Community Operations team, the first six months were pure content moderation to get a good grasp on the various issues in the market.
What kept me in T&S was the immediate sense of purpose. You're quite literally catching predators before they reach the next victim and you can see that these are real people, they're not abstract. Once you've sat in those queues and understood what the stakes are, it's very hard to do something else.
2. You've worked across very different content environments - social video, dating, and review platforms. What's the single biggest lesson you carry from one that the others hadn't fully learned?
Every company I've worked at has had a different organizing principle for T&S. At Facebook and TikTok, it was scale, where you're operating on a volume that makes efficiency the only viable approach. At Bumble, the focus shifted entirely to member safety and preventing real-world harm to the individual. At Trustpilot, it was the integrity of the platform, making sure reviews were trusted and how mistakes impacted the reputations of the businesses being reviewed.
What I've learned is that whichever one you optimize for, you will find its limit.
Scale-first thinking treats edge cases as acceptable loss. For example, if you're getting 1% wrong out of 100 million decisions, most people would look at that as a success. The problem is that in T&S, edge cases are often the first sign of a new attack, or a gap in your policy that someone is exploiting. By the time it shows up in your metrics, you're behind. But swing too far the other way, and every edge case becomes a lengthy policy discussion. Processes slow down, and you end up with something that works beautifully for unusual scenarios and falls apart under real volume.
The balance between the two is genuinely the hardest line to walk in this job. There's no formula. You're always making a judgment call about which signal matters most right now, and that judgment has to be informed by both sides at once.
3. Trustpilot is also a Checkstep customer, so you've seen our product from the other side of the table. What does that experience tell you about what platforms actually need from a T&S vendor - versus what they think they need?
When we say "platform" here, I think it's worth separating T&S teams from the non-T&S decision makers in the room, because they're not the same conversation. Most experienced T&S professionals understand their tooling well. The gap I've observed tends to sit at the organizational level, where the AI conversation has picked up enormous momentum and there can be pressure to move quickly toward implementation before the foundational work is done. T&S teams have been working with machine learning models long before it became accessible to the public. We're the teams who trained them in the first place. But that organizational pressure means the normal T&S process may get compressed or skipped entirely. That's not a criticism; it reflects how fast the space has moved and how much genuine opportunity there is.
So from that perspective, what a good vendor actually delivers is threefold. They come with built-in processes and tools that lighten T&S teams' operational load, freeing them up for the work that genuinely needs their judgment. They give T&S teams direct control over policy, detection, and moderation workflows so a minor change doesn't require a lengthy engineering cycle to implement. And they act as a genuine thought partner on the foundations: making sure policies are consistent and well-defined before building any detection on top of them.
4. The threat landscape has shifted significantly - coordinated inauthentic behaviour, AI-generated fake reviews, deepfakes in dating apps. What's the trend you think the industry is least prepared for right now?
I feel that we're still relying on outdated fraud patterns now that AI-generated user behavioral histories have become a reality. We've spent a lot of time improving detection for AI-generated content, which is not the main indicator for a “fake” user anymore - everyone uses AI to some degree! What I think most platforms aren't ready for is coordinated inauthenticity where the behavioral fingerprints also look real. Attacks where not just the content but the account history, the engagement patterns, the timing of the activity seem benign because they've been generated specifically to pass behavioral screening. At that point, the signals we've traditionally used to distinguish organic activity from coordinated behavior start to break down.
5. There's a growing conversation about the "arms race" dynamic in content moderation: bad actors adapt as fast as platforms improve. Do you think the industry is winning that race, and what would it take to actually get ahead?
Maybe “winning” isn’t the right frame here, I feel it implies a finish line that doesn't exist. But I do think the industry is often playing the wrong game. For years, we’ve been talking about how to be more proactive rather than reactive, but our proactive measures are optimized for yesterday's attack signals.
Getting ahead of bad actors would require sharing more resources across the industry. We do this to some extent when it comes to child safety or NCII for example, but not across the board. Bad actors move across platforms. Someone running a coordinated campaign on one review site is almost certainly doing it elsewhere. Platforms that build genuine cross-industry signal sharing beyond the existing coalitions, will be able to adapt to new threats much faster than platforms trying to do it alone. The challenge is that some companies think signal sharing means admitting you have a problem, which nobody wants to do publicly. That tension hasn't been resolved, and until it has, everyone's operating on a time delay. One of the benefits of the T&S vendor ecosystem expanding over the past couple of years is that these types of cross-industry signals are built into the services vendors provide.
6. AI-powered moderation has gone from promising to table stakes very quickly. Where do you see it genuinely transforming T&S operations, and where is the hype outrunning the reality?
Where I feel the hype is when it comes to calibrating model performance. Accuracy scores don't always equal operational effectiveness. A model can perform well on your test set and still fail in production because it's enforcing policies that your own team interprets inconsistently. The model will learn the inconsistency and scale it. Precision and recall look right, and then you put the model into live queues and see that reviewers are overturning half the decisions. This isn’t because the model is “wrong”, but because the policy was never clear enough to operationalize in the first place. That's not an AI problem, it’s a policy problem that AI exposes.
7. "Human in the loop" is a phrase that gets used a lot, but in practice it means very different things. What does meaningful human oversight actually look like in a high-volume moderation environment?
Human oversight starts before the automation even runs. If two people on the same team would classify the same borderline case differently, no amount of automation can fix that down the line. Automation scales both your high quality and low quality labels, so the first job of meaningful oversight is making sure the policy is solid enough to automate against.
From there, it's about where humans sit in the moderation system. Clear, high-confidence violations or benign content can be automated, gray area cases can be routed and prioritized using automation while having humans make the final decision. The mistake I see most often is trying to automate gray. You end up with a faster machine for making bad decisions.
But the most underinvested part is what comes after moderation where your reviewers are doing genuine pattern analysis, not just auditing individual cases. Standard QA sampling will tell you about average error rate but it won't catch a systematic failure, a new attack type your model hasn't seen, or a policy gap someone is quietly exploiting. Meaningful oversight catches those things early, feeds them back into the model, and closes the loop before it grows.
8. There's sometimes a tension between protecting moderator wellbeing and maintaining quality at scale. How have you approached that in your past roles?
Having started my T&S journey purely focusing on content moderation for 6+ months, I don't approach this from the outside. I always keep this experience with me, knowing how frustrating that extra mouse click can be, or the shock you get after getting an intense piece of content in your queue.
The tension gets framed incorrectly in most discussions. It's presented as a trade-off where you have to choose between protecting your moderators or achieving high quality or efficiency. What I've found is that the platforms where moderators are least protected also tend to have the worst quality. Turnover is high, institutional knowledge walks out the door, and you're constantly training people who don't stay long enough to get good at interpreting your policies at scale. The cost of wellbeing infrastructure is a lot smaller than the cost of running a function where your most experienced people leave every couple of months.
In my opinion, the best approach here is to give your moderators control and agency over their own wellbeing. Moderation tools should have multiple wellbeing features built into it that are controllable by the individual moderator. People have different comfort levels and the tools they work with should reflect that.
9. You've spent your career on the platform side - building and running T&S teams directly. What made you want to make the jump to a vendor, and why Checkstep specifically?
I don’t see this as a big jump to the vendor side of T&S, I feel the distance between the two sides collapsed for me before I ever joined. I’d been a Checkstep customer at Trustpilot, watching the product develop over two years, giving feedback to the Checkstep team about what we needed and seeing a lot of that feedback finding its way into the product.
When I first encountered Checkstep, I recognized the interface immediately. It's built on the same mental model as the internal tools I'd used at Facebook, and I'd spent years working on how to make these tools work better in the moderation environment. At some point it stopped feeling like I was working with a vendor and started feeling like I was watching something I cared about get built without me in the room, so the move felt obvious. I believe the product is genuinely useful in a space where a lot of customer service products get re-purposed into moderation workflows, and that just doesn’t enable T&S teams to perform at the right level.
10. What do you hope to bring to Checkstep's customers that only someone who's been platform-side can offer?
Platform-side practitioners know what a 2am escalation looks like. They know what it feels like when a policy decision made on Tuesday is generating a hundred edge cases by Thursday, and none of the documentation accounted for it. They know the difference between a process that looks clean in a spec and a process that holds up under volume.
There's also something harder to teach: knowing how mature a T&S function is within the first conversation. I can usually tell pretty quickly whether a team has a dedicated policy person, whether they've thought through their appeals process, whether their leadership treats Trust and Safety as a cost center or a strategic investment. That read changes everything about how you approach the relationship, what questions to ask, what to prioritize, what risks to surface. It’s quite hard to develop that level of insight from the vendor side, you get it by being on the other end and seeing how decisions are made throughout the years.
What comes through clearly in talking to Deniz is that the best T&S work doesn't happen in spite of operational pressure. It happens because someone in the room has felt that pressure firsthand and built something that holds up under it.
That's the perspective she brings to Checkstep's customers: not a theoretical framework for what good looks like, but a practitioner's instinct for what works, what scales, and what actually protects our digital world.
Wondering how your Trust & Safety operations stack up? Book a 1-1 with Deniz.
Whether you're building the function from scratch, navigating a period of rapid growth, or just want a second opinion on where the gaps in your moderation strategy might be, Deniz is offering no-obligation consultations to help Trust & Safety teams get an honest read on where they are - and where to focus.
If you'd like to take advantage of this limited time offer of a 1-1 with Deniz, click the link below to schedule a time to meet.
