Episode 20
Safety as a Growth Driver
with Jeff Dunn, COO of the Digital Trust Council
Show Notes
Jeff Dunn is the COO of the nonprofit Digital Trust Council, and is a long-time trust & safety leader, advocate and mentor. He has been Hinge's VP, Trust & Safety and Chief Risk Officer, and spent 7+ years in Trust & Safety at Google, with earlier work at ICANN and Harvard on child protection, election integrity and platform operations.
He co-founded the Trust & Safety Network, which provides free resources for the people who keep everyone safer online.
Chapter Timestamps
- 00:00Introduction
- 5:01Building a stronger and more consistent trust-and-safety community
- 8:44Safety as a business-growth driver, not merely risk mitigation
- 13:53Dating-app safety: moving beyond verification toward proactive prevention
- 16:42How scammers behave on dating platforms and how product design can help
- 20:00Cross-platform scams and the need for industry signal sharing
- 25:09AI for online-to-offline safety—and the privacy trade-off
- 27:31DoorDash example: obvious signals and the business value of trust
- 29:38AI-enabled voice fraud, targeted impersonation, and media literacy
- 33:21Location data: essential for safety but valuable to scammers
- 34:58User reports, offline incidents, and human review on dating platforms
- 38:33Regulation, transparency, and competing on measurable safety outcomes
- 44:24Trust-and-safety careers: salary trends and the return of liberal-arts skill
Resources & Links
- Jeff Dunn
- Digital Trust Council
- Trust and Safety Network — LinkedIn and trustsafety.net
- Rob Leathern
Transcript
There may be transcription errors: we apologize for those in advance.
Rob: Hi, welcome to Won't Fix. My guest today is Jeff Dunn. He's the COO of the Digital Trust Council and a longtime trust and safety leader at Hinge, Google, ICANN and Harvard. He also co-founded the Trust & Safety Network. As you'll hear, Jeff gives back a lot to the trust and safety community. He posts trust and safety jobs regularly, and his Trust & Safety Network is a great way to connect with others working on these difficult problems. I really enjoyed our conversation, and I think you will too. As always, please send me your feedback. Jeff, how are you doing?
Jeff: Rob, doing great. Thanks for having me.
Rob: It's good to see you. I was thinking of you recently when I talked to Becky Holmes about romance scams. I don't know if you had a chance to listen to that, or saw her book about Keanu Reeves, which was quite interesting. So why don't we start with you telling me a little about what you're doing and what you've done. I know you originally from the trust and safety community. Tell us more about your professional journey.
Jeff: Thanks for having me on. I can talk for days about that, but I'll give you the 30-second version. Back before it was even called trust and safety, I was at a startup that I built, ran and sold. That was really exciting. It was around education. On the day Steve Jobs announced the iPad, I thought, this is going to play a big role in education. So I created a startup around helping teachers learn how to use technology. That's so quaint looking back on it now, but here we are. From there I went to Harvard, then to ICANN, the nonprofit that governs the internet, where I got a taste for regulation. I liked the taste for a while, and then it started to sour, just because it was taking me all around the world and I was pretty busy. While I was there, a recruiter reached out and I moved over to Google. They had this fancy new team called Product Quality Operations, which tells you nothing about what the team actually does. I remember the day, probably 10 or 15 years ago now, when we decided to name the team Trust and Safety. We even had a whole newsletter called The Ampersand, because there was a huge debate: should it be "Trust and Safety" with the word, or "Trust & Safety" with an ampersand? These are the important things I remember. At Google I bounced around setting up all sorts of teams that are now industry standards for trust and safety. I learned a lot. Then I moved over to Hinge, where I've been the chief risk officer and SVP of operations, largely making sure Hinge can be as safe as possible for daters. Romance scams are part of the reason I wanted to come on and chat with you. They continue to be top of mind and have been for quite a while.
Rob: Absolutely. Tell me more about the Digital Trust Council and what that's all about.
Jeff: Oh yeah, I buried the lede. I'm currently the chief operating officer of a nonprofit called the Digital Trust Council. For those in the US who know it, we're developing basically the LEED-style certification for digital trust. It's called the Digital Trust Certification, and it will certify everything from AI models to digital products and services. Think of it as a Consumer Reports for this. You can pull up a simple website we're actively building and look up a product you're considering, whether you're a parent or a professional doing procurement, or just looking for something for your son or daughter. You can see what we found about it. We're also aggregating information from a lot of partners, so you can see what 100 different red teamers and 200 different certifying bodies think about these products and models. Over time, we'll offer certification, so companies can pay to have their models and products certified. That will be a really robust effort, and we're already designing all the operations for it. That's why I'm there. It's been really exciting. We just held a summit at Harvard a month or two ago with a bunch of world leaders and fancy people. It was hugely successful and helped us change the course of what we're up to. More to come on that.
Rob: It's interesting you mentioned naming trust and safety teams. Since I worked on the business integrity team at Facebook, I often have to explain what that is. If I say it's what a lot of other companies call trust and safety, it makes more sense to people. Funnily, soon after I joined, we talked about changing the team's name, because we wondered whether people really understood what business integrity was. Then I got a message saying we couldn't change it, because other teams had now been named "something integrity" too. So our chance to change the name was gone.
Jeff: I've said this to many people, on my team and elsewhere: the hardest part of trust and safety is naming things. It's not necessarily keeping all the bad actors off. When you work at a big company, you realize every single team has already taken your name. Every project has already used it. These are the problems we face.
Rob: Exactly. I used to say that what we did was more trust, and what the other teams did was more safety. We had all kinds of conversations like that.
Jeff: And you're like, I want to be the safety team. I'm pretty sure we all should be.
Rob: One thing I appreciate about you, Jeff, is the regular job postings you share for trust and safety roles. You're very active in helping build a community around trust and safety. Outside of the Digital Trust Council, talk about some of the things you're doing or have done, and why.
Jeff: I mentor about a dozen people at any given time, and honestly, I do it because it fills my cup. I don't charge anybody a penny. Giving back has helped the people I mentor directly, but it goes further. By sharing jobs and creating trustsafety.net, the T&S Network, and other things, I've been trying to push or pull the industry forward, depending on your perspective, so it matures a little faster. Frankly, we need more perspectives in the room. Like you, I'm from the early days of trust and safety, when it felt a little Bay Area–centric and a lot of policies and products were developed for people down the street in Palo Alto. Now trust and safety work protects billions of users. And it's not just one product with billions of users; there are countless products with billions of users. So all those products need to work better together. Google, Meta, TikTok and smaller companies need the same standard approach to trust and safety, so they can have fungible policies. That way you as a user — not Rob the trust and safety genius, but Rob the general user — aren't surprised when you join a new platform and its trust and safety is all over the map. What a mess: they're not secure, they don't do ID verification, they don't do all these things. If we have best practices and more perspectives joining the industry, everybody wins. That's why I share jobs. I aim for every Tuesday. I typically manage every other Tuesday at best. I'd give myself a B for effort.
Rob: I'd give you an A for effort, but your standards are very high, which is great. I remember one of the engineering leaders I worked with saying that trust and safety, integrity, whatever you call it, used to not be seen as a primary area, and then it became far more important. It became more technical. Machine learning and other AI technologies became central to it. So it's really evolved. I've compared it before to security. Now most companies have a chief information security officer, and I remember when that used to just be called IT. There's been a huge evolution, and I can only see this discipline becoming more important. It's about figuring out tradeoffs around risk, for both the business and users. What percentage of revenue are you willing to spend to protect users proactively versus reactively? Users versus corporate customers? Does that reflect the changes you've seen over the years?
Jeff: That's very similar to a conversation that, frankly, led to me becoming a chief risk officer. I couldn't agree more that this industry is growing and needs to keep growing. But it's also changing. A couple of things. First, AI safety and AI governance are a small to medium-sized chunk of trust and safety, and they're going to eat our lunch. If AI models eventually get good enough to replace frontline moderation, appeals, even escalations, then I don't see why trust and safety doesn't become AI safety that happens to do customer support and a few other things on the side. But whether you look at trust and safety today or whatever it looks like in a few years, it's not just about managing risk for the business. It's funny that Facebook called it business integrity, because that's just about protecting the business. I have a hot take, which I've talked about with others: safety is a growth driver, not a risk mitigation function. It's not just about lowering the money you spend or lose on cybersecurity incidents. Look at companies I've worked at, like Hinge and even Google. You probably remember the adpocalypse. The ads ecosystem was a disaster for quite a while, because the ads were downright dangerous. Invest in safety, and suddenly advertisers feel it's safe to advertise again. That's long in the past now, but if you make the platform safe for all your users, advertisers and developers, the business will grow, almost guaranteed. You still need a decent business that does something useful. But if you invest a proper amount in safety, there's a pretty clear line between safety and growth.
Rob: I think it's important for growth teams to understand trust and safety and user protection teams, and vice versa. There are lots of interesting ways to do that, and I think it's still a real opportunity inside some of these companies. There's a lot more work to be done, and empathy-building on both sides of the house.
Jeff: What's worked well in my experience is 20% projects, as Google called them, plus mentorships and shadow sessions. At Hinge I had product, operations and everybody in trust and safety, but they weren't there to make money, necessarily. Their OKRs weren't "make X dollars this quarter." They were "protect revenue" at best, but mostly "protect users." The growth teams, on the other hand, are all about making money. The best approach is to have them sit really close together, so growth people adopt a safety mindset and safety people adopt a bit of a growth mindset. You can't do that if they're not in the same meetings, the same rituals, the same office. I see it at big companies a lot: that's the island of safety over there, and don't talk to the ads team, which is on that other island. No sharing notes.
Rob: I think that's right. There needs to be collaboration, undoubtedly. Let's talk about romance scams and dating. What's important here, and what are the more recent trends? I know there's been a lot of discussion about ID verification, age verification, AI filtering and so on. What's the state of the art in online dating safety?
Jeff: I actually don't think ID verification and age assurance are where we should spend most of our time these days, because they're becoming the floor, as they should. These tools have existed for quite a while. They're proven effective, regulations require them, and everybody's doing them. You won't differentiate your business by doubling down on age verification. It's a must-do — certainly not a won't-fix — and you need to keep investing in it no matter what. Where you can win users and grow your business is by getting ahead of scammers and policy-violating actors, or bad actors, whatever you want to call them. Basically, the people who want your money, your time, your attention, your technology and your devices. So how do you get ahead of them? We did a number of things in my Hinge and Google days that I'm happy to talk about. Some of the biggest, going back to integrating with growth, have been getting safety signals into the algorithms. If you're bailing water with a leaky bucket as reports rack up and bad actors pile in through some loophole they hope you won't notice for a while — and that definitely happens — you can play whack-a-mole and keep investing in stopping them that way. But at some point, you need to go upstream, as far up the funnel as possible. What we've found at a few different companies is that the biggest wins come from integrating safety signals, like a graph of user information, into the recommender itself. The recommender almost does the job for you, instead of a giant filter at the front door with a bunch of rules and signals to track. When you build safety into the product, that's when the magic happens. That's been the biggest win I've seen over the past year or two, at Hinge and at other companies. For anyone who wants more detail — certainly not on a broad podcast — I'm always happy to share best practices on what kind of machine learning and AI to use.
Rob: Absolutely. Is there a high correlation between signals from someone who's not a scammer but is acting in bad faith — exploiting the system to get lots of matches, say — and what scammers do? Scammers are also trying to get lots of matches, though obviously for a very different purpose. Do many signals work for both: preventing platform abuse, which is a bad experience for other users, and stopping actual scams and fraud?
Jeff: On one hand, dating apps are full of people who are more open than usual because they want to meet their significant other. They put themselves out there and share a lot of information. That makes dating apps a honeypot for scammers. On the flip side, dating apps are actually pretty terrible for scammers. It takes a lot of effort. You can't just broadcast one TikTok and have everyone click a link you shared. You have to methodically talk to each individual and get to know them. It's hard, and it should be downright impossible. At Hinge, I'm pleased to say we've built the product to be potentially the hardest for scammers to achieve their objectives. But where there's a will, there's a way, and AI makes things easier to scale. It used to be pick two of three: fast, cheap or good. Now you can pick all three and reach many people with personalized approaches. That's where trust and safety teams have to find the tells, and there are still a lot of tells. I won't say what they are, because that's the secret sauce. But it's safe to say a scammer's behavior is quite different from a typical user's. Scammers are typically power users. They may set up dormant sleeper accounts on a brand-new phone or device and let them sit for weeks or even months, so that when they turn one on, it looks like an account in good standing. That's a pretty standard scammer approach. Once they turn it on, they're power users, and there's a burst of signals. Throughout, there are always signals and patterns we can detect. As long as product and safety teams are talking, you can build an experience that helps the average user, who is obviously the vast majority. Ghosting is a big issue for good-intentioned users. You send out a bunch of likes and hope to hear back, but no one responds. So build functionality that makes it worthwhile to respond before you get the next one: your-turn limits, match limits, those kinds of things. That also does a lot to stop scammers.
Rob: That makes sense. One question I often have, though it's tricky to figure out and you may not know: are the same scammers running similar scams across multiple platforms? Sometimes it's the same scams with different scammers, sometimes both. And I don't just mean other dating apps. Are they crossing into different modalities, like buying TikTok ads to run romance scams the way they would on a dating platform? Are they multi-platform, or highly specialized, going deep on whatever works on one platform?
Jeff: In my 10 to 15 years in trust and safety, I've seen that scammers and bad actors run some of the most efficient businesses you can imagine. They spend a lot of time figuring out what works, then double and triple down on it. So yes, they operate on many platforms, but they have scarce resources to balance, like any business. The approaches need to differ by platform, and that used to be expensive. Now, with AI, they can set up agents and bots that match with people, wait a random amount of time, analyze a profile and then say things that feel authentic and real. That's why we need more liveness detection and new ways to confirm someone is the person in the pictures and the person sending the messages. It's a big challenge. The scams that come through dating apps look extremely similar to marketplace scams. They're a little less like ad scams, because ad fraud and bad ads, as we called them at Google, are a scale game: you hope some percentage stays up. On dating apps, scammers have to invest a lot more in each attempt.
Rob: That said, I've recently seen a bunch of ads that look a lot like people trolling for a relationship. Some may be real. Some look like sexual solicitation, but it's hard to tell. Often the call to action is "message me," either directly in the ad or by asking you to reach out on Telegram, WhatsApp or some other so-called private platform. So in some cases, the algorithms seem to make dissimilar platforms serve similar goals, and scammers can apply experience from one to the other. But I don't have data on who these actors are or whether they're the same. And as we both know, scammers copy each other. When they keep seeing something, they assume it works and try it. One way to catch scammers is that they keep repeating what works, but that also makes it hard to tell different groups apart.
Jeff: On dating apps, my home base, we have a bunch of signals and best practices. But wouldn't it be great to have the knowledge the ads teams at Google and Meta have? And wouldn't it be great if all the marketplaces shared theirs? I'd love to see signal sharing done properly at the industry level. It would need to be nonprofit- or NGO-driven; I don't think government necessarily needs to be involved. Scammers learn from each other in real time, and platforms should too. Instead it's: we messed up, someone got through, we fixed it, and we didn't share how with any other company or even another team. Going back to our discussion about islands: the safety island is fighting fires, and growth is chasing revenue with blinders on. Now take that to the company level. Companies aren't sharing signals. There are a few really strong efforts, like the anti-scam coalition that Match, Snap and a few others started. But we have a long way to go before we can respond in real time as an industry, not just as a bunch of separate companies.
Rob: One other thing that makes dating tricky, and interesting from a safety perspective, is that like ride sharing, it's both an online and an offline safety problem. I once talked to someone at Uber, where they'd patented an algorithm that used phone signals to detect whether you were drunk and might wander into the street and get hit by a car. That's a very interesting application of AI and machine learning to a safety problem, though maybe not the obvious one. Have you seen important AI applications in this online-to-offline area, or ones we should be thinking about?
Jeff: Now I'm so curious. Are they just using the accelerometer to measure your gait and wobbliness?
Rob: It's in a patent somewhere, so in theory you could look it up on Google Patents.
Jeff: I could have AI look it up and figure it out. Even easier.
Rob: That's right.
Jeff: To answer your question, AI will play a huge role here. It would be folly to say otherwise. We all collect a lot of signals as platforms and have set up many machine learning rules to analyze them. For lack of a better term, LLMs and other models make it easier to join all the different data streams and data types. So it's not just "you're a little off course" or "you've been stopped for a while in traffic, tap to confirm you're fine." It could be: we noticed you were walking at this pace toward your Uber, but now you're walking very differently and in an unexpected direction, and we know your home is over here. That definitely borders on the Big Brother creepy factor. Privacy and safety are a constant balance. But AI can play a big role in analyzing additional signals and giving helpful guidance, as long as it's not creepy. That's the trick.
Rob: Sometimes I wish we had simple uses. Here's an example. This past week I ordered DoorDash and paid the extra $2.99 for direct delivery, so the driver wouldn't stack orders. Then I noticed the driver had messaged me saying they were dropping off another order first. Maybe the AI should read that message and see that they're not allowed to do that. Combine it with a very rare occurrence: this driver had no ratings whatsoever.
Jeff: Not off to a great start.
Rob: I checked in to see why it was taking so long. It was 40 minutes late, and the driver was stopped at some other location, obviously doing another delivery. These signals seem so obvious. Why not send an alert to the driver? Or maybe they do, and the driver just ignores it.
Jeff: It reminds me of the good old days when delivery meant Chinese food or pizza, end of list. "You'll get your pizza in 30 minutes or less," and people chose a business because they knew they'd get it in 30 minutes or less. But if you're Rob ordering on DoorDash, you're now thinking: I paid extra and got it even slower than expected. I'm going to try Uber Eats. I'm going to try Caviar. They all own each other, but there are other options. That's a good example of trust and safety as either a growth driver or a loss leader, depending on how you view it.
Rob: It's interesting. I agree we'll see these technologies show up in more and more places and become baked into our expectations. But these AI tools will also be used to cause harm. I was talking to law enforcement recently, and there are lots of demos of people changing their faces in real time — I'd love your perspective on that. But they said they're seeing much more use of voice technology to leave convincing voicemails that suggest a real person just missed your call. The scammers weren't necessarily attempting live calls, where they could get tripped up in a number of ways. It was a planned message that sounded legitimate, with the same voice from one instance to the next. I'm curious what you've seen, both the cool-but-not-great uses of this technology and the simpler ones where AI plays an unfortunate role in scams.
Jeff: Fast, cheap and good, right? It used to be pick two. Now it's pick three, and if you want a fourth, no big deal. We're already there. To borrow from the past: it used to be, oh my goodness, Google launched, how will we ever learn anything? Now it's, AI launched, how will we get the information we want? With rapid advances in voice, it's gone from a novelty to something super easy to use to fool people. It used to be quite hard to figure out, Rob, where you live or what your favorite things to do are — things only friends or family would probably know, but that you've likely shared somewhere online. If I were fanning out a bunch of agents, I'd say: Rob is my target. Let's find where he lives, which is pretty easy, then his resume, his family members, the food he posts on Instagram, the events he's been to recently. Then put together a bunch of compelling texts to all his friends and family, each with its own angle: "Hey, this is so-and-so. I'm here with Rob, and he's really pissed about his DoorDash." With all those details, your partner might know you ordered DoorDash right then and think, yeah, this checks out, it must be real. And it's a scammer the whole time. That's a random example, but it's never been easier, unfortunately, to stalk someone. We need much more online education and media literacy so people know how to detect this. At this point, I don't pick up the phone. If there's a voicemail, I look up the number to make sure it's real, and I check with my partner and friends before acting. Maybe I'm paranoid, but being on the front line and seeing scams every day makes me not very interested in picking up the phone or answering random emails.
Rob: I think that's right. Many years ago, I had a public website where I posted my location, somewhat manually, so colleagues could see I was in the New York office that day. I wouldn't do that today. It seems like a bad idea.
Jeff: Your Foursquare account is alive and well somewhere. I can't imagine something like that doing well these days.
Rob: Foursquare, yes. I know Dennis Crowley quite well; we worked together at Jupiter many years ago. I'm a longtime fan, but it also became clear how challenging it can be for people to have this kind of information out there.
Jeff: On the flip side, knowing where someone is can be a huge safety win, whether you're out on a date or your child is on the school bus home and you want to make sure it's not going off course. It'd be great to have an AI model that says, hey, the school bus went off course; it's either delayed or really off course, and you should get involved. Those are serious, big safety wins. But having your or your child's real-time location available all the time is a honeypot for scammers too.
Rob: One thing I always find interesting, since I talk to regulators about trust and safety topics all the time: you and I have worked at big companies with the narrative that if users see something bad, they can report it anytime. Bad ad? Report it. In reality, some super-reporters are very good at reporting bad content. Others, to use the ads example, just don't like a particular ad, or ads in general, so their reports aren't very useful. I think LLMs are already making user reports more useful, and will keep doing so, because they can separate the wheat from the chaff much more effectively. In dating, though, I imagine the dynamics around user reports and people causing problems are very different. How do you think about user reporting? There's also brigading, where people on the platform get mass-reported. I'm sure that's an issue on the dating side too. How do you think about this side of trust and safety for in-person and dating-related issues?
Jeff: It's a big challenge for ride-share and dating apps, because it involves things you can never see. It happened offline. Someone had a terrible experience in an Uber or Lyft. Someone had a terrible experience on a date at a restaurant. You're taking one person's perspective, you try to get the other person's, and figuring out where the truth lies in the middle is objectively hard. There's no objective data to go on. That's the challenge. If someone is harassing another user on your platform, you can pull up the chat log and say, yep, sure enough, you're a terrible person. Depending on severity, that's a strike, a warning or a ban. But whenever you connect humans in person — which is really important, and frankly something we all need more of, not less — you risk losing a lot of information. I won't call it data. I'll take you at your word, Rob, but we might say you didn't provide enough information to meet our evidentiary standards. If you just write "this person is a jerk," what do you expect me to do with that? A detailed report of what happened, with supporting information, is at least more likely to be legitimate. But anyone can write a long report full of made-up details about things that never happened. Figuring that out is the tricky part. That's why at Hinge, and I know elsewhere too, we have humans in the loop for every single ban. Appeals are reviewed only by humans. I'm also proud that Hinge has had a 24/7 global escalations team, on standby all the time for someone going through the worst moments of their life. It's important to have that capability and mitigate harm as much as possible. But at the end of the day, platforms' hands are somewhat tied, because we don't have all the information we need to make the right or most accurate decision.
Rob: Have regulations like the UK Online Safety Act or the EU's Digital Services Act changed how platforms handle these issues, appeals, account suspensions and so on? What has meaningfully changed in the last few years?
Jeff: There are a few recent ones, so I'll lump them together: the AI Act, and going back to the DSA. There are many more reporting requirements, which is great. The VLOPs, the very large online platforms, now have to report much more information than they ever did. I was responsible for creating the Article 40 researcher-access database at Google, so I think it's very useful to have all that information out there. But is it changing how businesses develop products or run their business? I don't think so. I wish greater transparency were leading to safer products, but I don't know that we can draw that line yet. It's a great start. For example, there's now a requirement that if you matched and talked with a scammer whom the dating app then banned, the app has to notify you that the person you connected with wasn't real and was a scammer. Hinge has done that for years. Now it's law, so all dating apps have to do it, which is great. Will that improve safety? It's a good step in the right direction, and it's more transparency. But the real magic will come when product teams are incentivized for safety, not just slapped on the wrist, so safety becomes, as I said, a growth driver. There's a lot of money on the table for being the safest product in your category: the safest ride-share app, dating app, marketplace or bank, however you define it. If you can show you have great features and you're there to help people at a moment's notice, the world is your oyster. But that takes a ton of investment, and right now there's no clear definition of trust as a growth driver.
Rob: I totally agree, but it's a chicken-and-egg problem: no one wants to share the underlying data that would let a third party assess them. When you buy a new car, you get a J.D. Power survey or something similar, facilitated by the manufacturers. They've all accepted that they might not come out as the best car. There's a whole set of processes around quality and safety, with third parties doing crash testing, and the National Highway Traffic Safety Administration. My supposition is that at some point it'll become clear that companies have to compete on safety, make this information available, or both. I think that's a better outcome than lots of what I'd call process requirements, where a regulator says you must handle things a certain way. I'd rather require transparency so people can assess the actual outcomes on a platform. That creates self-correction and feedback loops that make things better. Knowing exactly how a platform processes things, or that it takes 48 hours to do X or Y — do I care? Does that make the platform safer? I have no idea.
Jeff: When you're downloading an app or standing in Best Buy deciding whether to buy something, if you're like most people, you're not trying to understand how quickly the company responds to DSA transparency reporting requirements. It won't change your opinion at all. And at big companies, all those numbers are huge simply because the platform is huge. One percent of 2 billion people is still a lot of people, even though it's a small percentage. So I totally agree. Getting companies to be transparent about how they operate and come into the light a bit more is just the start, not the end. The DSA is not the end of regulation. If anything, it formalized a lot of regulation that needed formalizing. We've seen this in cybersecurity, aviation and medicine. Those industries are much further along, much bigger and employ many more people, and they have far more standards, certifications, requirements and laws, like HIPAA. I think the equivalent is coming for companies that host user-generated content, including video, audio, podcasts like this one, and even lightweight ephemeral posts. It'll be a busy few years for many product people on the safety front. We already see it with AI. There's a lot of interest in how safe models are, so companies are investing heavily in whatever helps them feel safer. But AI is a bit of a black box, so who wins that race is TBD.
Rob: Building on what you just said, Jeff, and thinking about the jobs companies are hiring for in trust and safety and adjacent fields: given AI progress, which skill sets are most important now that weren't around, or weren't as big a deal, a few years ago?
Jeff: I've taken a data-driven approach to this, so I have some pretty strong answers. Alongside building the Trust & Safety Network, I've quietly been monitoring job postings for about a year and a half. I have a pretty large Postgres database of every trust and safety job posted in that time, and I've been analyzing pay — what companies actually pay — because that's very useful to a lot of people. Spoiler: it's been going down. The average trust and safety salary has dropped about 2 to 5 percent from a year ago. As for skills, I say this a lot to my partner and others: liberal arts is back. For a long time, everyone needed to learn to code. Boy, how times have changed. Now I hear people ask why you'd even bother learning to code, since Claude, ChatGPT and others can do it for you. That's kind of a wild thing to say. If you can't look at code and understand it, if it just looks like gibberish, you're probably limiting yourself. But that's a whole other conversation. The skills I see are liberal arts skills, and they're what I look for when hiring. I've interviewed and hired hundreds, if not thousands, of people, and I always test critical and strategic thinking. More specifically, I want to see how people connect dots. I don't give long case studies. I don't like making people do homework; I don't think it's fair to candidates. Instead, I ask them to connect the dots informally, pretty casually, in conversation. Say a scammer is attacking us with a man-in-the-middle attack. I listen for whether they say, "I'll go in and solve it myself," or show they're collaborative. Will they think about signals we haven't considered? Will they build their own solution and then share it with others? How much will they work with others versus figure it out alone? No one says "I'll do it all myself" directly, but you can infer it when someone insists they have all the skills and are the smartest, because they want to look like they know what they're doing. That screens as selfish and not a good collaborator. They're not connecting the dots, and they're not strategic thinkers. When people on my team ask whether they're ready for a promotion — everyone wants a promotion, or at least more money — my answer is usually that I want to see strategic thinking, especially if they want to become a people manager. I want them to identify a gap in our business plan, our strategy or our coverage of a certain harm vector, and develop a solution. Maybe it's bringing on a new vendor. Maybe it's more transparency reporting. Maybe it's hiring a whole team they want to run. Convince me. Show me you've spotted a real problem, that other people agree, and what your solution is. It doesn't need to be perfect, but I want to see that thinking. That takes a lot of liberal arts–style thinking — I went to a liberal arts school — rather than "I'll just code my way out of this."
Rob: I love that answer. I also have a liberal arts background. Well, I did computer science at a liberal arts school, so I guess both.
Jeff: Oh, both. I got one software engineering job offer early on, and I thought, I don't think I'd be a good software engineer. Thankfully I took a different path.
Rob: And we're all grateful for it. Jeff, this has been an amazing conversation. I really appreciate you taking the time to catch us up on what you've been working on. I'll share links to everything you mentioned in the show notes.
Jeff: Sounds good. I appreciate it, Rob.