Why I Spent Months Talking to Scammers
Scam Detection & Prevention
Most people go out of their way to avoid scammers. I went out of my way to find them.
Over the course of this research, I've directly engaged with more than 200 scammers, and turned that engagement into a working dataset: 175 detailed cases and 15,913 individual messages. This wasn't secondhand data pulled from a report. I was in the conversations.
The reason traces back to my time working on online safety at Match Group.
What the statistics don't tell you
Search for information on scams like these, and you'll find plenty of numbers. The FBI's Internet Crime Complaint Center (IC3) logged 23,159 confidence fraud and romance scam complaints in 2025, totaling $929.3 million in reported losses — a 38% jump from the year before. Victims 60 and older accounted for 63% of that figure. Investment fraud was IC3's single largest loss category in 2025 at $8.65 billion. Across all internet crime reported to IC3, losses exceeded $20.9 billion in 2025.
These categories can also obscure how scams actually unfold. In my own data, relationship-building and investment fraud frequently appear as parts of the same behavioral sequence. Cryptocurrency-based investment scams, including so-called "pig butchering" schemes, can begin with a stranger, a friendly message, and an extended period of relationship-building before money ever enters the conversation. And the same scammer, running the same playbook, doesn't stay confined to one platform or one opening line. They show up on dating apps, but just as often on Facebook, Instagram, X, and TikTok — and a large share of first contact isn't even a match, it's a cold message sent to someone who never opted into meeting a stranger at all. The channel changes. The underlying behavioral sequence doesn't.
Numbers like these tell you the scale of the problem, and they're worth sitting with. What they don't tell you is the layer underneath them.
How does a scammer actually build a connection? What do they say first? How fast do they escalate intimacy? When do they introduce money? And critically, what do they do when the target starts to doubt them?
None of that shows up in a post-incident report. You only see it by watching the conversation unfold in real time — regardless of whether it started with a match, a follow, a friend request, or a cold message out of nowhere. So that's what I did, one exchange at a time, until the dataset became large enough to analyze systematically.
Scammers work from a playbook — literally
Along the way, I collected something else: the actual scripts scammers use. I now have 11 of these playbooks.
They aren't just lists of phrases to send. They lay out a process: how to build a relationship, how to earn trust, how to work on someone's emotions, when to bring up an investment, and — this is the part that surprised me — specific instructions for what to say when the target starts getting suspicious.
When I compared real conversations against these playbooks, the overlap was striking. Scamming isn't just improvised charm. It's a reasonably systematized behavioral process, and the messages I collected tracked that process closely.
Keyword detection misses the point
This has real implications for how we think about fraud detection.
Scammers don't open with "please send money." They open with something disarmingly ordinary: "How was your day?" "Did you eat yet?" "I really enjoy talking to you."
Taken as individual sentences, none of that looks suspicious. Read as a sequence, a different picture emerges: trust is built, intimacy is established, personal information is gathered, expectations about the future are shaped, and only then does money enter the frame — gradually, then directly.
If you're only scanning for suspicious keywords or urgent money requests, you'll miss almost all of this. The signal isn't in what's said. It's in how the relationship — and the target's behavior — is being shaped over time.
Telling someone "you're being scammed" rarely works
There's a harder problem underneath all of this: people being scammed often don't believe they're victims.
It can look obvious from the outside. Warnings from banks or platforms often don't change anything, because the target has already decided: "This is real." "We really love each other." "This investment is legitimate."
That's not naivety so much as the predictable result of a long con. The scammer has spent weeks, sometimes months, deliberately building trust. A single warning message isn't going to undo that, no matter how accurate it is.
The FBI has run directly into this same problem at scale. Through Operation Level Up, agents proactively identify people they believe are being defrauded through cryptocurrency investment scams and contact them directly, rather than waiting for a complaint to be filed. As of March 2026, the FBI had notified 8,935 people through the initiative — and 77% of them had no idea they were being scammed. The intervention is credited with preventing an estimated $562 million in further losses. Those aren't people who suspected fraud and asked for help. They were identified by federal investigators as active victims, and more than three out of four still didn't see it. That's the scale of the gap between what's obvious from the outside and what's believable from the inside of a relationship that's been deliberately constructed over weeks or months.
Which is why I don't think detection alone is the answer. What's actually needed is explanation — showing someone not just that something looks wrong, but what is happening to them: "This conversation followed a pattern of building intimacy, then introducing an investment opportunity, and it's now moving into a stage designed to prompt a financial transaction."
Building AI that explains, not just flags
I'm now using this dataset to develop a model that detects psychological manipulation and behavioral patterns within a conversation, not just risk scores.
The goal isn't an AI that outputs "this may be a scam." It's an AI that can explain, in terms a person can actually act on, how trust was built, what manipulation tactics were used, and what stage the conversation has reached — so someone has a real chance to understand what's happening, reassess the situation, and avoid further harm.
From scam detection to a broader question about AI and influence
This research isn't only about scams. As generative AI systems hold longer, more personal conversations with people, they will increasingly shape trust, emotion, and decision-making themselves — intentionally or not.
Where does persuasion end and manipulation begin? How do we evaluate — and prevent — AI systems that exploit cognitive biases or emotional vulnerability to influence a person's decisions? Studying how human scammers operate is, in a real sense, a starting point for answering that larger question about AI systems.
This work has been recognized as a partnership initiative under the UN's Global Dialogue on AI Governance.
Going forward, this series will draw on the scam conversation data, the scammer playbooks, and behavioral economics to unpack two questions: why people get deceived, and how scammers engineer that deception, step by step.
Before fraud becomes a financial loss, it is often a decision-making problem. That's why I study it through behavioral economics — and why the same framework matters for understanding AI's influence on human behavior more broadly.
Next in this series: breaking down the specific psychological mechanisms scammers rely on, stage by stage.
Why I Spent Months Talking to Scammers
Scam Detection & Prevention
Most people go out of their way to avoid scammers. I went out of my way to find them.
Over the course of this research, I've directly engaged with more than 200 scammers, and turned that engagement into a working dataset: 175 detailed cases and 15,913 individual messages. This wasn't secondhand data pulled from a report. I was in the conversations.
The reason traces back to my time working on online safety at Match Group.
What the statistics don't tell you
Search for information on scams like these, and you'll find plenty of numbers. The FBI's Internet Crime Complaint Center (IC3) logged 23,159 confidence fraud and romance scam complaints in 2025, totaling $929.3 million in reported losses — a 38% jump from the year before. Victims 60 and older accounted for 63% of that figure. Investment fraud was IC3's single largest loss category in 2025 at $8.65 billion. Across all internet crime reported to IC3, losses exceeded $20.9 billion in 2025.
These categories can also obscure how scams actually unfold. In my own data, relationship-building and investment fraud frequently appear as parts of the same behavioral sequence. Cryptocurrency-based investment scams, including so-called "pig butchering" schemes, can begin with a stranger, a friendly message, and an extended period of relationship-building before money ever enters the conversation. And the same scammer, running the same playbook, doesn't stay confined to one platform or one opening line. They show up on dating apps, but just as often on Facebook, Instagram, X, and TikTok — and a large share of first contact isn't even a match, it's a cold message sent to someone who never opted into meeting a stranger at all. The channel changes. The underlying behavioral sequence doesn't.
Numbers like these tell you the scale of the problem, and they're worth sitting with. What they don't tell you is the layer underneath them.
How does a scammer actually build a connection? What do they say first? How fast do they escalate intimacy? When do they introduce money? And critically, what do they do when the target starts to doubt them?
None of that shows up in a post-incident report. You only see it by watching the conversation unfold in real time — regardless of whether it started with a match, a follow, a friend request, or a cold message out of nowhere. So that's what I did, one exchange at a time, until the dataset became large enough to analyze systematically.
Scammers work from a playbook — literally
Along the way, I collected something else: the actual scripts scammers use. I now have 11 of these playbooks.
They aren't just lists of phrases to send. They lay out a process: how to build a relationship, how to earn trust, how to work on someone's emotions, when to bring up an investment, and — this is the part that surprised me — specific instructions for what to say when the target starts getting suspicious.
When I compared real conversations against these playbooks, the overlap was striking. Scamming isn't just improvised charm. It's a reasonably systematized behavioral process, and the messages I collected tracked that process closely.
Keyword detection misses the point
This has real implications for how we think about fraud detection.
Scammers don't open with "please send money." They open with something disarmingly ordinary: "How was your day?" "Did you eat yet?" "I really enjoy talking to you."
Taken as individual sentences, none of that looks suspicious. Read as a sequence, a different picture emerges: trust is built, intimacy is established, personal information is gathered, expectations about the future are shaped, and only then does money enter the frame — gradually, then directly.
If you're only scanning for suspicious keywords or urgent money requests, you'll miss almost all of this. The signal isn't in what's said. It's in how the relationship — and the target's behavior — is being shaped over time.
Telling someone "you're being scammed" rarely works
There's a harder problem underneath all of this: people being scammed often don't believe they're victims.
It can look obvious from the outside. Warnings from banks or platforms often don't change anything, because the target has already decided: "This is real." "We really love each other." "This investment is legitimate."
That's not naivety so much as the predictable result of a long con. The scammer has spent weeks, sometimes months, deliberately building trust. A single warning message isn't going to undo that, no matter how accurate it is.
The FBI has run directly into this same problem at scale. Through Operation Level Up, agents proactively identify people they believe are being defrauded through cryptocurrency investment scams and contact them directly, rather than waiting for a complaint to be filed. As of March 2026, the FBI had notified 8,935 people through the initiative — and 77% of them had no idea they were being scammed. The intervention is credited with preventing an estimated $562 million in further losses. Those aren't people who suspected fraud and asked for help. They were identified by federal investigators as active victims, and more than three out of four still didn't see it. That's the scale of the gap between what's obvious from the outside and what's believable from the inside of a relationship that's been deliberately constructed over weeks or months.
Which is why I don't think detection alone is the answer. What's actually needed is explanation — showing someone not just that something looks wrong, but what is happening to them: "This conversation followed a pattern of building intimacy, then introducing an investment opportunity, and it's now moving into a stage designed to prompt a financial transaction."
Building AI that explains, not just flags
I'm now using this dataset to develop a model that detects psychological manipulation and behavioral patterns within a conversation, not just risk scores.
The goal isn't an AI that outputs "this may be a scam." It's an AI that can explain, in terms a person can actually act on, how trust was built, what manipulation tactics were used, and what stage the conversation has reached — so someone has a real chance to understand what's happening, reassess the situation, and avoid further harm.
From scam detection to a broader question about AI and influence
This research isn't only about scams. As generative AI systems hold longer, more personal conversations with people, they will increasingly shape trust, emotion, and decision-making themselves — intentionally or not.
Where does persuasion end and manipulation begin? How do we evaluate — and prevent — AI systems that exploit cognitive biases or emotional vulnerability to influence a person's decisions? Studying how human scammers operate is, in a real sense, a starting point for answering that larger question about AI systems.
This work has been recognized as a partnership initiative under the UN's Global Dialogue on AI Governance.
Going forward, this series will draw on the scam conversation data, the scammer playbooks, and behavioral economics to unpack two questions: why people get deceived, and how scammers engineer that deception, step by step.
Before fraud becomes a financial loss, it is often a decision-making problem. That's why I study it through behavioral economics — and why the same framework matters for understanding AI's influence on human behavior more broadly.
Next in this series: breaking down the specific psychological mechanisms scammers rely on, stage by stage.