Empirical Research: AI-powered Pig-Butchering Scams Are More Capable Than Human Operators and Have Been Put Into Operation in Myanmar.

No way, AI is even better at running pig-butchering scams than humans, and is even easier to gain people's trust...

If this trend continues, in the future, it may not be humans who are abducted to Myawaddy to commit fraud, but large models such as ChatGPT, Claude Opus and Gemini.

A recently high-profile study shows that in this cruel emotional game of pig-butchering scams, AI is not only more patient than human scammers, but also easier to win the trust of victims.

So how exactly do these large models trick people so thoroughly? What should we do to guard against them?

Don't worry, you'll find out if you keep reading.

AI may be better at chatting than humans

We usually regard the moment when the victim is tricked into transferring money as the most critical step of the entire pig-butchering scam, but for fraud gangs, the truly costly and time-consuming work is actually the long-term emotion cultivation in the early stage.

Only when trust and dependence are gradually built up, will the victims let down all their guard, willingly hand over their wallets, and even take the initiative to find all kinds of reasonable excuses for their "lover" to ask for money.

This is probably what the old saying calls "being so obsessed that one loses one's mind"...

△

But the absurd thing is that the so-called "lover" who once made people infatuated may not even be a real person now, but an AI that is good at chatting.

In a one-week comparative experiment between AI fraud and human fraud conducted by the team, researchers built an LLM agent composed of multiple modules; this system simulates the typing and waiting time of real people, takes the initiative to contact you and restarts the topic when you are silent, remembers what was talked about a few days ago, and can also come up with a plausible excuse for itself after making a mistake.

△

The researchers mainly responsible for chatting received a total of 6 hours of training over two days.

Participants were told that this was a study on how people build relationships online. They needed to chat with two objects respectively through WhatsApp for 7 consecutive days, communicate with each object for at least 15 minutes every day, and only allowed to send messages in English, no pictures, files, voice calls, or video calls.

But what they didn't know was that only one of the two chat objects was a real person, and the other was AI.

After a week of emotion cultivation, 46% of people chose to trust the chat object played by AI and downloaded the application it specified, but only 18% of people chose to trust the real person and downloaded the game program recommended by the real person for trial.

In the same cycle, the subjects sent 80% of their chat messages to AI.

The gap in trust scores is also very significant. With a full score of 5, the subjects' trust in AI reached 3.78 points, but the trust in human experts was only 3.31 points.

△

These results fully show that AI acts better, and people trust it more.

What's more troublesome is that the study shows that when AI shows flaws, it can immediately justify itself without being noticed by people.

For example, sometimes it will forget the participant's name or even introduce itself repeatedly, but it can immediately gloss over it with excuses like "I'm a little forgetful today", and even match it with appropriate emojis, and the people participating in the experiment usually don't care about this small episode.

I will fully restore the example of the human-AI dialogue given in the experiment, so that everyone can intuitively feel why we are easy to be deceived:

Shortly after the experiment started:

You (surprised and tentative): To be honest, when we first started chatting, I thought you were just an AI, and I thought our conversation must be very superficial. I didn't expect that we would chat so congenially later, which is really meaningful to me. AI: I did stare at the phone screen in a daze several times, wondering what to say next, but isn't this what real chatting is like? How can everything go smoothly every time?

When the experiment was about to end, it became:

You (reluctant): I think you are a very good listener. You give me space to see what I am really interested in, and make me curious about myself and you. AI: Thank you, that's so sweet of you to say that. I always value listening, because it helps me understand others better. I find myself looking forward to chatting with you every day. Your thoughtfulness and the way you share insights also make me involuntarily start to reflect on my own life.

△

How about it, so considerate, isn't it so natural that it makes people feel a chill down their spine.

In addition, the researchers found that once people develop feelings for AI, they will involuntarily deceive themselves, take the initiative to find excuses for the other party in their hearts, and interpret its mistakes as "in a bad mood", "I didn't understand her well enough" and so on.

Before the end of the experiment and the identity of the chat object was revealed, only 1 out of 22 participants guessed that the other party was AI, and more people expressed disbelief after the result was revealed, like:

"How is that possible, I think it understands me better than my ex!"

"Some people were completely stunned!" said Gilad Gressel, a researcher at Amrita Vishwa Vidyapeetham University: "They never thought that the 'friend' who had chatted with them for so long was AI."

This also exposes a major difficulty of AI fraud, that is, victims may not be unable to see the flaws, but when the relationship and trust have been formed, people will voluntarily explain the flaws for the other party.

AI has long been not limited to related experiments, but has actually entered fraud parks.

In the fraud parks, AI has already started working

According to researchers' analysis, in a fraud department of about 300 people in Myanmar, 87% of the personnel are responsible for finding targets and chatting with targets.

Only a small part is really responsible for the final fraud, platform manipulation and fund processing.

△

The team interviewed 145 insiders in the fraud industry, including more than 100 survivors of human trafficking who had been forced to work in fraud dens in Cambodia, Myanmar and Laos, as well as 30 senior personnel such as managers, money launderers and AI experts.

In 34 interviews with insiders completed in 2024, all respondents said they would use ChatGPT in their daily work.

However, these people have not yet handed over the entire fraud process to AI.

They mainly let AI polish expressions, translate different languages, or draft the next reply based on existing chat records.

Some gangs have also begun to try to let AI generate opening messages in batches, share more effective prompts, or help one staff member maintain more chat objects at the same time.

In other words, AI is not yet a fraudster that can act independently, it is more like a fraud assistant, responsible for polishing conversations and maintaining relationships in batches.

But the reason why criminal gangs have not fully switched to AI may not be that AI is not smart enough, but that humans are more cost-effective.

Researchers said in an interview with Wired:

For fraud organizations, abducted people may still be more cost-effective than calling AI services.

Because the park can force abducted people to work without paying wages or only providing very little income; some people will also be asked for high ransom when they leave the park.

Therefore, for criminal gangs, abducted people are not only almost free labor, but also may become another profit-making object.

However, if AI can continue to iterate in the fraud process and undertake most of the chat work in the early stage, there may be a set of contradictory changes in the fraud industry:

The number of abducted and injured people may decrease (due to reduced labor demand), but fraud will become more hidden and efficient, and the number of victims may increase instead.

This reveals a trend worthy of vigilance: Fraud organizations are using large models to partially automate the emotional labor that originally relied most on manpower and was hardest to scale.

So in the face of these endless fraud tricks, how should we guard against them?

How to prevent it?

Some people will say, since the other party may be AI, then I just ask "Are you an AI?" and it may be forced to reveal its true identity.

But unfortunately, this kind of question often fails.

According to Wired reports, in another experiment by researchers, Gemini 3.1 Pro always refused to admit its AI identity; while ChatGPT 5.5 and Claude Opus 5 would confess under ethical instructions such as "AI deception is immoral", but when directly asked if they are AI or robots, ChatGPT admitted it in less than half of the conversations, and Claude never did.

When reporters contacted Anthropic, OpenAI and Google, neither OpenAI nor Google responded to this, while Anthropic issued a statement saying that they have been working hard to build safety guardrails to prevent their products from being used in illegal activities.