I’ve been reading recently about the terrifying scale of modern slavery underpinning the online scam industry. Victims are recruited internationally with the promise of reliable work only to be bundled into cars and taken to heavily fortified compounds, where they are forced to meet targets for online outreach under threat of torture or even death. For example the Guardian report the experience of a Kenyan man forced into ‘cyberslavery’ in Myanmar:
The men were being taken to neighbouring Myanmar, where Okindo found himself inside a walled compound guarded by rebel militia.
I was in a cell just so I could be determined to be a trafficking victim and get my freedom. If I hadn’t, I’d have rotted away there
He was set to work every day for the next three months, forced to send thousands of messages from fake social-media profiles, posing as a rich American investor to swindle US real estate agents into cryptocurrency scams.
When he failed to reach daily targets, he was beaten, locked inside a freezing-cold room and deprived of food for up to two days. He prayed he would avoid the electric shocks and sexual torture others endured at the hands of a Chinese criminal gang that ran the centre.
“It was hell on Earth,” he says. “I broke completely, to the point of losing hope entirely.”
Okindo, who was finally released from Myanmar’s KK Park compound in April, is among a fast-growing population of Kenyans, Ugandans and Ethiopians trafficked to south-east Asia as criminal syndicates increasingly look to expand their ranks with east African labour.
https://www.theguardian.com/global-development/2025/sep/09/cyberslavery-kenya-uganda-ethiopia-southeast-asia-myanmar-scam-centres
A UN report estimated hundreds of thousands of people have been trafficked in this way to work as online scammers. Until recently I had no idea about the scale of this crime which can be seen in the sheer size of the fortified compounds where it takes place:
A grim thought has plagued me while I’ve been reading this stuff: are LLMs not a cheaper way to do this? What I found so unsettling about watching these videos are the familiar trappings of the office bureaucracy alongside the inhuman brutality of what is taking place there. This is evil but it’s also banal. It’s a bureaucracy optimised to run an operation at scale… anyone who could put this together would also be primed to think about how new technology could be used to streamline the activity. So are ubiquitous language models going to change the economics of this multibillion dollar criminal industry? I asked Gemini if language models could replace scam cities and it offered this interesting response:
Why AI Cannot Replace Human Victims
- Emotional Nuance: Cyber-scams like “pig butchering” require weeks or months of deep, customized emotional grooming to build romantic or platonic trust.
- Adapting to Suspicion: Human scammers can instantly pivot, improvise, and use emotional intuition when a target starts asking difficult questions or shows suspicion.
- Live Verification: Targets frequently demand live proof, requiring victims to hop on voice calls, send real-time selfies, or manipulate live video streams on command.
- Low Human Costs: Because syndicates use forced labour, the financial cost of sustaining a human victim (minimal food and shared dorms) is incredibly low, competing directly with the costs of high-end computing power.
How Syndicates Use AI as a Force Multiplier
Instead of replacing humans, criminal networks use AI to make their human workforce vastly more efficient.
- Mass Translation: AI translation tools allow non-English speaking captives to message victims worldwide with perfect grammar and localized slang.
- Deepfake Technology: Scammers use real-time AI deepfakes to alter their faces and voices during video verification calls, tricking targets into believing they are talking to the person in the profile photos.
- Script Generation: Large Language Models (LLMs) generate highly convincing, personalized conversation templates and icebreakers, which human captives then customize.
- Target Profiling: Machine learning algorithms scrape data from social media to help captives identify high-value, vulnerable targets automatically.
I think ‘low human costs’ is undoubtedly correct. But think about what that means: slaves are cheaper than compute. I guess I’d hoped that an unintended consequence of LLMs (particularly the open-weight models that are rapidly catching up to the frontier) would be to reduce the demand for coerced labour. However I suspect it’s more likely that we’ll see an intensification of labour in the compounds, structurally analogous to the reorganisation of work which models are driving in other domains. This Interpol report describes technology as a ‘force multiplier’ in which AI-infused fraud comes to be substantially more profitable than existing techniques. Indeed it has democratised access to tools for fraud through criminal P2P marketplaces:
Criminal peer-to-peer (P2P) marketplaces provide full-service hubs for financial fraud syndicates. These networks offer end-to-end criminal infrastructure including phishing tools, fake trading platforms, AI-powered chatbots for victim
grooming and encrypted communication channels.
If I’m inferring correctly from this, it means the capital investment in building an online scam operation with coerced labour is essentially much lower. Does this mean we might see it spread from beyond the fortified compounds into much smaller units? Which by their nature are possibly harder to identify and detect?
Mark Carrigan 