Can AI Fight Human Trafficking? What Works, What Doesn’t, and What Remains Unproven

Cover image: Can AI Fight Human Trafficking? What Works, What Doesn't, and What Remains Unproven

Oguzhan O. Demir, PhD
oguzhan.demir@gc4ss.org

Senior Expert
Global Center for Security Studies

Introduction

Emerging technologies, particularly artificial intelligence have become a vital part of our lives. We chat, learn, practice, compare, analyze, and explore numerous things with AI tools. From data analysis to graphic design and from video generating to complex translation tasks, AI has been a useful friend for making things easier, faster, and more reliable.

AI has also been a quite useful actor in countering transnational threats and crimes, including human trafficking. Non-governmental organizations, researchers, private companies, and law enforcement authorities have been working on AI tools that can detect human trafficking activities and perpetrators and save victims for a long time.

In this working paper, the use of emerging AI based technologies in counter human trafficking efforts by governmental and non-governmental organizations as well as researchers are discussed. Three anti-trafficking tools which have been utilizing AI technologies are selected as cases. After describing the AI capabilities of each tool, the analysis critically examines their effectiveness and practical limitations.

How is AI Used in Countering Human Trafficking?

“Data analysis” and “predictive modeling” are two of the most popular types of AI capabilities. To identify anomalous behaviors including frequent travel, repeated trips to risky locations, and suspected money transfers, machine learning models analyze datasets such as financial transactions, travel itineraries, immigration records, and maritime or flight logs. Moreover, some AI tools can combine geographic mapping and machine learning red-flag activities to predict transit routes of human trafficking, and to identify and intercept victims. In addition, some AI models are designed to analyze behavioral and identity indicators of cargo vehicles/vessels such as dark activity records to score global cargo and maritime vessels by risk tier. It helps border/customs agencies to carry out targeted inspections.1

Other popular types of AI techniques are “Natural Language Processing (NLP)” and “covert ad decoding.”2 To identify and find probable hidden keywords, emojis, and linguistic patterns that can point to a form of exploitation, NLP algorithms routinely search internet sex ads, classifieds, and dark web forums. Also, machine translation tools help decode cross-border trafficking networks and identify connections between their regional operations.

AI is also developed in “computer vision” and “image recognition”.3 In online victim identification, to find missing or trafficked children, computer vision has been utilized as a method to examine digital evidence in the context of gathering and screening online platforms. Further, satellite imagery and automated identification system data can be tracked by AI tools to monitor fleets/vessels that are associated with forced labor or illegal transit. Interestingly, X-ray images of vehicles and containers are scanned by machine learning and computer vision models to look for hidden illegal imports or altered structural compartments.

Examples of AI-Based Tools to Counter Human Trafficking

Spotlight

Spotlight is an AI-driven web-based domestic sex trafficking detection tool and investigative platform launched in 2014.4 Developed by non- profit organization Thorn5, and now operated by Canary, it has been used by over 8,000 law enforcement agencies across the US and Canada. It utilizes NLP and entity resolution to ingest and cross-reference numerous online commercial sex ads every day. It is claimed that every day 300,000 sex classified ads are posted in the US.6 By using the coded terminology and specific phrases, the system detects linguistic markers that indicate a child victim. Spotlight integrates existing law enforcement databases and tools which allows investigators to link disparate ads through phone numbers and locations patterns and then prioritizes high-risk cases. The investigators then can focus on them and rescue victims.7

Spotlight partners with several high-tech brands AWS, OKTA, Single Store, Microsoft, Bugsnag, Google, and QuestionPro. According to AWS8 which provides open search service capabilities for Spotlight, data from 150,000 ads per day are processed and analyzed based on risk profiles provided the by the law enforcement. As a result of this initiative, Spotlight claims that over 28,500 victims and over 28,900 traffickers have been identified.

Spotlight uses a combination of computer vision, natural language processing (NLP), and graph analytics to analyze millions of online commercial sex advertisements and flag potential child trafficking victims. Because online ads are often written in coded language with altered photos to evade detection, Spotlight’s AI is specifically trained to see past intentional complication.9

Spotlight’s AI Capability

Spotlight has the capability of “computer vision.” As traffickers frequently crop, mirror, blur, or add stickers/emojis over faces and tattoos to prevent reverse-image searches, Spotlight uses specialized “visual AI” to bypass these tricks. First, it uses “facial recognition and comparison”. The system scans people’s faces in online ads and compares them against missing-child databases10 as well as across millions of historical and current ads to track a victim’s movement across different cities. Second, it utilizes “perceptual hashing” or “image fingerprinting”. Even if an image is resized, filtered, or cropped, perceptual hashing identifies underlying visual similarities to link duplicate or slightly altered photos back to the same source. Spotlight also “matches background with objects”. When faces are hidden, computer vision models analyze secondary clues such as hotel room bedspreads, wallpaper, tattoos, or distinctive jewelry to link different ads to the exact same physical location or person.

Spotlight also utilizes “Natural Language Processing” in analyzing texts. Escort ads rarely state a victim’s age explicitly. They rather use slang, emojis, and coded phrasing to signal youth or vulnerability. Using Semantic & Keyword Classification, NLP models trained on thousands of verified trafficking cases scan ad text for linguistic patterns, coded slang, and emojis associated with minors or coercion. Spotlight also uses automated risk scoring. Based on the text, posting patterns, and metadata, the AI assigns a triage score to all new ads. High-scoring ads are pushed to the top of an investigator’s dashboard, so detectives review the most urgent leads first. The AI automatically extracts unstructured details like phone numbers, payment app handles, and city names from messy blocks of text and turns them into structured, searchable fields.

Finally, machine learning models, behind the scenes, cluster ads that share phone numbers, email addresses, IP metadata, or posting schedules, automatically building a visual map of a potential trafficking ring. Finally, Investigators can set automated watchlists for specific victims or selectors while the AI continuously monitors incoming web scrapes and sends real-time alerts if a known phone number or face suddenly appears in a new city.

Infographic: Spotlight’s main AI-based tools — facial recognition, perceptual hashing, background matching, NLP, risk scoring, machine learning clustering and automated watchlists
Spotlight’s main AI-based tools.

Spotlight: What Works Well

Spotlight ingests and analyzes tens of thousands of daily online escort advertisements using natural language processing, perceptual matching, and facial recognition. Law enforcement users consistently report a 50% to 65% reduction in the time required to sort through leads and identify victims.

Traffickers frequently move victims across city and state lines. Spotlight links disparate data points such as phone numbers, photos, locations, and ad text into a single view. It allows over 8,000 investigators across the U.S. and Canada to collaborate and track movement patterns in real time.

By cross-referencing online commercial sex ads against databases of missing children in partnership with the National Center for Missing & Exploited Children (NCMEC)11, the tool flags potential minor victims quickly. Investigators can also set automated alerts to notify them the moment a new ad featuring a known victim is posted.

The software is provided free of charge to vetted law enforcement officers with statutory authority, removing budget barriers for smaller local departments.

Spotlight: What Doesn’t Work Well

Spotlight relies heavily on scraping publicly accessible classified ad sites and open web forums. As trafficking activity increasingly shifts toward encrypted messaging apps, private social media groups, and peer-to-peer payment platforms, open-web scraping loses visibility.

No single algorithm catches every indicator. Law enforcement analysts have noted that running the exact same search selectors across multiple tools often uncovers illicit ads that Spotlight missed on its own.

Civil liberties advocates and researchers raise concerns over warrantless mass scraping of online data and the use of facial recognition technology, which can misidentify individuals or sweep consensual adult sex workers into law enforcement databases without due process.

Spotlight is strictly a lead-prioritization tool rather than an automated solution. It still depends entirely on local police capacity, trauma-informed victim services, and traditional detective work to verify leads, execute safe recoveries, and prosecute offenders.

AI flag is not probable cause for an arrest, and software cannot conduct a welfare check. Human officers must still verify the identity of the person in the ad, obtain subpoenas or warrants, locate the victim safely, and coordinate with social services.

Whooster NightWatch

Whooster NightWatch12 is a free law enforcement investigative tool developed through a partnership between the data-fusion company Whooster and the counter-trafficking nonprofit DeliverFund.13 NightWatch focuses on targeting traffickers and criminal networks by connecting phone numbers, emails, and identities to illicit advertising data. NightWatch operates inside Whooster’s OWL Intelligence Platform and merges two primary data streams.

The system connects two massive sources of information to uncover hidden leads. The first is DeliverFund’s “RedLight” Dataset. This analyst-curated database holds billions of data points and millions of current and past commercial sex ads pulled from the open web, adult platforms, and the dark web. The second is Whooster’s Data Fusion. This stream pulls together public records, live telecom data, social media profiles, criminal histories, and asset records.

Because these two sources run together in the background, investigators don’t even have to be looking for human trafficking to find it. When an officer runs a routine check on a phone number or email, even while working an unrelated case like auto theft or narcotics, NightWatch automatically flags any ties to illicit sex ads and generates a real-time trafficking risk score.

NightWatch’s AI Capability

NightWatch relies on two powerful tools to connect the dots. The first is dynamic trafficking risk scoring. When an investigator searches a phone number or email, the system’s algorithms look at a mix of clues like how often ads are posted, where they appear, which sites are used, and telecom metadata. They are used to generate a live score showing how likely that contact info is tied to trafficking.

The second tool is entity resolution. Traffickers constantly switch between burner phones, fake emails, and VOIP numbers to cover their tracks. To beat this, machine learning scans billions of public records, telecom logs, and DeliverFund’s “RedLight” ads for overlapping patterns, helping investigators realize that two completely different phone numbers belong to the same person or criminal ring.

Infographic: NightWatch’s main AI-based tools — dynamic trafficking risk scoring, entity resolution, NLP and automated data extraction
NightWatch’s main AI-based tools.

NightWatch uses natural language processing (NLP) and automated scrapers to gather messy, unstructured text from the open web, adult platforms, and the dark web. It pulls out key details like phone numbers, payment handles, and travel routes, and organizes them into a clean, searchable format.

What sets this apart from standard automated scrapers is the human touch. DeliverFund pairs its AI tools with former intelligence analysts who manually double-check, clean up, and organize high-priority trafficking clues before they ever enter the database.

NightWatch: What Works Well

NightWatch proves especially valuable in a few key areas of modern policing:

Uncovering hidden rings during routine work: Because the tool is built right into standard investigative workflows. It even works via SMS in the field when officers don’t have cell service. Patrol officers and detectives looking into drugs, weapons, or gangs frequently come across overlapping trafficking networks by accident.

Tracing traffickers and telecom footprints: While victim-focused tools analyze photos and ad text to locate missing youth, NightWatch excels at following the digital trail behind the scenes. It traces burner phones, email addresses, and financial footprints back to the real identities of both traffickers and buyers.

Catching what other tools miss: Because DeliverFund builds and curates its own independent dataset, analysts who cross-check phone numbers across multiple platforms regularly spot illicit ads in NightWatch that other systems completely overlook.

NightWatch: What Doesn’t Work Well

Even with its investigative strengths, NightWatch comes with a few important limitations and trade-offs. DeliverFund specifically designs its intelligence to take down criminal networks rather than focus only on victim recovery. Because of this, the platform lacks dedicated tools for cross-referencing missing-child databases or using facial recognition to prioritize underage victims.

NightWatch works best when an investigator already has a specific lead, like a phone number from a tip or a sting operation, to plug into the system. It isn’t built for broadly browsing newly posted local ads by sight.

Combining commercial data-broker records with scraped adult ads raises serious civil liberties questions. Critics point to concerns over how long every day, non-criminal data is stored, as well as the risk of falsely tying innocent people to criminal rings through recycled phone numbers.

TraffickCam

TraffickCam14 matches background details in online exploitation photos against a database of millions of hotel room images. These images are crowdsourced by travelers and scraped from public sites. Its neural networks identify specific room features like carpet patterns, lamp styles, and headboard shapes allow investigators to pinpoint the exact hotel where a victim is being held. It is used daily by the National Center for Missing and Exploited Children (NCMEC).

TraffickCam uses deep learning and computer vision to perform fine-grained visual search that is essentially acting like a specialized reverse-image search engine. Developed by researchers at Saint Louis University, George Washington University, and Temple University15, the system matches background details from investigative photos against millions of indexed hotel images.16

TraffickCam: Core AI Techniques

Deep neural networks are trained by TraffickCam using large datasets such as “Hotels-50K.”17 The AI transforms an image’s distinctive visual elements, like carpet patterns, headboard textures, wall art, and lighting fixtures, into numerical vectors (embeddings) rather than just identifying generic items. The AI locates the closest mathematical matches in the database when an investigator uploads a query photo.

To preserve the victim’s privacy and stop the AI from focusing on human traits, analysts mask out the victim before sharing an investigation photo. The method combines AI in-painting and unique training approaches to ignore masked parts and infer the surrounding room information because these blacked-out areas might mislead ordinary image-matching models.

While separate rooms within the same hotel may have entirely distinct designs, standard hotel chains frequently utilize the same furnishings in hundreds of locations. To address this, the platform employs object ensembles models that have been trained to identify and contrast certain, distinguishing objects, such as framed artwork, drapes, or tile patterns, that reduce a match to a precise feature.

Infographic: TraffickCam’s main AI capabilities — visual feature matching, masked image handling, object ensembles, explainable results and robust matching across conditions
TraffickCam’s main AI capabilities.

TraffickCam employs visualization tools to show precisely which elements of the two photographs (such as a particular lampshade or carpet weave) triggered the match because law enforcement cannot rely on a “black box” result while gathering evidence. This makes it possible for human analysts to promptly confirm the accuracy of a match.

While investigation images are frequently dim, grainy, clipped, or taken at strange angles, promotional photos on travel websites are properly lit, arranged, and captured with wide-angle lenses. The AI learns to identify the same room under radically varied lighting and camera circumstances by mixing commercial travel photographs with daily, unprocessed smartphone photos that passengers submit through the TraffickCam App.

TraffickCam: What Works Well

The tool is quite helpful in identifying hotel chains. Based on established color schemes, bedding designs, and furniture templates, the AI is quite good at identifying the brand or chain (such as Hilton, Holiday Inn, or Sheraton) even when an image is cropped or indistinct.

The neural networks perform extremely well when trained on artificially blacked-out images. For privacy reasons, analysts can fully remove a victim from a picture, and the AI can still reliably extract background context from the frame’s remaining edges.

By blending sleek promotional shots with messy, dim smartphone photos from the TraffickCam app, the AI successfully bridges a major visual divide. It has learned to match a grainy, awkwardly angled selfie with a bright, wide-angle photo from a travel site.

Instead of leaving investigators to manually scroll through thousands of local listings on Expedia or Google, TraffickCam speeds up the search. It ranks the closest visual matches in seconds, giving analysts clear heatmaps to quickly double-check and verify the results.

TraffickCam: What Doesn’t Work Well

Even with these breakthroughs, TraffickCam still runs into some real-world roadblocks. Everyday travelers who use the app usually stay at major chains, business hotels, or popular tourist spots. Meanwhile, trafficking often happens in independent, roadside, or budget motels that have little to no online footprint. They rarely get crowdsourced photos. Simply put, if a motel isn’t in the system, the AI can’t match it.

Spotting a specific hotel chain is relatively easy, but pinpointing the exact location is much tougher. A budget room in New York can have the exact same carpet, curtains, headboard, and mass-produced art as one in Texas. Without a unique architectural quirk or a glimpse outside a window, the AI can only name the brand, not the street address.

Also, evidence photos are often blurry, dark screenshots of screenshots, or tightly cropped around a person. When all you can see is a blank white wall or a plain bedsheet, the AI has no distinct textures or fixtures to latch onto.

Hotels remodel every few years. If the database only has older photos of a room, and the property later swaps out the carpets and wallpaper, the AI won’t be able to connect the dots until new guests upload updated pictures.

Finally, as the system is built specifically around hotels and motels, it hits a wall if traffickers move victims into private homes, unindexed short-term rentals like Airbnbs or vehicles.

Comparison of AI Capabilities of Counter-Trafficking Tools

FeatureTraffickCamSpotlightNightWatch
DevelopersSaint Louis, George Washington, & Temple UniversitiesOriginally Thorn, now operated by CanaryWhooster & DeliverFund
Core AI DomainComputer Vision & Fine-Grained Visual SearchPredictive Modeling, NLP, & Perceptual HashingData Fusion, Entity Resolution, & Behavioral Risk Scoring
Primary Data SourceCrowdsourced hotel photos + public travel site imagesOpen-web classified ads, online forums, & missing child recordsIllicit web/dark web ads, telecom records, OSINT, & criminal/public databases
Main Investigative GoalWhere was this photo taken? (Physical geolocation)Who is this victim and are they a minor? (Victim identification)Who is running the network? (Trafficker & buyer attribution)
Accessibility & Access ModelHybrid: Public mobile app for uploading room photos; search database is restricted to vetted law enforcement and anti-trafficking organizations.Strictly Restricted: Available exclusively to verified law enforcement agencies and approved partner NGOs; zero public access.Strictly Restricted: Available exclusively to vetted law enforcement, intelligence agencies, and authorized prosecutors/investigators.

Conclusion

As emerging technologies help us in our daily lives, they have also potential to help law enforcement communities in different formats. Three AI based detection and investigation tools that have been examined above could help save and protect victims to a satisfactory extent. They have merit in the fight against human trafficking. What is special in these tools that they have been developed by non-profits/researchers, supported by tech firms in social responsibility manner, and got accepted widely by governmental bodies which integrated their systems to these tools.

There are several technical shortcomings of these AI tools. Despite them, ethical concerns are also raised as major criticism issues towards these tools as it is the case in almost all AI applications. Critiques argue that warrantless mass scraping of online data can misidentify consensual sex workers into databases of law enforcement. Also, warrantless use of facial recognition technology, which is a common practice around the world, results in a large-scale violation of personal privacy laws.

Furthermore, although these tools help law enforcement spare time in investigation by making the publicly available data into actionable, analytic, and operational source of information, it does not mean that these tools are proves of successful crime control.18 A system may identify a victim or suspect, but then we need to know if the identification resulted in a successful prosecution or prevented a criminal network. Even a successful prosecution and disruption does not necessarily show us that the technology reduced the future criminal activity. That is out of the scope of this paper and requires attention by researchers.

Notes

  1. Nikkel, 2025, AI and Human Trafficking, https://theexodusroad.com/ai-and-human-trafficking/ ↩
  2. Li, R., Tobey, M., Mayorga, M. E., Caltagirone, S., & Özaltın, O. Y. (2023). Detecting human trafficking: Automated classification of online customer reviews of massage businesses. Manufacturing & Service Operations Management, 25(3), 1051-1065. ↩
  3. Ijiga, A. C., Olola, T. M., Enyejo, L. A., Akpa, F. A., Olatunde, T. I., & Olajide, F. I. (2024). Advanced surveillance and detection systems using deep learning to combat human trafficking. Magna Scientia Advanced Research and Reviews, 11(01), 267-286. ↩
  4. https://www.spotlight.ngo/about-us ↩
  5. AWS, (2016). Thorn: Digital Defenders of Children at AWS, https://youtu.be/UUiDpgAzsqU ↩
  6. https://www.spotlight.ngo/about-us ↩
  7. Runyon, N. (2025). AI puts Spotlight on victim identification in fight against domestic minor sex trafficking, https://www.thomsonreuters.com/en/institute/articles/spotlight-trafficking-victim-identification ↩
  8. AWS, (2021). Using Technology to Combat Human Trafficking https://aws.amazon.com/blogs/publicsector/using-technology-to-combat-human-trafficking/ ↩
  9. https://www.spotlight.ngo/about-us ↩
  10. https://www.missingkids.org/home ↩
  11. https://www.missingkids.org/home ↩
  12. https://whooster.com/nightwatch/ ↩
  13. https://deliverfund.org ↩
  14. https://www.traffickcam.com ↩
  15. Dr. Abby Stylianou at Saint Louis University, Dr. Robert Pless at George Washington University, and Dr. Richard Souvenir at Temple University. ↩
  16. A. Stylianou, J. Schreier, R. Souvenir, and R. Pless. (2017). “TraffickCam: Crowdsourced and Computer Vision Based Approaches to Fighting Sex Trafficking,” IEEE Applied Imagery Pattern Recognition Workshop (AIPR), Washington, DC, USA, 2017, pp. 1-8, doi: 10.1109/AIPR.2017.8457947. ↩
  17. https://github.com/GWUvision/Hotels-50K ↩
  18. Giommoni, L. (2024). Why We Cannot Identify Human Trafficking from Online Advertisements. Journal of Human Trafficking, 1–15. https://doi.org/10.1080/23322705.2024.2435198 ↩

How to cite