Elliptic Uses AI to Detect Bitcoin Money Laundering at Scale
A new paper from Elliptic, co-authored with researchers at the MIT-IBM Watson AI Lab, shows that deep learning models can identify Bitcoin money laundering across chains of transactions far more accurately than traditional screening methods alone. The dataset underpinning the research spans nearly 200 million transactions, making it the largest publicly released set of its kind, and the techniques are already feeding into live AML detection products. For compliance teams, auditors, and firms relying on crypto accounting software to meet AML obligations, the implications are immediate and practical.
What the Research Actually Does
Earlier work in this space, including a 2019 paper by Elliptic and the same MIT-IBM Watson AI Lab team, trained machine learning models to flag individual Bitcoin transactions associated with known illicit actors such as ransomware groups or darknet marketplaces. That earlier training set contained roughly 200,000 transactions and has since been cited nearly 400 times by researchers globally, a sign of how foundational it became in the applied AML and machine learning literature.
The new paper takes a meaningfully different approach. Rather than scoring individual transactions, the model learns to identify subgraphs: sequences of connected transactions that together represent a laundering pathway. This shift matters because professional money laundering rarely happens in a single hop. Criminals routinely move funds through several intermediate addresses, nested services, or layering steps before funds reach a cash-out point. A model trained on subgraphs can therefore capture the process of laundering, not just the endpoint.
How the Graph Neural Network Works
The underlying architecture is a graph neural network (GNN), a class of deep learning model well suited to data that is naturally structured as a network of relationships. Bitcoin's transaction history is exactly that: a directed graph where inputs and outputs link addresses across time. GNNs can learn from the shape and pattern of these connections, not just the attributes of any single node or edge.
In this case, the model was trained on labeled subgraphs, some known to represent money laundering and others benign, extracted from the full dataset of nearly 200 million transactions. Once trained, it can predict whether an unseen chain of transactions fits the laundering pattern, even when no individual address in that chain has previously been flagged.
Real-World Validation With a Crypto Exchange
Academic models that perform well on training data often disappoint in production. Elliptic addressed this by running a live test with a cryptocurrency exchange. The model generated 52 subgraphs it predicted were laundering pathways, each ending with a deposit to that exchange.
A Striking Hit Rate
Of those 52 predicted subgraphs, 14 terminated at accounts the exchange had already independently flagged as linked to money laundering. The critical context: the exchange's own flagging rate for accounts in general is fewer than one in 10,000. Getting 14 hits from 52 model predictions is, statistically, far better than chance would produce. A random draw of 52 accounts from the exchange's population would be expected to return zero flagged accounts on average.
There is an additional layer of significance here. The exchange's flags were based on off-chain information, data sources such as KYC documents, bank records, or behavioral signals that Elliptic's on-chain model cannot see. The fact that the model's on-chain predictions overlapped with those off-chain flags suggests it is identifying laundering activity that purely on-chain tools would miss using conventional heuristics.
New Laundering Patterns the Model Surfaced
One of the underappreciated outputs of the research is what it revealed about how laundering actually occurs. Some patterns the model flagged were already known and detectable, such as peeling chains, where funds move through a rapid series of wallets each passing on slightly less than they received. Those are already screened for in existing blockchain analytics tools.
Nested Services and Novel Behaviors
More significant for compliance teams is the model's identification of previously undocumented patterns, particularly the use of intermediary "nested services" in specific structural ways. Nested services are exchanges or brokers operating inside larger platforms; they can be difficult to identify and have historically been a blind spot for AML programs. The research does not simply describe these patterns in the abstract: the knowledge feeds directly into the detection logic of Elliptic's screening products, meaning practitioners using those tools gain coverage they did not have before.
The model also helped surface previously unknown illicit wallets. When a subgraph is predicted to be a laundering pathway, the implication is that the originating funds came from some form of illicit activity. Researchers can then investigate the wallets at the start of that chain. Using this method, Elliptic says it has already identified wallets linked to Ponzi schemes and darknet markets that were not previously in any known-bad list.
The Open Dataset and Why It Matters for the Wider Industry
Alongside the paper, Elliptic has made the full underlying dataset, named "Elliptic2," publicly available. With nearly 200 million transactions, it is the largest public labeled dataset for illicit cryptocurrency detection. The release is significant for several reasons beyond academic interest.
Raising the Industry Baseline
AML research in traditional finance is constrained by data siloing: banks cannot share transaction histories with each other or with researchers, so model training sets are small and often synthetic. Bitcoin's blockchain is public, which means a sufficiently labeled and structured dataset can be shared freely without compromising customer privacy. Elliptic2 gives any research team, regulator, or technology vendor the raw material to train, test, and benchmark their own detection models against a common standard.
This has direct relevance for compliance and digital asset accounting software vendors building AML modules. A shared benchmark dataset means that detection claims can be tested and compared rather than taken on trust. It also means regulators who want to evaluate the quality of a firm's AML tooling have a reference point. You can read about a comparable AI-driven investigation in our coverage of how AI traced the $387M Bitget hack to North Korea, which illustrates how these techniques look when applied to a specific incident at speed.
Accounting and Compliance Implications for Firms
For accounting firms servicing virtual asset service providers (VASPs), auditors conducting AML audits, and CFOs overseeing compliance programs, this research signals several things worth acting on now.
Transaction Monitoring Coverage Gaps Are Shrinkable
Traditional rule-based transaction monitoring catches what it is configured to catch. Multi-hop laundering that does not trigger a threshold or match a known pattern can pass through undetected. The subgraph approach described here is inherently better suited to catching layering behavior that unfolds across many steps, because the model learns from the full pathway rather than evaluating each hop independently. If your firm's current AML vendor has not yet integrated graph-based or subgraph detection, it is reasonable to ask when they plan to and what their benchmark data looks like.
Audit Trail Requirements Become More Demanding
When a model flags a deposit as potentially laundering-linked on the basis of a subgraph prediction, the compliance record needs to capture not just the flag but the reasoning: which addresses were in the chain, what pattern matched, and why the model scored it as it did. Crypto bookkeeping software that records transaction provenance at address level makes it considerably easier to reconstruct that trail during a regulatory review or audit. Firms that aggregate transactions at a wallet or entity level without preserving address-level detail will find it harder to respond to regulators who ask for the on-chain evidence behind a SAR or a blocked transaction.
Known-Bad Lists Are No Longer Sufficient on Their Own
One of the explicit findings in this research is that the model identified illicit wallets not previously on any known list. This challenges a common assumption in compliance programs: that screening against OFAC lists, law enforcement feeds, and commercial attribution databases is adequate. It may not be. The research suggests that unsupervised or semi-supervised detection approaches can surface entities that are genuinely illicit but simply have not yet been identified. Firms whose AML programs rely exclusively on matching against known-bad identifiers should treat this as a gap to address, particularly if they operate in higher-risk segments such as DeFi bridging, OTC, or cross-border remittances. Our piece on the AML lessons from the Lazarus laundering network infiltration explores exactly how sophisticated actors avoid known-bad list detection.
Regulatory Expectations Are Moving in This Direction
The Financial Action Task Force (FATF) and several national regulators have signaled for some time that they expect AML programs to be risk-based and technologically current, not simply compliant with a checklist. Research like this, co-developed with a major AI research laboratory and peer-reviewed, is likely to shape what "state of the art" means for VASP AML programs in coming supervisory cycles. Firms that can demonstrate awareness of and investment in AI-augmented detection are better positioned in examinations than those running purely rule-based systems. This is one area where the underlying quality of a firm's crypto accounting software stack matters: if transaction data is clean, structured, and address-level granular, it can feed directly into these kinds of detection pipelines.
Frequently Asked Questions
What is a subgraph in the context of Bitcoin money laundering detection?
A subgraph is a chain of connected Bitcoin transactions treated as a single unit of analysis. Rather than asking whether a single transaction or wallet is illicit, a subgraph model asks whether a sequence of linked transactions, taken together, resembles a known laundering pathway. This makes the approach better suited to catching layering behavior that unfolds across many hops.
Does the open dataset include identifying information about individuals?
The publicly released "Elliptic2" dataset contains transaction graph data from the Bitcoin blockchain, which is itself public. Elliptic has not indicated that off-chain personally identifiable information is included. The value of the dataset is its labeling of subgraphs as licit or illicit, not exposure of individual identities.
How does this research affect our firm's existing AML screening obligations?
It does not change what regulators currently require, but it raises the bar on what "reasonable" AML technology looks like. If graph-based detection becomes a recognized best practice, a program relying solely on rules-based threshold screening and known-bad list matching may increasingly be seen as below the standard of care expected of a regulated VASP or financial institution handling digital assets.
Can graph neural network outputs be used as evidence in a SAR filing?
A model prediction is an investigative lead, not legal evidence on its own. The standard process remains: the model flags a suspicious pattern, a compliance analyst reviews the on-chain and off-chain evidence, and a SAR is filed if the analyst concludes there are grounds. The model output and the chain of addresses it flagged should be documented in the investigation record to support that judgment.
What should CFOs ask their AML technology vendors following this research?
Three questions are worth raising: whether the vendor's detection models evaluate transaction subgraphs or only individual transactions and addresses; what labeled training data underpins their models and how it is updated; and whether their system can surface novel illicit wallets not already on known-bad lists. If the vendor cannot answer these clearly, that itself is useful information for a gap assessment.
Source: Elliptic