No source supports a ranking of "AI cryptocurrencies," because the label has no technical or legal definition, so this page covers a documented case instead: a 2019 research benchmark that tested machine learning on Bitcoin transactions to flag money laundering. The study labeled 203,769 transactions, and the best model in it was a Random Forest, an older method, which beat a graph neural network. The result shows what AI can add around crypto, and how far a model trained on past data is from catching new schemes.
Key takeaways
- The Elliptic data set contains 203,769 Bitcoin transactions and 234,355 payment flows, and only 2 percent of transactions carry an illicit label.
- A Random Forest scored an illicit-class F1 of 0.788, ahead of a graph convolutional network at 0.628 and logistic regression at 0.481.
- Every model performed poorly after a dark market shut down at time step 43, even with retraining after each step.
- The authors describe an industry-standard false positive rate of upwards of 90 percent in anti-money-laundering screening.
- This page is general information, not investment advice, and it does not rate, name, or recommend any token or project.
What the data set contains
The paper "Anti-Money Laundering in Bitcoin: Experimenting with Graph Convolutional Networks for Financial Forensics" was submitted to arXiv on July 31, 2019 and presented in a workshop at the 25th SIGKDD conference. Its authors, led by Mark Weber and Giacomo Domeniconi, worked with a data set shared by the firm Elliptic. Each node is a transaction, and each edge is a flow of bitcoin from one transaction to the next. A transaction is labeled licit or illicit by the category of the entity that initiated it. Licit categories include exchanges, wallet providers, and miners. Illicit categories include scams, malware, ransomware, and Ponzi schemes.
Of the 203,769 transactions, 4,545 (2 percent) are labeled illicit and 42,019 (21 percent) are labeled licit. The remainder has no label. Each transaction carries 166 features: 94 describe the transaction itself, such as fees and the number of inputs, and 72 summarize its immediate neighbors. The data covers 49 time steps spaced about two weeks apart.
How the models were tested
The authors trained on the first 34 time steps and tested on the last 15, a split that follows time because real screening faces transactions that have not happened yet. They compared logistic regression, a multilayer perceptron, a Random Forest, and graph convolutional networks (GCN), which let each transaction take information from its neighbors. The focus was the illicit class, since a model that labels everything licit would still look accurate on a data set that is 98 percent licit or unlabeled.
The authors also state the practical problem: screening systems in industry produce many false alarms, and a better model should cut those without missing more criminals.
Results for the illicit class
| Method | Precision | Recall | F1 |
|---|---|---|---|
| Logistic regression, all features | 0.404 | 0.593 | 0.481 |
| Multilayer perceptron, all features | 0.694 | 0.617 | 0.653 |
| GCN | 0.812 | 0.512 | 0.628 |
| Skip-GCN | 0.812 | 0.623 | 0.705 |
| Random Forest, all features | 0.956 | 0.670 | 0.788 |
The Random Forest led on F1, and its recall of 0.670 means about a third of illicit test transactions went unflagged. The authors note that the supplied features are already informative, which helps simpler methods. They also report that adding neighbor information improved accuracy for all three feature-based models, and that EvolveGCN, a version that tracks change over time, reached an illicit F1 of 0.720.
The dark market shutdown
At time step 43, a dark market closed suddenly. The paper reports that all methods performed poorly afterward, and that even a Random Forest retrained after every test step, with the true labels available, could not reliably catch the new illicit transactions. The authors call robustness to such events a major open challenge.
Anyone reading a vendor's fraud-detection claim can take a lesson from this. A model's score reflects patterns in earlier labeled cases. If criminals change tactics or a service vanishes, those patterns shift, and the figures from a past test stop describing how the model performs today.
What changed since 2019
The same research group has since released a larger data set. A paper first posted to arXiv on April 29, 2024 and presented at the KDD 2024 machine learning in finance workshop introduces Elliptic2. It holds 122,000 labeled subgraphs of Bitcoin clusters inside a background graph of 49 million node clusters and 196 million edge transactions. Its authors, who include Mark Weber from the 2019 study, argue that money laundering is a subgraph problem, because the pattern sits in a group of linked transactions and a single transaction says little. The 2019 figures above therefore describe an early, transaction-level test. As of October 2026, no later result from that line of work was checked for this page, and the 2019 table should not be read as current tool performance.
Questions to ask about any claim of AI crypto monitoring
- What data was the model tested on, and were the labels checked by people?
- Was it tested on later time periods than it trained on?
- Are precision and recall both reported, not only an overall accuracy?
- What happens to its performance when behavior changes?
- Who reviews each flag before action is taken against an account?
FAQ
Is there a list of the best AI crypto projects?
No reliable list exists. The category is a marketing label, the projects differ in what they do, and any ranking would depend on price movement that no source can forecast.
Does the 2019 result apply to today's tools?
Not directly. It is a dated measurement on one data set of Bitcoin transactions, and a 2024 successor data set (Elliptic2) was built around subgraphs instead of single transactions. The 2019 paper still raises useful questions about labels, time splits, and false alarms.
Why did a simpler model beat a neural network?
The authors suggest the engineered features already carried much of the useful signal. They propose combining tree-based methods with graph networks as future work and did not test it.
Can a flagged transaction be treated as proof of wrongdoing?
No. A model output is a lead. The labels in the data set rest on heuristic reasoning about entities, and the paper itself describes the labeling as informed by heuristics.
Related reading
- What a blockchain can and cannot add to an AI system
- AI in Canadian financial services
- Ethical dilemmas that AI raises
- Who owns an AI system
More in AI in finance.
Last verified: October 2026
Sources
- The Shape of Money Laundering: Subgraph Representation Learning on the Blockchain with the Elliptic2 Dataset, Bellei C et al., arXiv:2404.19109, KDD 2024 MLF Workshop
- Anti-Money Laundering in Bitcoin: Experimenting with Graph Convolutional Networks for Financial Forensics, Weber M et al., arXiv:1908.02591, KDD 2019 Workshop on Anomaly Detection in Finance