Darknet markets trade without any law or authority protecting buyers, yet fraud within them is rare. How does an unregulated market hold deception down so far? Andrei and Veltri answer with signalling theory, which holds that a seller can convey trustworthiness to buyers in two ways. The first is costly signals, which cannot be faked because they require real investment: here, a reputation accumulated from past customers' reviews, seniority in the market, and use of escrow, where the market holds the buyer's money and releases it to the seller only once the buyer confirms delivery. The second is cheap signals, which anyone can produce instantly at almost no cost: here, the product description, its length, the variety of its vocabulary, and its positive tone. Every seller sends both kinds at once, so the authors set the two against each other for the same sellers, analysing a scraped dataset from AlphaBay, one of the largest darknet markets ever, which ran from December 2014 to July 2017, using two Generalised Additive Models. Sellers carrying the costly signals were reported as fraudulent less often, while those who invested most in their product descriptions were reported as fraudulent more often.
Research Objectives
- To identify which signals are associated with lower or higher fraud, a gap in a literature focused on reputation and sales.
- To test whether costly signals, escrow and positive reputation, are associated with less fraud.
- To test whether cheap description signals, length, linguistic diversity and sentiment, are associated with more fraud.
- To test whether seller lifespan signals credibility.
Methodology
- Secondary analysis of a scraped AlphaBay dataset from the Darknet Market Archives: 114,385 items, 6,033 sellers, 1,270,000 reviews, collected 26 to 28 January 2017. Zero-sale listings were excluded; the descriptive table reports 1,147,768 reviews analysed.
- Fraud was measured by dictionary: a review counted as a scam if it contained 'scammer' or a related term such as 'fraudster'.
- Predictors: percentage of positive reviews, escrow (yes or no), days selling, and three text metrics, description length relative to the market average, linguistic diversity, and a scaled AFINN sentiment score.
- Two GAMs, Tweedie family and log link: M1 without text variables, M2 with them. Controls: price, product category and class, shipment type, number of reviews.
Key Findings
- Escrow was associated with an estimated decrease in expected fraud of 20.69 percent in M1 and 21.67 percent in M2; transactions outside escrow were significantly more likely to be scams.
- Percentage of positive reviews had a significant non-linear relationship with fraud, generally inverse. Overall 97.7 percent of reviews were positive and fraud was rare.
- Cheap signals ran the other way: linguistic diversity (1.55) and sentiment score (0.89) were significant and positive.
- Description length was not significant (0.03, p = 0.063), so the authors treat H3 as only partially confirmed.
- Seller lifespan was significant but irregular: positive over the first 100 days, negative from 100 to 600 days, then falling steeply after 600, where confidence intervals widen. Text variables raised deviance explained from 61.9 to 62.3 percent.
- The authors explain the mechanism behind these results: escrow acts as a secure intermediary holding funds until both parties fulfil their obligations, so a seller who defrauds faces a higher risk of losing money. A good reputation brings higher sales and premium prices but is difficult to build and easy to lose, raising the cost of opportunistic behaviour and making reputation a reliable signal that sustains the social order of the market.
Recommendations
- Escrow, which the authors note is not widely adopted in mainstream commerce, could mitigate fraud there and help new sellers lacking reputation.
- Platforms should pair secure payment mechanisms with reputation systems.
- Anti-fraud campaigns should teach buyers to be sceptical of listings using enthusiastic language but no costly guarantee such as payment protection or reputation.
- Future research should use longitudinal designs, since cross-sectional data cannot capture change over time.
Key Takeaways
- Signals costly to produce, escrow, reputation and seniority, are associated with less fraud; signals cheap to fake are associated with more.
- The authors call product description an ambiguous signal: varied vocabulary and positive tone were associated with more fraud, not less.
- For Thailand and ASEAN this is a transferable mechanism, not local evidence: AlphaBay was not an Asian market, but escrow-style payment holds and scrutiny of persuasive listings are options Thai platforms could weigh.
References
Andrei, F., & Veltri, G. A. (2025). Signalling strategies and opportunistic behaviour: Insights from dark-net markets. PLOS ONE, 20(3), Article e0319794. https://doi.org/10.1371/journal.pone.0319794
Full text (Open Access): https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0319794
