cs.CRAug 25, 2026

Anatomy of a Scam Call: What 10,000 real scam and spam calls reveal about how phone scammers operate

Authors: Ethan Traister, Ankit Raj, Jiaqi Gan, Xingyu Shen, Tyler Wu, Yuchen Zhou, Tommy Duong, Kidus Zewde, 

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Telephone fraud is pervasive and costly, but its inner workings are rarely observed at scale. We analyze a complete corpus of 10,211 inbound scam and spam calls -- 913 hours of audio and 330,956 transcribed turns from 5,780 distinct numbers -- collected over 54 days by an AI voice-agent honeypot that answered callers and kept them talking, and introduced in a companion data descriptor. We separate outright scams, which solicit sensitive information, from the larger stream of predatory but legal lead generation ("spam") that feeds them. Scam operations keep office hours (6.6x more calls per weekday than weekend day); thousands of disposable numbers run a small catalog of recycled scripts (thirty opening clusters, half the traffic in the top five); and callers solicit identity anchors -- a home address and a date of birth -- far more often than payment credentials, pressing through persistence and manufactured authority rather than overt threats. Our central experiment asks: does it matter who picks up? Every seeded lead carried one of ten fictitious identities drawn uniformly at random, so the identity a fraud operation reaches is fixed before the caller exists. Across 1,823 randomized calls, scammers spent about 15% more conversational turns per decade of the target's apparent age (rate ratio 1.15, 95% CI 1.08-1.23; randomization p = 0.005) -- yet what they asked for did not change (26.3% of calls reached a request for sensitive information; odds ratio 0.99 per decade, 95% CI 0.90-1.08). A second experiment casts early detection as a benchmark: from a scammer's opening lines alone, on a caller-disjoint split, escalation is predictable at 0.72 ROC-AUC from the first line and 0.87 by the eighth, and a plain bag-of-words classifier matches a fine-tuned on-device language model. Telephone fraud emerges as a templated industry that varies how hard it works a target, but not what it wants.

Figures & tables

Figure 1: The study at a glance. A single closed corpus of answered calls, labeled once by an automated pipeline, supports five questions. Three are descriptive; one is a predictive benchmark scored on a caller-disjoint held-out split; one is a randomized experiment whose treatment—the identity a lead carries—was assigned before the caller ever bought the lead, and which is therefore the only question we answer causally.
IdentityAgeSexHome cityForms seededCallsNumbers
Emma22FMinneapolis, MN4,663162104
Tyler23MDenver, CO4,61011189
Chloe24FAustin, TX4,558190155
Madison26FCharlotte, NC4,694212148
Layla29FHouston, TX4,745161119
Amir33MDearborn, MI4,622151109
Table 1: The ten randomly assigned identities. “Forms seeded” counts lead-generation submissions carrying that identity between 1 and 20 July 2026; “Calls” counts the substantive inbound calls it received, and “Numbers” the distinct originating numbers behind them. All ten lines share one area code, one speech-to-text engine, one language model, one text-to-speech vendor, and an identical behavioral protocol; only the identity differs.
Figure 2: Scam calls follow business hours. Left: calls by hour of day (converted to U.S. Eastern time), with the 9-to-5 business window marked for both the East Coast (09–17 ET) and the West Coast (12–20 ET). Right: average calls per active day, by day of week—weekdays (navy) dwarf weekends (grey), a 6.6 ×\times difference.
Figure 3: A few scripts, many numbers. The twelve largest opening clusters (labeled by their distinctive terms), showing the number of calls (navy) and the number of distinct originating phone numbers (orange) running each. Each script is placed from dozens to hundreds of different numbers. Two clusters carry no single vertical and are marked “mixed.” The callback-style openers are not evidence of a prior relationship: no real person completed these forms, and the callers bought the lead downstream (Section 3.2).
Figure 4: Scams reveal themselves early—but model capacity does not help. From only the first k scammer lines, each model predicts whether the call will later reach a request for sensitive information, on a single caller-disjoint held-out fold of 1,276 calls (no originating number shared between training and test). Left: ROC-AUC; right: average precision (AP), with the 17.5% base rate marked. TF-IDF with logistic regression (navy) matches or beats both fine-tuned small language models at every prefix length and leads clearly by the eighth line. All five rungs are scored on identical data, so they are directly comparable.
Figure 5: What they want and how they get it. Left: how often each sensitive item is requested, colored by identity (blue) versus payment (orange); identity anchors dominate. Right: pressure tactics, ranked by how many calls use each; impersonation and persistence dominate, while overt threats are rare.
Figure 6: Who picks up changes how hard scammers work, but not what they want. A: the seeding pipeline attached each of ten identities to a statistically indistinguishable number of lead-generation forms, so assignment behaves like a coin flip. B: mean scammer turns per call against the stated age of the identity that answered; whiskers are 95% bootstrap intervals clustered on the originating number, the curve is the fitted negative-binomial trend, and p comes from enumerating all 10! reassignments of the age labels. C: the share of calls reaching a request for sensitive information, which does not vary with age; the labelled outlier is the one identity whose scripted biography is unusually compliant. D: the difference is concentrated in the long tail—older identities produce far more very long calls, while the median call is unchanged.
Persona-level outcomeρ\rho vs. agep
Scammer turns per call+0.8300.005
Scammer words per call+0.8060.007
Share of calls past 10 minutes+0.7210.023
Call duration (mean)+0.5760.088
Talking time (median)+0.5880.081
Share of scam-labelled calls-0.6850.035
Table 2: Every persona-level outcome we examined. Spearman rank correlation between an identity’s stated age and its mean outcome across the ten identities, with p-values from enumerating all 10! reassignments of the age labels. The last three rows are balance and placebo checks rather than outcomes. Because the analysis was specified after collection ended, we list all of them; under a Holm correction across the five engagement outcomes, scammer turns and scammer words remain significant and the duration-based measures do not. Two further families were tested the same way and are summarized in the text rather than tabulated: the nine individual request types, none of which varies significantly with age, and the eight call-ending categories, of which one—the caller appearing to believe the scam had succeeded—is nominally significant ( ρ\rho = +0.71, p = 0.027).
More in this topic

Financial Fraud Detection

Financial fraud detection identifies suspicious transactions, accounts, claims, and behavior to prevent losses and guide investigation.

See all 139 papers

Explore similar work

CardsList
Aug 25, 2026

A Corpus of Real Scam- and Spam-Call Conversations from an Active Voice-Agent Honeypot

Real conversations between fraudsters and their targets are among the most informative artifacts for studying telephone scams, yet also the scarcest: passive honeypots overwhelmingly capture automated messages and hang-ups, large-scale studies characterize call metadata rather than dialogue, and manual scam-baiting does not scale. We present a dataset of real scam-call conversations collected by an active voice-agent honeypot. Dedicated numbers are seeded into the lead-generation channels fraud operations harvest; inbound callers are answered by a low-latency conversational agent that adopts a plausible target persona and sustains the interaction while every call is recorded, transcribed, and automatically labeled. Over an initial 53-day window we captured 10,015 inbound scam and spam calls (6,601 with two or more turns): roughly 895 hours of audio and 328,869 transcribed turns from 5,665 distinct originating numbers. Under a holistic classifier the substantive calls are predominantly predatory-but-legal lead generation ("spam", about three in five), while about one in seven is an outright "scam" (949 in this snapshot). Each call carries a turn-level transcript, three-channel audio, per-turn latency telemetry, and layers of automatic labels, including a holistic scam/spam/legitimate judgment corroborated by independent human review (75% agreement on the binary decision). We describe the collection system, the record structure, and technical validation of the corpus's realism and label quality, including that the agent is recognized as non-human in only about 5% of engaged calls. We also benchmark established scam-detection methods, where detectors trained on published synthetic dialogue collapse in precision on real traffic.

Ethan Traister, Dennis Tsang Ng, Siyu Zhang +7

Financial Fraud Detection
AskPDF
Sep 10, 2026

The Machines Are Calling: Measuring Automated and Synthetic Voices in Unwanted Inbound Calls

In February 2024 the U.S. Federal Communications Commission (FCC) placed AI-generated voices under the Telephone Consumer Protection Act (TCPA). Yet no peer-reviewed measurement says how much unwanted call traffic is placed by a machine, or how much of that machine speech is synthesized rather than played from a recording. We report both with a disclosed pipeline. An interactive voice honeypot (language-model personas on real U.S. numbers, the caller recorded on its own track) recorded 10,987 calls over 66 days. Three instruments read each opening: an audio fingerprint that finds the same recording played on other calls, a commercial synthetic-speech detector on the caller's first ten seconds, and blinded listeners who check what it flags. Of the 7,233 greeted calls we analyze, 13.8% open with a recording we also heard on another call, and 13.1% with fresh audio the detector labels synthetic. A further 9.9% open with a caller who never spoke after our greeting, 54.2% with fresh audio the detector labels human, and 9.0% could not be scored. Machine-voiced openings are therefore at least 26.9%, a further tenth of calls are silent connections we read as machine-placed, and replays of a recording make up 45% of the detector's own rate (29.3% of 6,192 scored openings). The same waveform played on two calls lands on opposite sides of the detector's threshold 13.6% of the time, and eleven listeners confirm 54.4% of what it flags. Synthetic openings concentrate in lead-generation spam (33.8%), not fraud (21.1%); 0.44% disclose automation. Prevalence tracks how long a bait number has circulated (59% against 19% in the same weeks): seeding history, not calendar time, explains the trend. Campaigns outlast their numbers: one recorded compliance notice opens calls in six campaigns, and one synthetic voice serves nine.

Xingyu Shen, Tommy Duong, Muduo Xu +8

Scam AI (Reality Inc.)

Topic pending
AskPDF
Sep 16, 2026

TeleAntiFraud 2.0: A Refreshable, Profile-Grounded, and Audio-Based Benchmark for Telecom Fraud Detection

Telecom fraud scripts evolve rapidly and are often designed to resemble routine service conversations, creating two key requirements for audio-based telecom-fraud evaluation. First, benchmarks must incorporate newly observed scam patterns without overwriting previously established test sets. Second, they must distinguish fraud from lawful, near-domain calls rather than relying on topic-separated negative examples. We present TeleAntiFraud 2.0, constructed with our Mixed-Tree Anti-Fraud Generation Pipeline and evaluated under a monthly frozen evaluation protocol. The pipeline transforms online fraud-case abstracts into profile-grounded scenarios, expands them through mixed-tree generation, realizes fraud and non-fraud dialogue paths under shared contexts, renders validated dialogues as role-matched speech, and freezes the resulting audio, labels, prompts, manifests, and provenance records for each monthly evaluation set. Each frozen set contains 900 Chinese calls, comprising 600 fraud and 300 near-domain non-fraud cases. Controlled text experiments show that three classifiers achieve perfect macro-averaged F1 (Macro-F1) when evaluated against unrelated or ordinary negatives, but drop to 0.65-0.68 with near-domain sibling negatives. Full-set audio and automatic-speech-recognition plus large-language-model (ASR+LLM) evaluations further reveal class-prior shortcuts, prediction collapse, and snapshot sensitivity. Together, these findings establish near-domain construction and collapse-aware reporting as core requirements for evaluating audio-based telecom-fraud models under realistic confusable conditions. The accompanying research artifact includes the construction code, evaluation scripts, manifests, and documentation. Our dataset and code are available at https://anonymous.4open.science/r/TeleAntiFraud-2_0-EEB2/.

Huiyuan Liu, Zhiming Ma, Yanxing Liu +11

People’s Public Security University of China · JD Technology · Chongqing Ant Consumer Finance Co., Ltd., Ant Group +3

Anomaly Detection
AskPDF