Conversational Scam Detection

Overview

Conversational scams often reveal risk across identity claims, trust-building, responses, and later requests rather than in one message. This series moves from process models and benchmark construction to prediction, memory, evaluation, and human safeguards, showing why stateful detection can improve early warning while remaining evidence-limited, privacy-sensitive, and prone to false alarms.

Learning outcomes

When scam risk depends on earlier claims, promises, responses, and later pressure, trajectory-aware state adds evidence that message-only checks miss.
PreScam makes progression measurable from structured victim reports, but its filtered, reconstructed cases are not a prevalence sample or full transcripts.
Early warning, future-action forecasting, and technique recognition are distinct capabilities and should be evaluated with different metrics and timing costs.
Summary memory can preserve distant evidence at bounded cost, yet compression, false alarms, privacy exposure, and incomplete evidence limit deployment claims.

Keywords

conversational scamsprocess-aware detectiontrajectorymemoryearly warningverificationfalse alarmsbenchmarkpsychological techniquesprivacy

Chapters

Conversational Scam Trajectories

Why context changes a request's meaning

8 min · Why context changes a request's meaning

8 min

Scam Kill Chain Phases

Track objectives, actions, and uncertainty

8 min · Track objectives, actions, and uncertainty

8 min

PreScam Benchmark Construction

What the filtering funnel makes measurable

9 min · What the filtering funnel makes measurable

9 min

Three Layers of Scam Detection

Separate evidence, representation, and inference

9 min · Separate evidence, representation, and inference

9 min

Scam Risk and Action Prediction

Separate early warnings from future moves

13 min · Separate early warnings from future moves

13 min

Scam Detector Memory

Balance context, cost, and evidence

13 min · Balance context, cost, and evidence

13 min

Scam Detector Evaluation Limits

Read recall, false alarms, and timing together

13 min · Read recall, false alarms, and timing together

13 min

Process-Aware Detection Limits

Create distance for verification, not obedience

12 min · Create distance for verification, not obedience

12 min

Preview every chapter · Unlock full text, audio, and video from chapter 4

Plans

Ch.1-3 FreeCh.4-8 Premium

FAQ

Why is a conversational scam different from a suspicious message?

A conversational scam unfolds across turns, so identity claims, promises, responses, and later requests can change the meaning of an otherwise ordinary message. Message-level checks still matter for explicit links, contact details, and brand claims.

What should a process-aware scam detector predict?

It should separately estimate whether a conversation is approaching an extraction phase and forecast likely future scammer actions. Those outputs have different failure costs, metrics, and timing requirements.

Does conversation memory make scam detection more reliable?

Conversation memory can preserve distant evidence that a recent-message window loses, but summaries are lossy and reported gains come from specific evaluation settings. Full history can cost more, while short exchanges may favor simpler context.

Can process-aware detection prevent scams on its own?

No. It is decision support that can create time for independent verification, but benchmark results do not prove deployment-wide recall, low false alarms, privacy safety, or faithful explanations.