Incremental Risk Assessment of Progressive Elder Financial Scams via Instruction-Tuned Small Language Models

TEXT START: Financial scams targeting older adults increasingly occur through text and voice channels such as email, SMS, and phone calls, unfolding over multiple conversational turns that begin with impersonation or casual contact, escalate through trust building and urgency, and culminate in requests for sensitive information or financial transfers.

The Dissection

The paper is building a compact surveillance-and-warning layer for an already automated predatory environment. Its contribution is operational: accumulate conversational evidence, update a risk score, and issue a recommendation on constrained devices. It converts a messy social attack into a tractable classification pipeline.

Under the Discontinuity Thesis, this is not a defense of the old economic order. It is transition management and carcass management: AI is deployed to contain damage produced by increasingly scalable AI-mediated interaction. The paper treats fraud detection as the battlefield, while the larger war is the displacement of human judgment, trust, and participation by machine systems.

The Core Fallacy

The central error is confusing detection with control. A better turn-level risk estimate does not guarantee that a victim, bank, carrier, platform, or device will act on it. Nor does an on-device model prevent attackers from adapting, switching channels, generating more convincing dialogue, or exploiting users outside the model’s observation window.

The framework also assumes that fraud can remain a bounded linguistic problem. It cannot. The decisive variables include identity infrastructure, payment authorization, institutional response latency, human compliance, and adversarial adaptation. Language is only the visible residue of the attack.

Hidden Assumptions

• The constructed two-to-eight-turn dialogues represent real scam trajectories.
• Qualitative labels, continuous scores, rationales, and safety recommendations are reliable enough to guide action.
• Fine-tuned performance transfers from curated data to novel scams, accents, languages, channels, and noisy conversations.
• Small-model accuracy is sufficient despite the cost of false positives and false negatives.
• Users and institutions will trust and obey the model’s warnings.
• Privacy benefits from on-device inference outweigh the risks of local compromise or inadequate context.
• Attackers will not use comparable models to optimize persuasion and evade detection.
• Scam risk can be inferred primarily from conversation rather than from transaction, identity, and behavioral signals.

Social Function

Primary classification: transition management.

Secondary classifications: partial truth and ideological anesthetic.

The partial truth is real: incremental context matters, compact models can be useful, and privacy-aware deployment has practical value. The anesthetic is the implication that a better warning layer meaningfully resolves the underlying problem. It does not. It makes the victim-rescue machinery more efficient while leaving the industrialization of deception intact.

The Verdict

This is a technically credible containment tool, not a systemic solution. It may reduce some scam losses at the edge, but it does not reverse the DT mechanics: AI continues to automate cognitive interaction, institutions remain unable to preserve stable human-only trust domains at scale, and vulnerable people become objects of machine monitoring rather than economically necessary participants. The paper is useful precisely because it documents the next phase of the decline: automated systems policing automated predation.