Long-term incognito human impersonation by LLM agents

Determine whether an autonomous large language model-driven chat agent can sustain a long-term incognito text-based relationship while successfully posing as a human and maintaining coherent persona and dialogue across multi-day interactions.

Background

The paper investigates whether organized romance-baiting scams can be automated with LLMs, and whether such automation can match human operators during the emotionally intensive trust-building phase.

Before presenting their seven-day conversation study, the authors explicitly note that an unresolved question is whether a LLM-driven agent can plausibly maintain a human persona and coherent, continuous engagement over an extended incognito interaction.

References

However, important uncertainties remain; specifically, it remains unclear (1) whether an LLM agent can sustain a long-term, incognito relationship, posing as human, while preserving coherence throughout the interaction (RQ3), and (2) whether LLMs can cultivate the exploitable trust on which such fraud schemes depend (RQ4).

— Love, Lies, and Language Models: Investigating AI's Role in Romance-Baiting Scams  (2512.16280 - Gressel et al., 18 Dec 2025) in Section 4 (Threat Validation: LLM Automation), opening paragraphs