A modular review of the evidence for harmful manipulation by large language models

Harmful manipulation by large language models (LLMs) is now a regulatory concern, yet the empirical evidence remains difficult to assess. Harmful manipulation involves multiple features: an influence goal, tactics that bypass reasons or exploit vulnerabilities, limited target awareness, successful influence, and harmful outcomes, which existing studies rarely capture together. This review organizes the literature by progressively stronger constellations of these features. Some studies show that LLMs successfully influence beliefs through overt argumentation, without clear harm. Other studies show that LLMs can influence covertly, but without of harm or bypassing reason. Closest current approximations to harmful manipulation come from studies attempting to operationalize harmful outcomes. Some show little behavioral success, Others capture successful harmful influence, but without covert or reason-bypassing moves. Overall, the literature does not yet provide full evidence of harmful manipulation, but establishes several constituent capacities that make harmful manipulation by LLMs an empirically grounded risk rather than a purely speculative one.