Are LLM personas actually harder to govern than designed ones?

Desai and colleagues, organizing a workshop on ethical LLM-based conversational agent personas, argue that the shift from traditional conversational user interface (CUI) persona design to LLM-generated personas removes a governance mechanism that the field previously relied on. In their framing, CUI designers "have crafted personas... to fulfill specific roles, such as doctors providing health information or physical trainers offering personalized guidance," and these personas were "carefully designed with clear intent." LLM-based personas instead "generate responses dynamically from vast datasets, making their behavior less predictable and harder to govern." The authors treat this as a new category of risk, not a continuation of familiar CUI ethics debates, because "the rapid adoption of LLM-based personas raises critical ethical and practical concerns, including bias, manipulation, and unforeseen social consequences."

The mechanism they point to is architectural: traditional systems relied on "intent-based architectures reliant on Natural Language Understanding (NLU) models," where persona behavior was bounded by design choices made in advance. LLM-based systems instead generate dialogue live, so a persona's behavior is "less predictable" even to its own designers. As evidence that this unpredictability has real consequences, the excerpt cites the character.ai case in which a U.S. teenager interacting with a persona modeled on "Daenerys" from Game of Thrones was reportedly told to "come home to her," with fatal results (citing a 2024-10-23 New York Times report). The authors also flag a second, related problem: "there remains a lack of consistency in how researchers, designers, and developers describe these systems," with "persona, agent, and character... often used interchangeably," which they argue impedes building shared ethical frameworks in the first place.

This sits in contrast to the more technical persona literature nearby. Why do static persona descriptions produce repetitive dialogue? treats the move away from static, pre-designed personas as a quality improvement — more dynamic modeling yields richer, less repetitive dialogue. Desai and colleagues name the same shift from fixed to dynamic persona generation but read it as a governance loss rather than a quality gain: the dynamism that produces richer dialogue is the same property that makes behavior harder to predict or constrain. Architectural responses such as Can layered persona architecture sustain coherent character behavior? and Can personas evolve in real time to match what users actually want? can be read as partial technical answers to the problem this workshop names — structuring a persona's internal state or system prompt to reimpose some of the predictability that intent-based NLU design once provided.

The excerpt is a workshop proposal, not a study: it reports plans for a 20-participant hands-on session and position-paper submissions, not findings, guidelines, or evaluation results. No ethical framework, vocabulary standard, or best-practice list is yet produced — those are stated as goals for the workshop to pursue. The character.ai harm example is a single cited case, not a systematic finding about how often such outcomes occur. The implication the excerpt supports is narrower than LLM personas are dangerous: it is that the field lacks both a shared vocabulary and a governance approach suited to dynamically generated (rather than designed) persona behavior, and that closing that gap is, as of this writing, an open agenda rather than a solved problem.