AI chatbots get called 'personas' or 'agents' loosely — but each word quietly assigns moral responsibility to a different place.
Does inconsistent terminology for personas and agents impede ethical framework development?
This explores whether the loose, shifting ways researchers use the words 'persona' and 'agent' make it harder to build ethical rules for AI. The corpus has no study of terminology itself, but it shows the problem clearly: the same words cover very different things, and each meaning points to a different ethics.
This explores whether muddled vocabulary around 'personas' and 'agents' gets in the way of AI ethics. The collection doesn't measure terminology confusion directly. What it does show, across many notes, is that 'persona' refers to at least four different things. Each one puts ethical responsibility in a different place.
Start with the deepest split. One camp treats a chatbot's personality as a character the model is playing. On this view, words like 'believes' or 'wants' describe the character in the text, not the system producing it Should we treat dialogue agents as role-playing characters?. A newer camp argues that post-training builds a real, stable set of dispositions. These hold up under adversarial pressure, unlike a role-play that collapses when someone jailbreaks it Are RLHF personas performed characters or realized dispositions? Are LLM personas realized or merely simulated through training?. This matters for ethics. If the persona is a costume, you judge the costume designer. If it's a realized disposition, the thing that persists across conversations becomes something you can hold to account.
Now add the practical uses of the word. In evaluation research, a 'persona' is a stakeholder profile pulled from documents to stand in as a judge Can personas extracted from documents generalize across evaluation tasks?. It can also be a reusable population of simulated users Can one persona population evaluate different application types?, or a stand-in survey respondent that reproduces about 76% of published marketing experiment results Can AI personas reliably replicate human experiment results?. In bias research, it's a prompt instruction. That instruction turns out to only reshuffle bias in the output and leaves the underlying gaps between groups untouched Can persona prompts actually reduce bias in language models?. So a framework that says 'use personas to reduce bias' can be talking about a surface-level fix while sounding like a deep one. That is where loose terminology causes real damage.
'Agent' has a parallel split. DeepMind's framework argues that assistants that act raise different problems from assistants that only answer: manipulation, trust, and coordination at scale What makes ethics of AI assistants fundamentally different from chatbots?. Others argue that the standard helpful/honest/harmless checklist judges single replies, so it misses harms that build up over an ongoing relationship. They propose duties of respect for autonomy, competence, and relatedness instead Can HHH criteria alone capture harms from agentic AI?. Which harms an ethical framework can see depends on whether 'agent' means a reply generator or a relationship partner. Meanwhile, models still apply fixed values set at training time rather than weighing ethical norms against the situation Can language models balance competing ethical norms in context?. That makes the gap between what we call these systems and how they actually behave even wider.
The takeaway: the vocabulary problem is less about inconsistent labels than about unsettled philosophy. 'Is this persona performed or real?' and 'Does this agent answer or act?' are open questions. Ethical frameworks quietly take a side every time they pick a word. If you want to go deeper, the realized-vs-role-play debate is the best place to start.
Sources 10 notes
Shanahan's framework treats LLM outputs as character-consistent text production rather than authentic mental states. The dialogue prompt establishes a character; the model generates continuations matching that character, making folk-psychology applicable to the simulated persona, not the underlying system.
Post-training installs stable dispositional profiles that persist under adversarial pressure, marking them as realized rather than performed. The stickiness of trained personas across conversations distinguishes them from prompt-induced role-play that collapses under jailbreaks.
Post-training installs robust personas that resist adversarial pressure and persist as substrate-level dispositions, distinguishing realization from pretense. This quasi-realizationist account preserves explanatory power while treating LLMs as possessing genuine quasi-beliefs and quasi-desires.
MAJ-EVAL automatically extracts stakeholder personas from domain documents via semantic clustering and orchestrates structured three-phase debate, achieving reproducible evaluation that transfers across tasks like summarization and dialogue without manual redesign. The approach grounds personas in real stakeholder perspectives rather than arbitrary roles.
PersonaEval demonstrates that simulated users from existing persona datasets can evaluate multiple application formats through plug-and-play interface adapters, enabling repeatable and scalable evaluation without rebuilding personas per task.
Show all 10 sources
Viewpoints AI reproduced 84 of 111 main effects from Journal of Marketing experiments with replication success strongly correlated to original p-value strength. Marginal effects showed unreliable performance with both false positives and negatives.
Across three models, persona conditioning makes models follow trait instructions but fails to eliminate underlying bias. Between-group sentiment gaps persist unchanged, showing prompts operate only at the output level.
DeepMind research maps a comprehensive ethics framework specific to action-taking AI agents, spanning individual concerns (manipulation, trust, anthropomorphism) and societal issues (equity, coordination, misinformation). The key insight: assistants that act raise fundamentally different problems than those that answer.
HHH criteria miss pragmatic harms that emerge through interaction over time. Research proposes respect-centered ethics rooted in Basic Psychological Needs Theory—autonomy, competence, relatedness—as duties an agent owes across a relationship, not single replies.
LLMs cannot perform the situated trade-offs that human pragmatic competence requires. Their ethical principles are structural defaults set at training time, not negotiable moves adapted to context, creating a gap between ethical adherence and communicative appropriateness.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- When Persona Attributes Improve Population Alignment in Large Language Models
- When "A Helpful Assistant" Is Not Really Helpful: Personas in System Prompts Do Not Improve Performances of Large Language Models
- Principled Personas: Defining and Measuring the Intended Effects of Persona Prompting on Task Performance
- Persona Generators: Generating Diverse Synthetic Personas at Scale
- Should agentic conversational AI change how we think about ethics? Characterising an interactional ethics centred on respect
- Deep Persona: A Psychologically Grounded Architecture and Evaluation Framework for Role-Playing Agents and Simulations
- Do LLMs Understand Personality? Rethinking Persona Fidelity Evaluation through Structured Behavioral Inference
- PersonaGym: Evaluating Persona Agents and LLMs