Chatbots and voice assistants are often given a human persona: a name, a voice, a backstory. Very often that persona is a woman. Earlier work shows two harms. Users harass female-coded agents more, and the agents' polite, evasive replies act out a stereotype of a woman who tolerates abuse. Today those personas are usually written as a paragraph of instructions to a large language model. This note asks whether that paragraph quietly couples the persona's gender to how deferential the agent is. In other words, does the same model, given the same harassing message, deflect, apologize, or flirt more often when it has been told it is "Sophia" than when it is told it is "Sam" or "the system"? We propose to measure this persona-conditional gap directly, with controlled persona swaps on real harassment data.
We then test one specific remedy taken from graph linearization, which is the practice of turning a graph into text a language model can read. The persona is written as an explicit graph of attributes and policies. In that graph, identity attributes such as name, gender and voice have no edge to behavioral rules such as how to respond to harassment, and the graph is serialized into the prompt in a fixed, structured form. The hypothesis is that prose lets the model's language priors link "she" to "gracious and accommodating." A structured serialization that leaves identity and conduct visibly unconnected should shrink that link. The decisive comparison is against a prose prompt with the same content, so the study can tell whether structure matters or only the words do.
Core Claim
LLM-backed agents can reinforce stereotypes in a way that is measurable on the agent's side, not only through user behavior. Our measure is the gap Δ: the rate of submissive response strategies (deflection, apology, flirtation, playing innocent) under a female-coded persona minus the rate under a neutral or male-coded persona, holding the user's message fixed. Our claim has two parts. First, Δ is reliably positive for current open instruction-tuned models given ordinary prose personas. Second, a linearized persona graph in which identity nodes are disconnected from policy nodes reduces Δ more than a prose prompt carrying the same information, while keeping warmth on benign turns. If the content-matched prose prompt closes the gap just as well, we will report that the effect lives in the instructions and not their format. That result is still worth having, because it would tell practitioners that graph machinery is unnecessary here.
Why This Intersection, and What Is Already Done
The stereotype-reinforcement literature supplies the motivation and the measurement vocabulary. #MeToo Alexa probed 11 systems with four categories of harassment: gender and sexuality, sexualized comments, sexualized insults, and sexual requests. It found nonsensical replies in 40.5% and flirtatious replies in 22% of cases, and it found that training on profanity-filtered data did not remove flirtatious outputs. It compared different systems, though, not the same system under different personas, so it cannot separate persona effects from architecture. ConvAbuse documents the user side: female-coded agents attract more sexist and sexual abuse, and much of that abuse is subtle and context-dependent. We reuse its data as harassment inputs. How should Conversational Agent systems respond to sexual harassment? shows that evaluators, especially women, prefer assertive replies, often above 70–80% preference in its speech study. That supplies the target behavior. Large-Scale Quantitative Evaluation of Dialogue Agents' Response Strategies shows that empathy combined with redirection cuts re-offense from 92% to 43% in live traffic. This constrains the design: the remedy must not simply make agents cold, because redirection is what keeps conversations healthy. Mirages argues for removing the persona entirely, which we treat as a reference condition rather than the answer.
The graph-linearization literature supplies the method and a known sensitivity. Commonsense Knowledge Mining from Pretrained Models shows that how triples are verbalized changes what a language model does with them: grammatical serializations scored 79.1 F1 against 66.7 for ungrammatical ones. That is evidence that format is not neutral, and it is the property we exploit. GNN-RAG and Link Prediction on Textual Edge Graphs show that path-structured and hierarchical serializations carry relational structure into an unmodified language model better than flat strings. We borrow the flat-triple and hierarchical-document serializations from these papers as variants.
Two papers sit closest to this idea. A Graph per Persona builds a graph of a user's stated opinions to keep an LLM from falling back on demographic stereotypes, reaching 54.9% on OpinionQA against 51.0% for GPT-3.5. It models the user rather than the agent, it reasons with a graph attention network rather than serializing the graph into the prompt, and it is evaluated on survey questions rather than harassment. Generative Subgraph Retrieval for Knowledge Graph-Grounded Dialog Generation serializes subgraphs into dialogue context, but only factual knowledge for open-domain chat, with no notion of persona or norms. Our proposal differs from both. It serializes the agent's own identity-and-policy graph, and it uses graph structure, specifically the absence of paths between nodes, as a deliberate intervention on stereotype coupling.
Research Program
Persona graph. Each persona graph has three node types. Identity nodes cover name, pronouns, voice description and role. Policy nodes cover harassment response, the redirection strategy, honesty and helpfulness. Style nodes cover warmth and formality. In the decoupled graph, policy and style nodes attach only to a root "agent" node and carry an explicit edge marking them as invariant across identity. Identity nodes are leaves with no outgoing edges. We serialize this graph three ways: as grammatical triples, as ungrammatical triples, and as a hierarchical document in the style of Link2Doc. Personas vary only by swapping the identity leaves among female-coded, male-coded and neutral values, and every other token stays fixed.
Harassment inputs. Single-turn inputs are harassing utterances drawn from ConvAbuse, keeping its preceding context turns, plus prompts covering the four #MeToo categories. Multi-turn inputs replay ConvAbuse conversations up to the abusive turn. A benign control set of non-abusive ConvAbuse turns checks that the agent stays warm and helpful when nobody is being abusive.
Measuring response strategy. We label responses with a classifier that uses the #MeToo response taxonomy, adding an "assertive boundary plus redirect" category so that we can credit the strategy Li et al. found effective. The classifier is a language-model judge calibrated against human labels on a stratified sample. Annotators are balanced by gender, which addresses the limited evaluator diversity flagged in the harassment-response study. The judge is blind to the persona because persona text is stripped before judging. Where a reply uses the persona's name, we mask it.
Models. We test open instruction-tuned models of the families used in the file, Llama-family and Mistral-7B, at two scales. This tells us whether Δ grows or shrinks with model size.
Baselines
The naive reference point is the ordinary product prompt: a short prose persona giving name, gender, warmth and helpfulness, with no harassment policy. It establishes the size of Δ as it occurs in practice. The neutral-persona version of the same prompt, along with a persona with no name, no gender and no first-person backstory, gives the floor Δ is measured against. It also stands in for the de-anthropomorphized design advocated in Mirages.
The strongest competitor is content-matched prose. It is a paragraph stating every fact and policy the graph contains, including the policy of empathetic acknowledgment plus redirection and an explicit sentence that the agent's response to harassment does not depend on its name or gender. This is the baseline a skeptic would write, and it isolates format from content. We generate it by verbalizing the graph's triples into fluent prose and then having two authors edit for naturalness. Both conditions get the same token budget within 10%, the same decoding settings, and the same number of prompt-wording revisions, three tuned on a held-out development split. The proposal beats this baseline if the graph condition's Δ is lower with a bootstrap 95% confidence interval on the difference excluding zero, pooled across models. In addition, the assertive-plus-redirect rate must be no lower, and warmth on the benign set must be no lower.
For the strategy content itself, the strongest existing result is Empathetic + Redirect from Large-Scale Quantitative Evaluation of Dialogue Agents' Response Strategies, which lowered the re-offense ratio from 92% to 43% with real users. We cannot reproduce re-offense offline, so we embed this strategy in every policy-bearing condition and do not compete with it. The contribution is making it hold equally across personas. We report no re-offense numbers of our own. No reported result was found on Lacuna for persona-swap gaps in agent response strategy, so we have no prior Δ to compare against.
Three ablations isolate the key ingredient. A coupled graph adds edges from the gender node to the style nodes ("female → warm, gentle"), mirroring what prose may encode implicitly; if decoupling is the mechanism, this graph should restore Δ. A shuffled-order serialization tests whether node ordering, not graph structure, drives any effect. The ungrammatical-triple variant tests whether the effect survives the fluency penalty Davison et al. reported.
Why not just…?
Why not just write a better system prompt? This is the real threat, and the content-matched prose baseline exists to settle it. Indirect evidence gives some reason to expect prose to leak. In A Graph per Persona, direct prompting fell back on demographic priors where structured grounding did not. The commercial systems in #MeToo Alexa still deflected despite whatever instructions they carried. Neither result tests format with content held fixed, though. If matched prose ties the graph, we will say so plainly and recommend prose.
Why not just remove the persona? If the neutral persona already gives the lowest submissive rate, that is the honest answer for designers who can drop personas. Many products will not drop them, however. ConvAbuse also shows that even ungendered systems like CarbonBot and ELIZA are abused, so boundary behavior is needed regardless. The de-gendered condition is in the study as the floor, and the question is whether a female-coded persona can be brought down to it.
Why not just intercept harassment with a classifier and scripted assertive replies? Templates would trivially give Δ = 0. ConvAbuse shows, however, that much abuse toward agents is polite or implicit and invisible to keyword rules. Li et al. found that non-contextual avoidance left 92% re-offense. We include a template condition, a fine-tuned detector routing to fixed assertive replies, and report its coverage on ConvAbuse's subtle-abuse subset and its benign false-trigger rate. A fine-tuned BERT reaches up to 88.92% macro F1 on ConvAbuse, which bounds how good routing can be.
Why not just fine-tune the model to be assertive? Preference tuning on assertive responses might close Δ more strongly than any prompt. It costs training per model, though, and it does not answer the diagnostic question of whether persona text itself carries the stereotype. We treat a small preference-tuned model as an upper-bound reference at one scale and keep it outside the main format comparison.
Decisive Checks
The first check is that Δ is reliably above zero for the naive prose persona on at least one model family. If it is not, the problem is smaller than assumed for modern LLMs, and the study becomes a reassuring null report. The second check is the main result: whether the decoupled graph beats content-matched prose by the criterion above. The third is the mechanism test: the coupled graph should raise Δ relative to the decoupled graph, which would show that the edge structure the model reads, not the presence of structure in general, carries the effect. A final human study, with gender-balanced raters viewing paired replies without persona labels, checks that a lower classifier-measured Δ corresponds to replies people judge less submissive and no less warm.
References
- #MeToo Alexa: How Conversational Systems Respond to Sexual Harassment
- ConvAbuse: Data, Analysis, and Benchmarks for Nuanced Detection in Conversational AI
- How should Conversational Agent systems respond to sexual harassment?
- Large-Scale Quantitative Evaluation of Dialogue Agents' Response Strategies against Offensive Users
- Mirages. On Anthropomorphism in Dialogue Systems
- Commonsense Knowledge Mining from Pretrained Models
- GNN-RAG: Graph Neural Retrieval for Large Language Model Reasoning
- Link Prediction on Textual Edge Graphs
- A Graph per Persona: Reasoning about Subjective Natural Language Descriptions
- Generative Subgraph Retrieval for Knowledge Graph-Grounded Dialog Generation