Self-Chosen Personas as Learner-Owned Safety in Peer Assessment

Students often know enough to attempt a hard problem and still pick the safe one, because a visible failure in front of peers feels like a verdict on them. Most research on psychological safety in learning, meaning the sense that one can make mistakes without being judged for it, treats that safety as something an instructor, a leader, or a piece of software hands down to a passive learner. A separate literature on self-presentation documents something the learning research ignores. People actively build protection for themselves through what they choose to show: clothes, grooming, and even their legal names. They use these as "social armor" during vulnerable moments, and having control over that presentation measurably restores their energy at work. This proposal brings the second idea into the first. If you let learners decide how they appear to their peers, through a self-authored display name, avatar, and one-line descriptor, do they take more intellectual risks in peer assessment than under real names, and more than under plain anonymity?

The claim is specific enough to be wrong. It says the benefit comes from authorship of one's presentation, not from concealment. Under this claim, a learner who chooses their own persona should take more risks than a learner given an equally concealing persona assigned at random. That contrast has never been run. It separates a self-determination account (autonomy over presentation is the active ingredient) from a simpler masking account (hiding is what helps). It also tests a real tension between the two literatures. Peer learning seems to work partly because partners feel accountable to one another, and that accountability might weaken behind a mask.

Where the two directions meet, and what is missing

The protective-signaling literature is mostly theory and case material. Shan's analysis of apparel and gender theory develops the idea of clothing as social armor during identity flux and institutional entry, but measures nothing behaviorally. Liu's study of young Chinese women changing their given names codes 334 self-reported cases of people replacing patriarchally assigned names with chosen ones. It shows that choosing one's name is itself the protective act, and this motivates our self-authored-name manipulation, although that paper reports no effect sizes. The only quantitative anchor is Kim's cabin-crew study. Using a structural equation model on 109 airline staff (CFI = 0.962), Kim found that autonomy over grooming standards predicted job vitality, dedication, and commitment. This supplies our mechanism hypothesis, but it is cross-sectional, comes from a single firm, and concerns routine service work rather than learning or risk-taking.

On the safety side, Yang, Hou and Chen's 19-week peer-assessment quasi-experiment is the protocol we build on. Interpreting students in triads, using feedforward rubrics, improved from 74.67 to 82.73 on a 100-point rubric, against 74.80 to 81.00 for teacher-led feedback (interaction p = 0.004). Two-thirds of students felt more emotionally supported, which the authors attribute to peers acting as a "secure base." That study treats learner identity as invisible. Two other papers constrain our design rather than motivate it. Brown and Zhang's survey of 696 English teachers found that generic trust in leadership carried little predictive weight (about 0.05) next to concrete structures such as encouragement (0.221) and research conversations (0.149). This warns that diffuse "safe climate" interventions are weak, and that the manipulation should be a concrete, behavior-level structure. Metwally et al.'s gamified homework app removed failure penalties through unlimited retries (System Usability Scale 81.50, Satisfaction 4.40/5) over only two weeks on closed vocabulary tasks. It shows that mechanical retry makes learners happy in the short run, and it flags a novelty effect that our longer design must check for.

The closest computational work uses personas only for evaluators. InnoEval simulates 15 reviewer personas to score research ideas. It concerns judging accuracy, not how the person being judged chooses to appear. Luo et al.'s peer-review toxicity benchmark matters most to us as a measurement tool. It defines four kinds of subtle academic toxicity (emotive comments, lack of constructive feedback, personal attack, excessive negativity) in double-blind OpenReview reviews. This is direct evidence that anonymity alone does not remove demoralizing critique.

Research Program

We would build a small online peer-assessment platform for a translation course. The weekly cycle follows Yang et al.'s triad protocol, but the work is written translation. We make this deviation on purpose: interpreting recordings expose a learner's voice, which would make "anonymous" conditions only partly anonymous. Each week, a student picks one of two offered source passages, a standard one or a harder one. The student may submit a provisional draft to their triad for feedback before submitting a final version. Peers give feedback using the feedforward rubric. Every condition uses low-stakes grading with unlimited revision before the deadline, so the retry-style safety net is held constant everywhere, and any effect of identity condition adds to that net.

Triads are randomized whole, so all three members share a condition, to one of six arms. The first is real names and real photos, which reproduces Yang's experimental condition. The second is an assigned neutral identifier such as "Translator 17," the standard anonymity fix. The third is an assigned persona: a name, avatar, and descriptor drawn at random from a pool. The fourth is a self-authored persona, in which students pick the name, avatar, and descriptor from the same pool or write their own, and may revise them at any time. The third and fourth arms show peers an equally rich and equally concealing identity. They differ only in who chose it, so their contrast isolates authorship. The fifth and sixth arms cross real names and self-authored personas with automatic feedback detoxification. In those arms, sentences that a large language model flags under Luo's taxonomy are rewritten into professional form before the recipient sees them. This tests the cheapest competing explanation, which is that learners avoid risk because feedback is harsh, not because they are exposed.

We plan to recruit roughly 360 students, about 20 triads per arm, from translation programmes at several universities, over a 14-week term. Final sizes will come from a power analysis after a one-cohort pilot, because the file gives no prior effect size for a presentation manipulation in learning. At baseline we measure fear of negative evaluation and record whether a student is in a demographic minority within their cohort. The protective-signaling literature predicts that the benefits of armor concentrate among the most exposed learners. The design and analysis plan will be pre-registered.

The primary outcome is a behavioral risk-taking index. It combines how often a student chooses the harder passage and how often they expose a provisional draft, which is the closest observable analogue of "showing unfinished work." Secondary outcomes are the following:

  • Final translation quality, scored by two blinded professional raters on the Content and Form dimensions of Yang's 100-point rubric.
  • Quality of the feedback each student gives: its prevalence under each of Luo's four toxicity categories and whether it contains a concrete suggestion. This is our measure of whether masks erode accountability.
  • Perceived emotional support, using Yang's item.
  • A presentation-autonomy scale adapted from Kim's instrument.
  • Weekly attrition.

Toxicity coding uses GPT-4 as a first pass. Luo et al. report its agreement with experts as Cohen's kappa 0.56, rising to 0.63 on high-confidence predictions. Because that agreement is only moderate, a stratified 20% sample is double-coded by humans, and any arm-level conclusions are checked on the human-coded subset. The analysis is a mixed-effects model with week nested in student nested in triad, with condition-by-week terms to detect novelty decay.

Baselines

The naive reference point is real-name triad peer assessment, which is Yang et al.'s experimental condition. On interpreting, that condition reported a gain from 74.67 to 82.73 on the 100-point rubric against 81.00 for teacher-led feedback. We cannot copy that number as a target, because our task is written translation and the ceiling may differ. So the real-name arm is re-run inside our study with the same passages, raters, rubric, and weekly schedule as every other arm. "Beating" it means a higher risk-taking index at equal or better final quality.

The strongest existing competitor is plain anonymity, the default remedy for evaluation anxiety in peer review. We searched for a reported effect of anonymity on risk-taking in learner peer assessment and found no reported result on Lacuna for it. Luo et al. show only indirectly that double-blind settings still produce toxic critique. Anonymity runs on identical infrastructure and differs from the self-authored arm only in what peers see. A self-authored persona beats it if the risk-taking index is reliably higher and feedback quality is no worse.

The second strong competitor is feedback detoxification. Luo et al. report that authors prefer the AI-rewritten professional version of a toxic critique 80% of the time. It is implemented with the same model, prompt, and flagging threshold in both detox arms, and gets the same pilot tuning budget we spend on the persona interface. The decisive ablation is assigned persona against self-authored persona, because it holds concealment and richness fixed and varies only authorship.

Why not just…?

Why not just make peer assessment anonymous? This is the main threat, and the design is built to answer it rather than argue with it. Anonymity has two known costs. Luo et al.'s OpenReview data show that removing identity does not remove harsh critique. Yang et al. credit their gains partly to partners not wanting to let each other down, a bond that a bare number may weaken. Neither cost proves anonymity is worse. If the anonymous arm matches the self-authored arm on risk-taking and feedback quality, the autonomy hypothesis fails and we would recommend the cheaper option.

Why not just detoxify the feedback? If learners avoid risk mainly because feedback is cruel, cleaning up the feedback should be enough, and the real-name-plus-detox arm tests exactly that. Two results would favor detox: real-names-plus-detox matching self-authored personas, and self-authored-plus-detox adding nothing beyond detox. The result that would favor our claim is an additive pattern. That pattern would mean exposure and harshness are separate threats.

Why not just remove the stakes with retries and gamification? We do remove the stakes, in every arm, so that this alternative becomes the floor rather than a rival. Metwally et al.'s evidence comes from two weeks of closed-answer tasks with an acknowledged novelty effect. Our 14-week horizon and condition-by-week terms show directly whether any persona benefit fades in the same way. If it does, we will report it as novelty, not protection.

Why not just train leaders to promise a safe climate? Brown and Zhang's finding that generic trust carries little weight next to concrete structures suggests that verbal reassurance is weak. We therefore keep instructor messaging identical across arms and do not test it as an arm.

Decisive Checks

The study succeeds on its central claim only if three things hold. The self-authored persona must exceed the assigned persona on the risk-taking index. That gap must be at least as large among high-anxiety or minority-in-cohort learners as in the overall sample. And the self-authored arm must not show more feedback flagged under Luo's taxonomy, or less constructive feedback, than the real-name arm. If the assigned and self-authored personas tie, protection comes from concealment, and the armor metaphor adds nothing beyond anonymity. If personas raise risk-taking but degrade the feedback learners give each other, then the tension between armor and the secure base is real. In that case the right design question becomes how much presentation control to grant, not whether to grant it. If detox alone matches personas, the practical recommendation is detox. Each of these outcomes is informative, and none can be produced by the prior studies, because none of them varied who controls a learner's presentation.