The Research Behind CoStudy

Most peer evaluation tools treat the evaluation as a measurement device: collect ratings, generate a score, move on. CoStudy is built on a different hypothesis. Evaluating peers requires students to examine the same behaviors in themselves — closing the gap between self-perception and reality. It's not just a way to measure collaboration. It's how students build the skills that make collaboration work.

The concept traces back to Sadler's work on evaluative judgment^{Sadler, 1989}: the capacity to make quality judgments about your own and others' work is a core learning outcome, not a byproduct of assessment. When students evaluate their teammates, they practice identifying what effective contribution looks like, articulating it, and calibrating their standards against the group's.

Cho and MacArthur found that students who gave feedback outperformed those who only received it.^{Cho & MacArthur, 2011} The act of reviewing, of having to articulate what's working and what isn't, forced deeper processing than simply reading someone else's comments about your work. Wu and Schunn confirmed this in a large-scale meta-analysis: providing feedback shows stronger learning gains than receiving it.^{Wu & Schunn, 2023}

This is CoStudy's core design philosophy. Every evaluation cycle isn't just collecting data for the instructor. It's training students to recognize, articulate, and improve collaborative behavior. The measurement and the learning are the same activity.

Edmondson's foundational work^{Edmondson, 1999} identified the conditions under which teams perform: members need to feel safe to take interpersonal risks: asking questions, admitting mistakes, offering honest feedback without fear of punishment. Google's Project Aristotle confirmed this at scale, finding psychological safety was the single strongest predictor of team effectiveness.^{Duhigg, 2016}

You've heard that before. Here's what most accounts leave out: Edmondson's conditions aren't personality traits or team culture vibes. They're structural. Safety emerges when the environment reduces social risk, normalizes vulnerability, and separates feedback from personal judgment. In student teams, where power dynamics, social hierarchies, and grading pressure are all amplified, these conditions don't happen by accident. They have to be designed in.

CoStudy maps three specific design decisions directly to Edmondson's conditions:

The conversation about AI in higher education has mostly been defensive: How do we detect AI-generated work? How do we prevent cheating? Those are real concerns, but they miss the bigger question: what skills do students need to work effectively with AI systems, and how do we teach them?

Tai et al. identified evaluative judgment, the capacity to assess the quality of work, your own and others', as a critical learning outcome in higher education.^{Tai et al., 2018} Bearman et al. extended this argument directly to the AI era: evaluative judgment is the skill AI cannot replace, and the one students need most when AI can generate plausible-looking output on demand.^{Bearman et al., 2024} Students who can't judge quality will accept whatever AI gives them. Students who can will know when to trust it and when to override it.

This is where peer evaluation becomes forward-looking, not just defensive. Every time students evaluate their teammates' contributions, they practice exactly the skills Tai and Bearman describe: assessing quality, calibrating standards, articulating what “good” looks like, and giving honest critical feedback. These are the same capacities needed to collaborate effectively with AI systems.

Peer evaluation also provides something artifact-based assessment has lost: direct observation of the collaborative process. AI can generate a paper or a presentation, but it can't fake who showed up to meetings, who contributed ideas, and who did the work. That process-level visibility remains authentically human.

It's the most common objection to peer evaluation, and it's the right question to ask. If students are rating general impressions (“How good a teammate is Alex?”), then yes, you get friendship bias, halo effects, and results that track social standing more than actual contribution.

But that's not what happens when you use behavioral criteria. Falchikov and Goldfinch's meta-analysis of 56 studies found that peer ratings correlate with instructor ratings at r = 0.69 when behavioral criteria are used.^{Falchikov & Goldfinch, 2000} That means when students evaluate specific observable behaviors instead of general impressions, their ratings line up with the professor's assessment nearly 70% of the time. This isn't a vibe check.

“Behavioral anchoring” sounds technical, but the idea is simple: rate what you observe, not what you feel about someone. Instead of “Was this person a good team member?” (which invites bias), you ask “Did this person come to meetings prepared with their assigned work?” or “Did this person respond to team messages within 24 hours?” The question targets a specific, observable action. There's no room for a halo effect when you're answering a factual question.

CoStudy's questions are designed around this principle. Every item targets an observable behavior. Combined with anonymization (so social pressure doesn't distort ratings) and multi-rater aggregation (so no single evaluator's bias dominates), the result is assessment data that holds up to scrutiny, not a popularity contest.

CoStudy's core assessment was originally developed through a composite analysis of peer evaluation instruments across higher education, refined in collaboration with a doctoral-level psychologist. The goal was to isolate the behavioral dimensions that most reliably differentiate effective from ineffective team contribution, and to ask about them in ways that reduce rater bias.

That foundation still holds, but teaming doesn't look the same across every context. An engineering capstone, a humanities seminar, and a clinical practicum each define “contribution” differently. That's why we work directly with professors to create custom question sets aligned to specific course objectives. The accountability and growth mechanisms should reflect what actually matters in a given learning environment.

We're also transparent about what's ahead: we are currently preparing a major overhaul of our primary assessment instrument. The goal is stronger psychometric grounding, tighter alignment with current teamwork research, and better adaptability across disciplines. We'll share more as that work matures.

Woolley et al. published a striking finding in Science: group intelligence isn't predicted by the smartest person in the room. It's predicted by social sensitivity, turn-taking equality, and gender diversity.^{Woolley et al., 2010} Teams with one star performer and poor collaboration consistently underperformed teams with moderate individual talent and strong collaborative dynamics.

Riedl et al. replicated this across 5,279 individuals in 1,356 groups, confirming the collective intelligence factor holds at scale.^{Riedl et al., 2021} The finding is robust: what makes teams effective is not the sum of individual ability but the quality of interaction between members.

The skills that predict team effectiveness (social sensitivity, equitable participation, constructive disagreement) are exactly the ones CoStudy's evaluation cycles develop and measure.