Trust is the missing piece in AI disclosure

Audio briefing based on Student Voice Weekly issue #9.

This Week

This week, the episode discusses genAI trust, QAA benchmarks, and response bias. 739 students show why trust matters for AI disclosure The main topics are grouped below by student voice practice, research, sector developments, archive context, and practical application.

Main Topics Discussed

Student Voice Practice

After my travel last week, it's good to be back in the office.

Research Spotlight

Sector Watch

QAA Subject Benchmark Statements, and what they mean for student feedback evidence
QAA's GenAI assessment focus groups show why student voice on AI needs more structure

From the Archive

Practical Application

Overall sentiment averages are rarely enough when teams need to understand whose experience is changing.

Subscribe to The Student Voice Weekly: https://www.studentvoice.ai/blog/newsletter/

Transcript

Hello, and welcome to Student Voice Weekly. I'm Dr Stuart Grey, founder of Student Voice, and today's theme is AI disclosure as a trust problem: students tell you what they did when they think you will treat them fairly.

Today I'd like to talk about something I think a lot of universities are misreading. We are treating GenAI disclosure as mainly a policy communication job. Write the guidance, publish the flowchart, tell students to declare their use, and then wonder why disclosure is low, or why declarations feel vague and defensive.

The key thing is simpler. Students do not disclose when they do not trust the system.

And by trust, I mean practical, day to day signals. Do staff interpret the rules consistently. Do students think they will be treated fairly if they are honest. Do they think disclosure triggers suspicion. Do they believe the process can tell the difference between acceptable support and misconduct.

That is the main story this week, and it is backed by a useful study. Xia and Wei tracked 739 students across 25 courses over four waves. The headline finding is straightforward: students were more willing to disclose AI assistance when they believed governance would treat them fairly.

What predicted that sense of fairness was not just whether the policy was written clearly. It was implementation fidelity, meaning how the policy shows up in real teaching, real assessment, and real conversations with staff.

That matches what students are actually saying. They rarely say, the document was unclear. They say, different tutors say different things. Or, I do not want to be accused. Or, I asked a question and got a weird reaction. Or, it felt like the marking punished it.

So why does this matter right now.

Because we are moving from the early AI panic phase into routine governance. The risk in the routine phase is that inconsistency becomes normal. Students learn that disclosure is not a neutral administrative step. It is a social risk. It exposes them to a judgement call, and they do not know how that judgement call will go.

And there is a second pressure. Course teams are being asked to demonstrate standards and comparability. Not just do you have a policy, but does it work in practice. Does it produce the behaviours you intend. Does it protect standards without creating a culture of fear.

A common mistake is to push harder on compliance messaging. More warnings, more policing language, more threat framing. In practice that can reduce disclosure because it raises the perceived cost of being honest.

Instead, ask a different question: where, in the lived assessment process, does trust get built or broken.

This shows up in ordinary moments. A student asks, can I use AI to plan. Can I use it to check my writing. Can I use it to debug code. They are not only asking about the rule. They are testing the reaction. Are you going to help me do this properly, or are you going to treat me like a risk.

So if you own AI guidance at institutional level, make sure you assume the real policy is the one enacted at course level. A central document does not create trust on its own. Staff confidence, consistent interpretation, and predictable processes do.

Here is a practical way to use the Xia and Wei study as a diagnostic. The concept to focus on is procedural justice. In plain terms, do students experience the process as fair, consistent, and safe.

You can turn that into something operational quickly. In a pulse survey, a module evaluation, or a targeted check in, ask a small number of pointed questions:

I understand what I am expected to disclose about AI use.
I believe disclosures are treated fairly and consistently.
I feel safe asking staff for advice about AI use.

Then look hard at the comments. The comments tell you what the fear is. Is it fear of accusation. Is it fear of inconsistent marking. Is it a sense that disclosure will be used against them. Or is it confusion about boundaries.

Now I want to connect this to the other research spotlight in the issue, because it affects how you interpret your evidence.

De Bruin, Owen, and Wu look at non-response bias in teaching evaluations. The reminder is that your feedback system is only as strong as the voices it captures. They show, using a randomised experiment, that evaluation design shapes not just response rates but who responds. Some groups are less likely to complete traditional evaluations, including lower-income students, students with lower GPAs, later-year students, and some minority ethnic groups.

Why does that matter for AI disclosure and trust.

Because if you build AI governance mainly on the feedback of the most confident and engaged students, you can end up designing for the wrong student. The students who feel least secure are often the ones least likely to add extra risk by disclosing more detail about their process. So make sure representativeness is part of your AI evidence plan. Do not only ask, did we get a decent response rate. Ask, who is missing, and what might their experience of risk and trust look like.

Now sector watch. There are two QAA signals worth noting.

First, QAA has revised Subject Benchmark Statements in a set of subject areas. Second, QAA is running focus groups on GenAI and assessment, explicitly bringing student evidence into the standards and assessment design conversation.

The key thing here is what this prompts internally. Course teams are being pushed again towards evidence-based curriculum and assessment decisions, including student evidence. That changes the conversation from, have we got a policy, to, can we show that students experience our assessment approach as coherent and fair.

So if you are a programme leader looking at benchmarks, do not treat AI as a separate policy track. It is going to show up in assessment design, skills development, and questions about what good academic practice looks like now that tools are normal.

And the QAA focus groups matter because they signal that student voice on AI is moving from local debate to sector-level expectations. If you are reviewing AI guidance, make sure you ask students before local rules harden into routine. Once staff have been teaching a policy for a year, it becomes difficult to change even if it is not working.

Here is what I do when I read student comments on this topic. I separate three signals that often get bundled together.

First, clarity comments. Students saying they do not know what is allowed, the guidance is vague, or different modules say different things.

Second, fairness comments. Students saying it depends who marks you, it feels inconsistent, or they are being assessed on something they were not taught.

Third, safety comments. Students saying they are scared to ask, they do not want to disclose because it will raise suspicion, or they feel assumed guilty.

Those three signals need different fixes. Clarity is mainly communication and staff briefing. Fairness is about consistent marking practice and assessment design. Safety is about culture and process, including what happens after disclosure and whether students think it changes how they are judged.

And this is the point I want to land. A university can have a clear policy and still produce unsafe experiences. If disclosure feels like self-incrimination, students will avoid it.

One thing to try this week.

In your next meeting about assessment, feedback, or AI guidance, do a quick two-column exercise. On the left, list the steps in the student journey where AI use might occur: planning, drafting, editing, referencing, coding, analysis, reflection. On the right, write what you currently ask students to do at that step: nothing, disclose, cite, reflect, or seek permission.

Then ask one simple question: at which step would an honest student feel most anxious about being misunderstood.

That is your trust pinch point. That is where you tighten consistency, give staff a usable script for responding to questions, and check what students are actually saying in comments, not just in tick-box declarations.

That is it for this week. The full set of links and summaries is in Student Voice Weekly.