Black and Asian students use AI most carefully, and punitive policies put them most at risk

I was looking at how students balance AI against traditional sources like textbooks and lecture notes – and Black students lean toward AI at 51 per cent compared to 33 per cent for White students, a gap larger than any gender, subject, or institutional difference the report discusses.

Black and Asian students are significantly more likely to see AI skills as essential to their futures, more likely to want their institution to provide AI tools, and more likely to use AI for summarising and studying textbooks.

The obvious question is whether this is really about ethnicity or whether it’s actually about something else – subject choice, socio-economic background, institution type, or geography.

But subject composition only seems to explain a fraction of the gap, socio-economic grade barely shifts the needle, and Black students in this sample actually have the highest ABC1 share of any ethnic group.

A London effect is real – students at London universities lean more toward AI, and Black students are disproportionately London-based – but I figure that accounts for roughly 3 points of an 18-point gap.

Even after estimating the impact of every available confounder, a big residual persists – and the motivational data suggest why.

Black students consistently cite AI as a source of access to support and content when traditional structures fall short — broadening available materials, getting help outside study hours, finding instant answers when no one is around.

Only 2 per cent of Black students express no interest in AI tools, compared to 8 per cent of White students. In other words, this isn’t a group experimenting with a novelty – it’s a group that has integrated AI into how they study, apparently because the alternatives are not meeting their needs.

At the Secret Life of Students last week, we published our own look at students and AI – although we ask different questions, on assessment, learning, the hidden curriculum, and belonging. The full report will be out in your Monday Briefing.

And guess what. Several of the demographic patterns that were buried in HEPI’s cross-tabulation data appear in our data too, and are often amplified.

Motivations

In the Kortext/HEPI data, Black students’ motivations clustered around access – broadening content, getting out-of-hours support, instant answers. Asian students were more intellectually driven, drawn to AI because they believed they learned more with it, but also more ethically conscious.

Our data can’t replicate this directly, but the attitudinal questions tell a similar story. The statement “AI helps me learn more effectively” draws agreement from 67 per cent of Asian students, 41 per cent of Black students, and 25 per cent of White students.

Asian students aren’t just using AI more – they believe, at much higher rates, that it is making them better students.

The access-and-support story from the HEPI/Kortext Black student data shows up differently here. Black students in our data are the most likely group to perceive a gap between what their course says it values and what it actually rewards – 52 per cent agree, compared to 43 per cent of Asian students and 28 per cent of White students.

They are also less likely than any other group to say they are “learning what I need to learn, not just producing what I need to produce” – just 52 per cent agree, versus 73 per cent of White students and 87 per cent of Asian students.

If you are a Black student who feels your course rewards something other than what it claims to value, and who doesn’t feel you’re learning what you need to learn, AI may be a rational response to a system you perceive as misaligned.

The ethics inversion

When students are asked what they consider when deciding how to use AI on a specific assignment, the groups that use AI most are also the groups that think most carefully about it. 94 per cent of Black respondents who answered this question ticked the assignment guidelines. Seventy-nine per cent ticked what their tutors had said. Seventy-nine per cent thought about what felt right to them ethically. Fifty-nine per cent considered whether they’d be able to explain their work if asked.

Meanwhile, for White students, 36 per cent checked guidelines, 31 per cent checked tutor guidance, 40 per cent thought about ethics, and 21 per cent thought about explainability.

It inverts a common anxiety about AI and assessment integrity. The students using AI most heavily are not cutting corners – they are the most ethically engaged, the most attentive to institutional guidance, and the most conscious of whether they could defend their work.

The vast majority of Black students who say they feel completely confident explaining their AI use to a tutor reinforces this – these are not students hiding what they’re doing. They’re doing it openly, thoughtfully, and – by their own account – within the rules. Asian students are almost as confident, at 87 per cent. White students – 55 per cent.

Belonging and AI have a notable relationship. Students who strongly disagree that they feel part of a community use AI at the highest rate by belonging category. The students who feel most disconnected from their academic community are the ones most likely to turn to AI in their work, just as HEPI found the most isolated students most likely to turn to AI for support.

In their own words

In the free text comments, BME students discuss a qualitatively distinct relationship with AI – they describe using it as a learning companion, using it to check understanding, interpret data, and navigate referencing conventions – and then worry about whether that specific, learning-oriented use crossed a line.

White students are more likely to describe using it for brainstorming and planning, with less of the push-pull anxiety. The other major free text difference is on assessment redesign, where BME students converge heavily on wanting practical, applied testing, while White students’ suggestions scatter more evenly across formats like portfolios, vivas, and open-book exams.

The tone and texture of the comments differ too. White students tend towards longer, more evaluative language – more likely to use words like “engaging” or “stimulating,” or to make broad philosophical statements about AI’s role. BME students are more likely to describe what they actually did with AI and ask whether it was OK – White students are more likely to declare a position on whether AI should or shouldn’t be used at all.

I used AI to help me interpret and describe the trends in data from a biology practical. I wasn’t sure if this crossed a line, as analyzing and explaining results is usually expected to reflect my own understanding. Black student

I’ve opted to never use AI, both from academic reasonings but also due to concern for the environment!!” White student

I would redesign one of the assessments to focus more on practical, real-world projects instead of only written assignments. For example, creating a working software application, solving a real computing problem, or developing a small system. Asian student

I consider my use of AI to always be acceptable. Guidance prohibiting it is unproductive. Students should be taught how to use and critique it effectively. White student

What the comparison tells us

Three things stand out from reading these two datasets side by side.

First, the ethnicity gap in student AI use is not an artefact of one survey, one question type, or one sample. The underlying finding is consistent – the relationship between ethnicity and AI is large.

Second, our data surfaces the ethical consciousness of the heaviest AI users. The assumption that more AI use means less careful use is not supported. The opposite appears to be true.

This matters for how institutions frame their AI policies. If the most prolific AI users are also the ones most attentively reading the guidelines, checking with tutors, and thinking about ethics – and if those users are disproportionately from ethnic minority backgrounds – then punitive approaches to AI in assessment carry equality risks that go far beyond who gets caught.

Third, the loneliness-AI relationship is not a single story. In the HEPI data, it’s a story about isolation and coping. In our data, it’s a story about confidence and capability. The difference almost certainly reflects what’s being measured – AI for emotional support versus AI for academic work.

But it complicates any simple narrative about AI and students. The students turning to AI for companionship are not the same population as the students turning to AI to improve their grades, even when some belong to both groups. Policy responses need to be equally differentiated.