Was this forwarded to you by a friend? Sign up, and get your own copy of the news that matters sent to your inbox every week. Sign up for the On EdTech newsletter. Interested in additional analysis? Upgrade to the On EdTech+ newsletter.

Due to the importance of the first topic on AI and assessment, I am sharing the full first section in front of the paywall. [-Phil]

For those of us in the US, that was quite a week. Somewhere between the polls, the results, and the election postmortems, I even managed to read a little about EdTech.

I finally got around to reading Wonkhe's report on assessment, and I should have read it earlier because it's really good.

Assessment has become one of the most fraught areas of higher education's encounter with AI. It is where some of the technology's most troubling consequences, including cheating and cognitive offloading, are playing out. It is also where some of our worst institutional instincts are on display, from increased surveillance to ham-fisted attempts at detection and punishment. And, perhaps most interestingly, AI is exposing longstanding weaknesses in how we assess learning, raising uncomfortable questions about what our assessments actually measure and whether they were doing a particularly good job of it in the first place.

Based on a large survey of UK higher education students, supplemented by focus groups, Wonkhe’s report gets beneath the AI adoption statistics that fill my inbox every week. The researchers ask whether students believe they have learned what their assessments supposedly demonstrate, whether grades reflect their understanding, and what influences their decisions about using AI.

The authors describe some of the findings as "uncomfortable," and I can see why. Almost half of students (47%) worry that their grades don't fully reflect what they know, while 38% say they sometimes submit work they cannot fully explain without going back to their sources.

One of the most interesting findings concerns what students believe their courses reward. Only around one in five think their courses primarily reward thinking and reasoning rather than producing outputs such as essays or exam answers. As one computing undergraduate put it:

And this appears to matter for AI use. Among students who say their courses primarily reward final outputs, 71% report using AI for assessment. Among those who say their courses primarily reward thinking and reasoning, the figure is 38%.

That's a substantial difference, although we should be careful about interpreting it. The study doesn't establish that emphasizing thinking and reasoning causes students to use AI less. But it does suggest that students' perceptions of what their courses value are associated with how they approach AI.

The report also finds that students who understand their assessment criteria report substantially less AI use than those who find the criteria unclear.

What I particularly like about this research is that it shifts our attention away from simply identifying which students are using AI and toward examining how the design of courses and assessments might influence their behavior.

The report is especially useful because it offers concrete recommendations that go beyond the usual invocation of one of the Three Bromides of AI in Learning: design more authentic assessments (the other two being "keep a human in the loop" and "we need more AI literacy").

Instead, the authors emphasize what they call "accountability moments": opportunities within assessment when students must actively demonstrate their understanding. The idea is to make thinking visible during the learning process rather than relying entirely on a final product. They also emphasize timely, actionable feedback so that students can use it to improve their work, rather than receiving comments after the opportunity to learn from them has passed. Formative assessment should actually be formative, rather than summative assessment in disguise.

I like these recommendations, although the report is less helpful on how to introduce accountability moments at scale, particularly in large courses with limited faculty capacity.

Another intriguing finding concerns belonging. I usually encounter belonging in the student success literature, where it is associated with persistence and the likelihood of stopping out. Here, the researchers find that students' sense of belonging is associated with striking differences in how they perceive AI and its potential benefits.

The differences are striking. Among students who strongly disagree that they belong to a community, 98% agree that AI helps them learn more, compared with just 28% of students who strongly agree that they belong. These findings concern students' perceptions of AI's benefits as much as their actual use of it.

The authors suggest that students who feel they belong may be less inclined to turn to AI because their educational experience already provides some of what AI substitutes for. I find that fascinating, although the evidence doesn't establish that belonging itself causes these differences. Students who feel a stronger sense of belonging may also differ in their engagement, motivation, academic experiences, or access to support. Still, I appreciate that the researchers are looking beyond individual decisions about whether to use AI and considering how the educational environment might shape those decisions.

The report is a welcome contribution to a conversation about assessment. If AI finally forces us to reconsider what our assessments reward and whether they genuinely demonstrate learning, something worthwhile might come out of this.