¿Qué examen es recomendable en la era de la IA?
We may need to rethink relying on traditional handwritten exams
Autor
Students sat traditional exams this year without the assistance of AI, a tool that many have come to rely on in their vidas cotidianas. Handwritten, supervised exams are one of the most reliable assessments of what students can achieve unaided. But this leaves us in an odd position in a world now saturated with AI, says Rosa suerte, Professor Emerita at University College London. For decades, Luckin has designed and evaluated technology and AI for use in education. “We end up certifying the skill of working without the tools they will use every day and certifying nothing about whether they can use those tools well,” she says.
Are educational assessments fit for their intended purpose?
El purpose of assessments is to give the learner and teacher an accurate picture of what the learner does and doesn’t understand, so both know where to focus their attention. This goal could be undermined by technology that is capable of completing homework, writing essays, and doing coursework.
A handwritten test should be just one instrument of assessment, Luckin argues, and not the only one. “Retreating entirely to 1995 does not solve the problem; it postpones it and narrows what we are willing to value.”
The best assessments make thinking visible, she says. They require the student to explain their reasoning aloud, defend a position under questioning, and produce work in stages with checkpoints.
“The best assessments make thinking visible.”
Alex Scharaschkin, chief assessment research officer at exam board AQA, agrees that it’s important to evaluate what’s behind the output. “Can we do more to elicit information about how students are doing along the way as they develop those skills, over and above looking at the final project?”
Assessing students amid the upheaval wrought by AI is keeping policymakers as well as teachers and educators busy. In England, the government is reviewing curriculum and assessment, with changes coming in 2028, but Scharaschkin expects that formal exams will still contain a large traditional component.
“But could we use digital technology to change the nature of some of these assessments – using more simulation-based or more spoken viva-type assessments, potentially with some AI mediation?” he asks. He adds that exams should probably also test whether learners are able to use AI tools intelligently and appropriately.
How important is knowledge in an age of generative AI?
Core curriculum knowledge remains valuable when students deploy AI tools, says Scharaschkin. The less students know, the less they can reliably outsource to AI, because they are less able to determine whether the AI response is reasonable.
Students need to understand the extent of their own knowledge. Luckin gives the example of a science student noticing when the outcome of an experiment doesn’t fit the hypothesis, or a history student explaining why one source is more trustworthy than another. Teachers might use the ‘exit ticket’ strategy developed by history teacher David Owen. In this strategy, as Luckin explains, students ask themselves at the end of a lesson what stopped their learning, what new question, idea or perspective they considered, and what they have understood and learned. The second question – which Owen refers to as the “middle ticket” – matters most because it asks students how they are learning, which is the bit that AI can’t do.
“Core curriculum knowledge remains valuable when students deploy AI tools.”
How can teachers work with AI?
AI detection is unreliable, so teachers can’t always spot who hasn’t completed their own work. Luckin therefore suggests that teachers “assess the process, not only the product”. Look at drafts, work in progress, and the AI prompts a student uses. Try to build in short spoken checkpoints. “A five-minute conversation…will tell you more than any detector, because understanding can’t be outsourced to a chatbot in real time.” To avoid ambiguity, she advises making it clear when AI is allowed, where it isn’t, and which tasks make the collaboration with AI the point. AI will be less useful for tasks that refer back to class material or to what is local, recent, or experienced, so these can be good tasks to set.
Some educators suggest asking students to critique an AI-generated answer that is confidently wrong about something they are studying. This could prompt a student to examine critically a source in history, or to test a result in science; the student does what AI can’t by deciding what is actually true.
Educating students about how a large language model works could help them learn to be judicious in their use of it. Teachers can explain that an LLM may appear to be thinking, while it is in fact producing words based on patterns, and that it can flatter users. “The ability of students to interact critically with those models is probably the key thing for effective learning,” says Scharaschkin.
“When used critically and effectively, AI tools can be useful at mashing up or recombining stuff that’s already ingested by the internet and giving it to you in another form,” he explains. “But they don’t do the important parts of intelligence, such as empathising with someone’s point of view, seeing something through somebody else’s eyes, or changing as a result of what you have learned. These things are going to be very hard to outsource.”
“Ultimately, students need to understand what knowledge is, where it comes from, and how to learn about the world in a flexible way.”
AI will be with us forever, and education and assessment must adapt, says Scharaschkin. AI is shaping our understanding of what it means to be educated. Without a settled model for assessment, and with academic honesty under the spotlight, many are arguing for clearer policies, better authentication of student understanding, and tasks that can show reasoning and understanding of the subject matter.
“There are certainly opportunities we are trying to explore in terms of how generative AI can help to improve how we do assessment,” says Scharaschkin. “But AI is also a threat because the longer-term impacts upon things such as students’ cognitive development are unknown, and that needs to be explored.”
Education must include AI literacy and data literacy, says Luckin, with a focus on learning mastery and knowledge mastery. Ultimately, students need to understand what knowledge is, where it comes from, and how to learn about the world in a flexible way.
Autor