AI Should Flag, People Should Decide: What Proctoring Teaches Us About AI in Education

Most conversations about AI in education start with the chatbot and whether students will use it to write essays. That is a fair question, but it tends to skip a quieter one that matters just as much to the people who run universities, certification programs and corporate academies. When an AI system makes a judgment about a learner, who is accountable for it? 

I think assessment is where that question gets answered first, because it is where the stakes are easiest to see. A grade, a certificate or a failed exam can shape a career. So I want to use one concrete case, AI-assisted proctoring, to make a wider argument about AI in education. My view is simple. AI should flag, and people should decide. 

The trust problem is bigger than the technology 

Every institution that moved exams online learned that delivering a test remotely is the easy part. Convincing students, faculty, employers and regulators that the result means something is much harder. 

AI looks like an obvious help here. It can watch more sessions than any team of invigilators, and it never gets tired. But speed and scale are not the same as trust. If a learner is accused of cheating by a system they cannot question, the institution has not solved its integrity problem. It has swapped it for a fairness problem, and quite possibly a legal one. 

That is why I think the useful question for decision makers is not only “how accurate is the AI?” It is “what happens after the AI raises a flag, and who owns that moment?” 

What AI-assisted proctoring does well 

Constructor builds online proctoring software that universities, schools, certification bodies, and companies use for online, hybrid, and in-person exams. The AI in it is good at one thing in particular: noticing. During a session, it looks at signals such as face and voice detection, gaze tracking, device detection, and background noise. It can match a candidate’s face against their ID document, and a secure browser locks the test environment and blocks restricted applications. 

What comes out of all that is not a verdict. It is a set of flags, backed by a recording and a timeline of events, that a person can look at. Constructor’s own rule is that no candidate is judged by an algorithm alone. The company treats the AI’s output as a signal for a reviewer designated by the institution, and that reviewer confirms or dismisses each flag. 

Institutions also choose how much people stay involved. Some let the AI monitor every session and send flags to their own team. Others have a reviewer check each recording after the exam. Others keep live proctors watching in real time, with the AI helping them decide where to look first. Constructor’s position is that the institution, not the vendor, should make that trade-off. 

The reason this split works is that machines and people are good at different things. A model can spot that someone looked away from the screen forty times. Only a person can work out whether that was cheating, a nervous habit, a disability, or a younger sibling walking into the room. 

Human review is not a formality 

Saying “a human makes the final decision” is easy. Making that true takes work. If a reviewer has two hundred flags to clear before lunch and no context for any of them, the human in the loop becomes a rubber stamp, and the AI is effectively deciding anyway. 

Meaningful review needs a few things. Reviewers need evidence they can actually inspect, not just a risk score. They need time and training. Learners need to know what the AI does and what people do, and they need a clear route to challenge an outcome. European data protection law already points this way, since GDPR limits decisions about people that rest solely on automated processing. 

The guidance we give our clients follows the same logic: review flags before any sanction, offer escalation and appeal, tell exam takers about the role of AI, audit flagged incidents regularly and keep a record of how oversight works. None of this is glamorous, but it decides whether AI in assessment earns trust. 

Fairness and privacy are design choices 

People sometimes talk about bias and privacy as risks that arrive with AI and have to be managed afterwards. I see them as product decisions made early. In our case that means demographic-blind flagging and regular bias audits on our side, and a recommendation that institutions audit their own flagged cases for fairness too. 

On privacy, the principle is to collect the minimum data each exam scenario needs, with the test taker’s consent. Not every exam needs a webcam, so there is a lockdown-only mode without video. Where data is stored matters as well, which is why regional hosting and on-premise deployment are options for institutions with strict rules. 

Regulation is moving in the same direction. Under the EU AI Act, Constructor acts as the provider of the AI system, with obligations to match. Beyond the law, UNESCO has spent the past few years arguing for a human-centred approach to AI in education, including guidance on generative AI and AI competency frameworks for students and teachers. That framing puts the learner’s rights at the centre, not the tool. 

Where AI in education is heading 

AI is moving into assessment design, grading, tutoring, course authoring and even timetabling. I expect the same tension to show up in each of them, between what AI can do quickly and what institutions can stand behind. 

My prediction is that the institutions that do well with AI will not be the ones that automate the most. They will be the ones that are clearest about where the machine stops and the person starts, and that can explain that line to a student, a parent, an auditor or a court. 

For buyers, that suggests a short set of questions worth asking any vendor, including us. What does the AI decide, and what does a person decide? Can a learner see and challenge the evidence? What data is collected, where does it live and for how long? How is bias checked, and by whom? 

If a vendor cannot answer those clearly, no accuracy figure should make up for it. AI can make education fairer and more efficient, but only if people stay responsible for the decisions that matter.