AI for Decision Makers
Unit 6: Education
Education uses prediction to sort, support, and certify people at moments that shape their lives. It also has a measurement problem at its core: we care about learning, but we observe grades, test scores, clicks, and completion. AI systems optimize whatever is measured, so in education the first question is always whether the measure is the goal.
The Institutional Landscape
| Application | What is predicted or produced | What remains a judgment |
|---|---|---|
| Automated essay scoring | The score a trained human rater would give | Whether that score reflects the skills the course is meant to teach |
| Early warning systems | Which students are likely to fail or drop out | Which intervention to offer, and to whom it will actually help |
| Adaptive practice | Which skills a student has mastered | Pacing, pedagogy, and when to involve a teacher |
| Admissions screening | Likelihood of completing or succeeding | What the institution is selecting for, and how to weigh other goals |
| Generative AI tutors and assistants | Explanations, feedback, worked solutions | When help supports learning and when it replaces it |
Decision: Assessing Students
Assessment shows why the human baseline matters. Teacher grading is noisy and can be biased. Students' own work is increasingly produced with AI help. Grades feed high-stakes decisions about progression and admission. Each stage below has a human failure mode as well as an algorithmic one.
Student Assessment Through the Five Stages
| Stage | Traditional | AI-Mediated | Risk |
|---|---|---|---|
| 🔍Framing | Teachers design assessments around learning objectives, with uneven validity | Assessments drift toward formats a model can score reliably | What is easy to score crowds out what matters to learn |
| 📊Prediction | Raters disagree; expectations about a student can colour grading | A model predicts the score a rater would give from features of the response | Models can reward surface features (length, vocabulary) and be gamed |
| ⚖️Decision rule | Grade boundaries and moderation set by examiners and teachers | Statistical standardization or automated cut-offs | Group-level adjustments can be imposed on individuals who are not typical of their group |
| 🎯Action | Feedback, progression, and certification | Automated grades, routing to modules, or flags for intervention | Life-shaping outcomes without a meaningful route to appeal |
| 🔄Feedback & adaptation | Teachers adjust instruction informally | Platforms optimize engagement and completion metrics | Engagement is not learning; students and teachers adapt to whatever the system rewards |
Case Study: England's 2020 A-Level Grades
The 2020 A-Level Standardization
With exams cancelled by COVID-19, schools in England submitted a predicted grade and a rank order for each student. The regulator, Ofqual, then standardized those grades with a statistical model anchored to each school's historical results. On results day, 39.1% of grades in England were lower than the teachers' assessments. After protests, the government reversed course within days and awarded the teacher-assessed grades, or the calculated grade where that was higher.
| Change from teacher assessment (England) | Share of grades |
|---|---|
| Lowered by one grade | 35.6% |
| Lowered by two grades | 3.3% |
| Lowered by three grades | 0.2% |
| Unchanged | 58.7% |
| Raised | about 2.3% |
Source: Ofqual results-day statistics, August 2020, as reported by Schools Week and Tes
The model existed to serve a stated policy goal: keep 2020 grades comparable with other years. Ofqual estimated that unstandardized teacher assessments would have raised A-level results by around 12 percentage points over 2019. To achieve comparability, the model distributed each school's grades according to its past performance. That systematically penalized strong students at schools with weak track records. Because the model was unreliable for small classes, teacher grades carried more weight there. Critics showed this favoured centres with small classes, which are disproportionately independent schools.
Lessons
- Imposing a school's historical distribution on individual students guarantees errors for students unlike their school's past cohorts
- Design details such as special rules for small classes have distributional effects that should be analysed before launch
- The failure was as much about legitimacy and recourse as about statistics: individuals could not effectively contest a grade derived from other people's past results
- The reversal was not costless: it replaced one set of distortions (downgrading) with another (inflation)
Early Warning Systems: Predicting Risk vs. Helping Students
Early warning systems flag students likely to fail or drop out so schools can intervene. They are prediction tools in the service of a causal goal. As Unit 3 showed, the students at highest risk are not necessarily those a given intervention helps most. A flag can also change how teachers see a student. Ask three questions of any proposal. What intervention follows a flag? What evidence, ideally from a randomized evaluation, shows it helps flagged students? Is the system evaluated on students' outcomes rather than on how accurately it flags them?
Automated Essay Scoring and Validity
Automated scorers can agree with human raters about as often as raters agree with each other. That establishes reliability, not validity. Agreement can rest on features like length and vocabulary that correlate with quality without constituting it. Les Perelman's BABEL Generator famously produced grammatical nonsense that some automated scorers rated highly. Once students know what a scorer rewards, they adapt (Unit 4's interdependence). Low-stakes formative feedback is a far safer use than high-stakes certification.
Generative AI: Assessment and Tutoring
Generative AI breaks a long-standing assumption: that a take-home essay or problem set is evidence of an individual's skill. Institutions now face a redesign question rather than a detection question. Which assessments must be supervised, oral, or iterative, and which can openly allow AI use?
Tutoring evidence shows why guardrails matter. In a randomized study of nearly a thousand high-school mathematics students, access to GPT-4 during practice raised practice performance sharply (48% with a standard chat interface, 127% with a tutor designed to give hints rather than answers). But on a later exam without AI, students who had used the standard interface scored 17% worse than the control group. Those who used the guarded tutor scored about the same as the control group (Bastani et al. 2025). Help that substitutes for thinking can raise measured performance while reducing learning.
EdTech Platforms as Markets
Platforms such as Khan Academy, Coursera, and Duolingo match learners with content at a scale no institution could. In the sense of Unit 1, they are market-making systems, and their design reflects their business model. Where revenue depends on engagement or retention, the platform is rewarded for time-on-app and streaks, which are imperfect proxies for learning. When procuring or recommending a platform, ask what it optimizes and what independent evidence shows about learning outcomes.
Key Takeaways
- In education the measured target (scores, engagement) is rarely the goal (learning). Check the proxy before trusting the optimization.
- The 2020 A-level episode was a value conflict between individual and cross-cohort fairness, decided inside a model without real recourse. Both the model and the reversal had costs.
- Early warning systems serve a causal goal; judge them by student outcomes and by evidence that interventions help those flagged.
- Agreement with human raters is reliability, not validity; keep automated scoring out of high-stakes certification unless validity is established.
- Generative AI can raise performance while reducing learning. Measure learning without the tool, and design assistance that prompts thinking rather than replacing it.
Evidence & Further Reading
- Ofqual (2020). Awarding GCSE, AS, A level, advanced extension awards and extended project qualifications in summer 2020: interim report; results-day statistics (August 2020).
- Bastani, H., Bastani, O., Sungu, A., et al. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122.
- Perelman, L. (2014). When 'the state of the art' is counting words. Assessing Writing, 21.
- Athey, S. (2017). Beyond prediction: Using big data for policy problems. Science, 355(6324), 483–485.