When AI Predict di Exam: Wetin 90% Accuracy Really Show About Broken Systems
When I dey work with organisations on Digital Transformation and Business Process Re-engineering, I dey always ask one question early: how much of your planning dey depend on historical pattern repetition instead of real forward reasoning?
One group of 18-year-old students for UK recently answer dat question with more force pass any boardroom presentation wey I don see.
Dem use 20 years of GCSE and A-Level exam papers, mark schemes, examiner reports, and curriculum documents, build AI system wey predict exam questions with over 90% accuracy. Last year di system reach 85%. Dis year e pass 90%. Di improvement no be by mistake. Na feedback loop dey work exactly as dem design am.
Most people dey read dis as story about AI. I dey read am as story about systems.
Di Architecture No New
Make we dey precise about wetin dem build. Di system dey lean on three layers: data ingestion layer wey dey process two decades of historical papers and examiner commentary, pattern recognition layer wey dey use large language models to model topic cycles and coverage gaps, and probabilistic output layer wey dey produce ranked predictions wey dem weight by historical recurrence.
Dis na Retrieval-Augmented Generation wey dem apply to temporal pattern prediction. Di same architectural logic dey show for financial forecasting, retail demand planning, fraud detection, cybersecurity analytics, and supply chain optimisation.
Di domain dey change. Di mathematics no dey change.
Wetin 90% Accuracy Actually Expose
If system wey dem train on historical data fit predict examination questions with over 90% accuracy, dat one no be evidence of sophisticated AI first. Na evidence of structurally deterministic assessment system.
Well-designed examination suppose dey highly resistant to historical pattern prediction. If 20 years of data fit tell you wetin go dey next year paper with 90% confidence, di system dey measure recall and pattern matching, no be genuine understanding.
Di students no break di examination system. Dem make wetin don already break dey visible.
Di same flaw dey inside business. Annual planning cycles wey dey reproduce last year assumptions. Performance frameworks wey dey reward behaviours wey dem value for previous competitive environment. Strategic intelligence processes wey no get mechanism to detect structural change. Dis ones no be different problems. Na di same problem for different domains.
Wetin Organisations Must Now Face
Di organisations wey go build durable competitive positions no be di ones wey dey restrict wetin dia people fit use. Na di ones wey dey redesign dia processes to reward wetin AI genuinely no fit replicate: novel synthesis across unfamiliar domains, ethical judgement for ambiguous situations, cross-domain creativity, and di discipline to ask better questions instead of generating faster answers.
If your strategy, planning, or decision-making fit dey predicted with 90% accuracy from historical data, AI no be your biggest challenge.
Originally published on LinkedIn by Keshavan Thiagaraja. Read the original on LinkedIn.