Assessment Has Always Been the Bottleneck
In Benjamin Bloom's original 1984 paper on mastery learning, he demonstrated that students who received one-on-one tutoring outperformed classroom peers by two standard deviations, what he called "the 2 sigma problem." The tutoring advantage came primarily from two things: immediate feedback and instruction calibrated to the individual learner.
For 40 years, that insight sat largely unactioned because the logistics were impossible. You cannot give 30 students individual tutors. What AI is beginning to offer is not a tutor, but something structurally similar: the ability to provide timely, specific feedback at a scale that was previously impossible in a classroom setting.
Formative vs Summative Assessment: Where AI Fits
Formative assessment happens during learning to inform instruction. Exit tickets, quizzes, think-pair-share, thumbs up/thumbs down checks, these are formative. Their value depends on speed: feedback that arrives three days after the lesson cannot adjust instruction for that lesson.
Summative assessment happens at the end of a learning period to measure achievement. Exams, final projects, standardized tests, these are summative. Their value depends on accuracy and comparability.
AI tools are making the most significant impact in formative assessment, where the speed advantage is most consequential. Here is how.
Exit Tickets at Scale
Exit tickets, brief end-of-class checks for understanding, are one of the most evidence-backed formative assessment strategies. A 2019 meta-analysis in the Journal of Educational Psychology found that regular brief assessments with immediate feedback improved learning outcomes by an average of 0.4 standard deviations across subject areas.
The obstacle has always been analysis speed. A teacher with 90 students across three class sections receives 90 exit tickets per lesson. Reading, categorizing, and planning instructional responses traditionally takes 30–45 minutes per class per day. With AI-assisted tools like Kahoot!'s AI analysis features or Formative, that analysis now happens automatically: the teacher receives a summary of misconception patterns within seconds of class ending.
Automated Short-Answer Grading
Multiple choice grading was automated decades ago. Short-answer and constructed-response grading, where students write 2–5 sentences explaining their thinking, has remained stubbornly manual because it requires understanding meaning, not just pattern matching.
Natural language processing has changed this. Current AI systems can:
- Score short answers against a rubric with inter-rater reliability comparable to trained human raters for many question types
- Identify the specific conceptual misunderstanding behind an incorrect answer (not just "wrong" but "this student confused X with Y")
- Generate individualized feedback comments based on the student's specific errors
This does not mean AI grading is perfect. It performs best on constrained prompts with clear rubrics and worst on open-ended analytical writing. Teachers using AI grading appropriately treat it as a first pass that they review, not a final decision.
AI and Bias in Assessment
One often-overlooked benefit of AI-assisted grading is bias reduction. Human graders are susceptible to well-documented biases: halo effects (grading work from high-performing students more generously), stereotype threat effects, and fatigue effects (earlier papers graded more carefully than later ones).
A 2022 study published in Computers and Education found that AI grading models showed significantly lower demographic bias in essay assessment than human raters, particularly along race and gender dimensions, provided the training data was carefully curated. This is not an argument for fully automated grading, but it is an argument for using AI as a check on human consistency.
Bloom's Taxonomy and AI Question Generation
Bloom's taxonomy categorizes cognitive tasks from lower-order (remember, understand) to higher-order (analyze, evaluate, create). A persistent challenge in assessment design is that most teacher-generated test questions cluster at the lower levels, recall and comprehension, because higher-order questions take longer to write and harder to grade reliably.
AI question generators now allow teachers to specify not just topic but cognitive level. A teacher can request: "Generate five questions about the American Civil War at the evaluate level of Bloom's taxonomy." This democratizes sophisticated assessment design that previously required significant training or time investment.
Addressing the Concerns
"AI grading will make teachers lazy." The evidence does not support this concern. Teachers who use AI assessment tools consistently report using the time saved to have more individual conversations with students, precisely the high-impact interaction that Bloom identified as the mechanism behind tutoring's advantage.
"Students will game AI feedback." Yes, students will try. But this is not new: students have always tried to reverse-engineer grading rubrics. The solution is assessment design that requires demonstration of genuine understanding, projects, oral defenses, and varied assessment formats, rather than abandoning AI tools.
"AI cannot assess creativity or deep thinking." This is currently true for summative assessment of complex work. AI is most reliable for formative assessment of constrained tasks. Using AI to grade final exams in philosophy is inappropriate. Using AI to analyze whether students can correctly identify the main argument in a paragraph is appropriate and useful.
The Near-Term Reality
The most realistic near-term scenario is not AI replacing teacher judgment in assessment, it is AI handling the mechanical, time-consuming parts of the assessment cycle so that teacher judgment can be applied more selectively and at higher cognitive levels.
The teacher's role shifts from reading 90 exit tickets to interpreting the AI's analysis of those 90 tickets and deciding what instructional adjustment is warranted. The teacher's domain knowledge, relationships with students, and pedagogical expertise remain essential. The AI handles volume; the teacher handles meaning.
For institutions considering how to integrate AI assessment tools, the key question is not "which AI tool is most accurate?" but rather "how do we build assessment processes where AI and teacher judgment complement each other's strengths?"