Rethinking Homework for the AI Age: Andrej Karpathy’s Playbook for Schools

Recently, a viral demonstration showcased an AI tool solving a handwritten math problem and presenting the answer in the student’s handwriting. If handwriting once served as the proof-of-work for assignments, it now resembles more of a stylistic filter. This clip encapsulates a challenging reality for educators: out-of-class work and AI assistance are becoming increasingly intertwined.

Andrej Karpathy, former leader at OpenAI and Tesla, offers a straightforward recommendation: eliminate the enforcement of AI-generated homework and restructure assessments to focus on what can only be evaluated within a supervised classroom setting. In a widely circulated online discussion, he declared AI detectors “doomed to fail,” advocated for shifting most grading back to in-class environments, and emphasized the need for students to be adept with and without AI.

This article translates Karpathy’s insights into a practical, research-informed guide for teachers, administrators, and interested readers. We will explore his advice, discuss the challenges of AI detection, and outline strategies for reimagining courses, assignments, and grading to ensure effective learning in the AI era.

Key Insights from Karpathy

Here are Karpathy’s main points, paraphrased and organized for clarity:

• Unreliable Detection: AI usage in take-home assignments will never be reliably detected. Treat AI detectors as fundamentally flawed and easily bypassed.
• In-Class Evaluation: Shift most graded assessments to in-class settings where teachers can supervise students directly.
• Dual Proficiency: Ensure that students are both proficient with AI tools and capable of independent work. Just as we teach arithmetic to make students capable of verifying their work, we must do the same for AI proficiency.
• Varied Evaluation Formats: Create a range of evaluation styles based on teacher discretion—ranging from no tools, to limited notes, to open book assessments, and AI critiques, depending on the educational objectives.

These suggestions stem from a public discussion and school board briefing that Karpathy mentioned, aligning with his overarching caution that while AI systems are powerful, they are also fallible and require human oversight.

The Challenges of AI Homework Detection

Even prior to the recent advancements in multimodal AI, text-only detection tools faced significant obstacles. OpenAI discontinued its AI text classifier in July 2023 due to its unreliable accuracy. Reports and independent tests uncovered high rates of false positives, particularly in shorter texts or those authored by non-native English speakers.

Consider the following changes:

• Multimodal Capabilities: Newer models can read a photo of handwritten work, solve the problem, and produce an answer imitating the student’s handwriting, making surface-level authenticity checks nearly irrelevant for many assignments.
• Image-Generation Advances: Tools like Google’s Nano Banana Pro, based on Gemini 3 Pro Image, increasingly facilitate better text rendering and layout control, enabling seamless integration of AI outputs into typical homework formats.
• Vulnerability to Manipulation: Existing detectors remain unreliable and can be gamed through paraphrasing or prompt adjustments. As a result, universities have scaled back on using them after instances of falsely accusing students, highlighting the ethical and legal risks associated with these weak signals.

In essence, the rapid improvements in AI models greatly outpace the reliability of detection tools, and this imbalance is unlikely to change.

A Deeper Educational Insight: Teaching with and without AI

Karpathy’s analogy to calculators underscores a fundamental principle: calculators are commonplace, yet understanding math remains crucial. We teach mental models so that learners can estimate, verify, and reason about numbers before leveraging calculators for efficiency. Similarly, students should learn to reason and validate their work before utilizing AI as an enhancement tool.

Practically, this translates to structuring course experiences that develop three complementary abilities:

1. Core Understanding Without AI
2. Conduct short, supervised quizzes and oral checks to assess conceptual understanding.
3. Use whiteboards for problem-solving and think-alouds.

4. Implement closed-tool exams focused on reasoning processes rather than just final answers.

5. Proficient Use of AI as a Tool

6. Create prompts that require students to plan their work with AI, critique AI outputs, and refine their results.
7. Assign tasks where students must set constraints for AI and verify its output against these constraints.

8. Require reflection memos detailing the student’s process versus the AI’s contributions and validation steps.

9. Meta-Skills for Verification and Judgment

10. Run error-spotting drills where students must identify flaws in AI-generated answers.
11. Engage students in calibration exercises where they gauge their confidence and then compare it to actual outcomes.
12. Facilitate source-checking activities, prompting students to track claims back to credible sources or primary data.

Reimagining Assessment in an AI-Driven Landscape

Here’s a variety of strategies for adjusting assessment practices. Select combinations that align with your objectives, grade level, and constraints:

Focus High-Stakes Evidence in the Classroom

• Use in-class exams with problems that are different from those seen outside of class.
• Implement timed whiteboard or paper tasks in which students narrate their problem-solving process.
• Conduct live coding walkthroughs or practical labs.
• Organize mini-vivas, short oral defenses of a solution or paragraph.

These formats provide insight into students’ thinking processes, rather than merely the final products they submit.

Utilize AI Constructively Outside Class

• Allow AI for preliminary work: Students can use AI for brainstorming or outlining solutions and must document the prompts and versions used. In class, they can refine and defend their outcomes.
• Assign AI critique tasks: Provide an AI-generated draft or solution and ask students to review, modify, and justify their changes.
• Implement flipped labs: Students prepare with AI-produced study guides at home and tackle new problems in class that require application of what they’ve learned.

Make Verification Part of the Process

• Require a brief verification plan with every AI-assisted submission detailing what was checked, how it was validated, and what considerations were made.
• Implement random spot checks: choose a small sample of submissions weekly for brief oral explanations or replicating one problem in class.
• Use rubrics that include a verification criterion to reward students for showing how they validated content, not just for achieving a correct answer.

Define Allowed vs. Disallowed AI Usage

• Set clear policy categories for assignments: no tools, notes only, open internet, provided-AI only (students can critique but not generate), or controlled access to AI.
• Maintain transparency logs: students should document their final prompts and reflections on any AI-assisted work.
• Ensure equitable access: establish policies that don’t penalize students with limited device access by providing managed sessions at school.

A Practical Roadmap for Educators

For those redesigning a course mid-year, consider a phased approach:

1. Immediate Triage
2. Transition from take-home essays to in-class assessments and oral explanations.
3. Substitute some essay prompts with tasks that involve analyzing AI outputs.

4. Introduce a verification criterion in upcoming assignments.

5. Next Term Redesign

6. Align learning outcomes with specific assessment modes: no-tool, limited-tool, or open-tool.
7. Develop a library of in-class variants to ensure take-home materials are for practice, not evaluation.

8. Create a standardized transparency appendix template for students to utilize.

9. Institutionalization for Next Year

10. Align department policies on permissible AI usage and documentation requirements.
11. Provide shared monitoring toolkits for low-tech assessments: seating charts, multiple versions, and whiteboard work sessions.
12. Conduct workshops for faculty focused on prompt design, verification strategies, and inclusive assessment.

Considerations of Fairness, Bias, and Trust

Relying on AI detectors presents tangible risks. Documented incidents show false positives disproportionately impacting certain student demographics, resulting in stressful investigations and potential delays in graduation. Even reliable vendors warn that low-accuracy flags should only inform discussions, not definitive conclusions.

Karpathy’s recommendation to embrace AI use in homework and transfer grading into the classroom circumvents the detection issue entirely. It shifts the institutional focus to creating conditions where students are motivated to learn without relying on AI, understanding that high-stakes evaluations will be observed, while also fostering fluency in AI for real-world applications.

The Technological Context: Why Now?

Two significant shifts converged in late 2025:

• Capable Multimodal AI: Technological advancements have enhanced AI’s ability to read and generate images, with systems emphasizing legibility and creative control. Consequently, an AI can now analyze a photo of handwritten notes and recreate polished pages in the student’s handwriting.
• Cultural Normalization: Viral instances have established an expectation that many take-home tasks can be convincingly AI-generated, impacting perceptions of academic integrity.

Simultaneously, the reliability of detection tools has faltered. The outcome is a compelling push towards redesigning assessments rather than escalating a futile detection competition.

Embracing AI-Positive Innovations in Education

Karpathy is not against incorporating AI in educational settings. He actively invests in AI education and founded Eureka Labs to assist learners in engaging effectively with modern language models. His stance is that we should harness AI as a learning aid while simultaneously upholding the integrity of alternative evaluation methods.

Here are proactive approaches that align with this philosophy:

• AI as a Mentor: Employ AI to simulate a Socratic tutor that asks probing questions, identifies knowledge gaps, and recommends practice exercises.
• AI for Practice: Allow students to rehearse oral defenses with AI-generated questions before formal assessments.
• AI for Feedback: Have AI suggest improvements to organization, clarity, or code style, then require students to analyze, adjust, or reject these recommendations with justifications.
• AI for Accessibility: Utilize AI-generated summaries, alternative explanations, or adjustments in reading levels while evaluating comprehension in class.

Adapting Assignments for the AI Era

Here are some adapted examples for common homework types:

• Traditional Essay to In-Class Argument: Students draft their arguments at home using any resources, including AI. In class, they deliver a 3-minute argument followed by a 2-minute cross-exam, assessed on clarity, evidence, and reasoning.
• Problem Set to Whiteboard Studio: Groups tackle a new variant of a take-home problem on the whiteboard while narrating their steps. Each group member must take turns at the board.
• Research Summary to Source-Tracking: Provide a brief AI-generated summary. Students must trace each claim to credible sources and highlight any discrepancies.
• Coding Task to Live Code Review: Students present AI-assisted code and need to explain their design choices and address a new issue in a monitored setting.

Addressing Common Concerns

• In-Class Assessment and Anxiety: Offer multiple smaller checkpoints, permit brief prep notes, and maintain predictable structures to alleviate pressure.
• Managing Large Classes: Rotate oral check-ins, conduct whiteboard studios with assisting TAs, or implement brief quizzes using random variations.
• Perceived Extra Workload: While it may require more effort initially, creating reusable variant banks and structured rubrics, as well as utilizing AI for prompt drafting, can help manage workload over time.

Ethical Considerations and Transparency

Even when incorporating AI, it’s essential to clarify authorship. Require students to disclose AI usage details, including prompts and verification steps. This transparency fosters trust and provides insight into their processes, while also instilling a professional habit that will serve students in AI-influenced workplaces.

In Conclusion

Karpathy’s message emphasizes the necessity to pivot away from unobservable homework assessments and toward structured evaluations that showcase human reasoning. Make evaluated work visible in the classroom, define AI use explicitly, emphasize verification, and reward judgment. This approach can help students harness powerful tools like AI without becoming wholly dependent on them.

---

FAQs

1) Are AI detectors acceptable to use?

They can serve as one potential indicator, but their inaccuracies and biases imply they should not be the sole basis for academic decisions. When utilized, they should be combined with human review and other factors, treating low-confidence flags as unreliable.

2) How should I convey new AI policies to students?

Utilize straightforward language, provide examples of permitted and prohibited uses, and require a brief transparency appendix detailing prompts used, tools accessed, and validation steps. Make it clear that most grading occurs in monitored environments.

3) How can I manage long projects within my course framework?

Decompose projects into sequential deliverables that include in-class assessments: proposal pitches, mid-project reviews, code walkthroughs, and final defenses. Reserve take-home assignments for exploration and practice, while evaluation is designated for supervised interactions.

4) How can I teach students to verify AI outputs?

Engage students in error-spotting drills concerning AI-generated content, run confidence calibration exercises, and assign source-tracing tasks. Mandate a verification plan for every submission enhanced by AI.

5) Where can I explore more of Karpathy’s perspectives on AI and education?

Refer to his recent online thread summarizing advice to a school board, and check out his ongoing initiatives to develop AI educational resources and curricula through Eureka Labs.