Evaluating AI Results: A Practical Guide for Instructors

Cristina Mitrea is a Lecturer IV in computational medicine and bioinformatics at the University of Michigan Medical School.

Generative AI tools like ChatGPT, Claude, Google Gemini, and Copilot have become part of our daily workflow for many instructors. We use them to draft syllabi, generate practice problems, outline lectures, and write code for student exercises. These tools can be very useful, but they can also produce output that looks polished and authoritative while being subtly wrong.

As instructors consider whether and how to use AI in their teaching, it is critical they consider how to use it responsibly, and how to prepare students to do the same.

This article is adapted from an Enriching Scholarship 2026 workshop I gave with my colleague Anthony J. Wing. The goal is to list some strategies you can apply right away, regardless of your discipline or the specific tool you use.

AI Outputs Are Drafts, Never Finished Products

As instructors, we have the responsibility to be prepared to critically and effectively evaluate the output of AI tools if we are to use that output in the teaching materials we share with our students. Treat everything an AI tool generates as a first draft that requires your review. All of it needs verification before it reaches students: text, code, data, citations, answer keys.

Large language models are based on calculating the most probable next word. They do not look up facts in a database, and they do not know when they are wrong. The result is output that reads fluently but may contain fabricated references, outdated information, or plausible-sounding claims that are factually incorrect. If you put your name on it, you are accountable. Even if you do not put your name on it, you are responsible for what you do with that output.

This is where the instructor’s own expertise becomes essential. To evaluate AI output effectively, you need to be a domain expert to the best of your abilities. AI can help you learn more, go deeper into your subject, and create better teaching materials with more relevant and varied examples. But the danger is cognitive offload: relying on AI to do the instruction rather than using it as a tool to strengthen your own. If you do not understand the material well enough to catch errors, the AI’s mistakes become your mistakes.

Where AI Helps Instructors

AI tools are strongest when you already know the material and need help with volume, structure, or variation. They work well for:

• Drafting and outlining. First-draft lecture outlines, lab guides, or reading lists that you then refine and verify.
• Creating problem variations. Multiple versions of practice exercises from a template you provide.
• Formatting and restructuring. Converting rough notes or bullet points into structured handouts or study guides.
• Generating synthetic datasets. Sample data for exercises when real data is restricted or unavailable.
• Writing boilerplate code. Skeleton code for student exercises that you review and test before distributing.

In each case, you are the subject-matter expert. AI accelerates the mechanical work while you remain responsible for accuracy and pedagogical intent.

To give a sense of what AI can do, I used Copilot in Visual Studio Code and created a working Python package for analyzing public COVID-19 data with just a couple of instructions in a README file (Devakumar, 2020; JHU CSSE COVID-19 Data; Jauhari, 2020). The AI wrote the code, created unit tests, set up the project structure, and produced a bar chart of recovery rates by country. The whole package worked. But verifying all of that code takes longer than generating it. To reduce the verification time, ask the AI to produce well-documented and modular code that can be easily understood and navigated. That way, you can check each piece independently rather than reading through one long script.

Where AI Does Not Do So Well

AI tools struggle in areas that matter most for teaching quality:

• Answer keys. AI can mark the wrong answer as correct, especially in multi-step reasoning, problems that require distinguishing between similar concepts, and numerical calculations. Always verify answer keys independently.
• Nuanced explanations. AI may oversimplify, present outdated frameworks, or omit important caveats that students need to understand.
• Citations and references. AI frequently generates well-formatted references with fabricated DOIs, nonexistent authors, or invented journal articles. Every citation must be checked.
• Tool-specific instructions. AI may give instructions for the wrong software version or confuse similar tools. This is especially problematic in technical courses.

The VERIFY Framework

When you receive AI-generated content, run it through the VERIFY checklist before using it in your course:

• V – Validate sources. Check every reference and link the AI provides. Confirm the authors, year, and journal match a real publication.
• E – Examine specifics. Verify concrete claims: names, dates, numbers, formulas, definitions. These are where errors hide.
• R – Recognize hedging. Watch for phrases like “it is generally believed that…” or “many researchers argue…” AI uses this language to sound cautious while still making claims that may be wrong.
• I – Identify recency issues. For any time-sensitive topic, verify with a current source. AI training data has a cutoff, and the tool will not tell you its information is outdated.
• F – Find contradictions. Ask the AI the same question in a different way or in a new conversation. If the answers conflict, neither should be trusted without independent verification.
• Y – Yield to authoritative sources. When in doubt, go back to the textbook, peer-reviewed literature, official documentation, and domain experts. AI is a drafting tool, not an authority.

This framework applies whether you are reviewing AI-generated text, code, or data. For code, also check library versions and imports, look for hardcoded assumptions, and verify the logic, not just whether the syntax runs without errors.

Here is a quick example. I asked GPT-5.5 for references to five articles from “Nature Genomics,” a journal that does not exist. The AI correctly guessed I meant Nature Genetics and returned five nicely formatted citations with titles, authors, page numbers, and DOIs. Everything looked right. But when I clicked the first link, it led to a completely different paper. The Validate and Examine steps of VERIFY would catch this in seconds, but only if you do the check. Citations are a big challenge for AI tools (Topaz, 2026).

Why “Show Your Work” Is Not Enough

You might expect that AI models with visible reasoning steps would be easier to trust. Researchers tested this by evaluating Chain-of-Thought faithfulness in models like Claude 3.7 Sonnet and DeepSeek R1 (Chen et al., 2025). When models were given hidden hints suggesting a specific answer, they often changed their answers to match the hint but failed to mention the hint in their reasoning. Faithfulness scores dropped on harder tasks, exactly where we rely most on this transparency.

Do not assume that because an AI shows its work, the work is trustworthy. Press the model for the connections it made. Better yet, make the connections yourself.

Designing AI-Resilient Assignments

Rather than trying to ban AI or rely on detection software, consider redesigning assignments so that AI cannot easily complete them. Five strategies that work across disciplines:

1. Require process documentation. Ask students to submit drafts, revision histories, or annotated decision logs. AI cannot fake a genuine process.
2. Use unique or live data. Base assignments on datasets, case studies, or scenarios specific to your course and not available online. AI cannot analyze what it has never seen.
3. Include oral components. Even a brief follow-up conversation about a submission reveals whether a student understands what they turned in.
4. Assign multi-stage submissions. Break large projects into checkpoints, where each stage builds on feedback from the previous one.
5. Embed classroom-specific context. Reference in-class discussions, guest lectures, or peer feedback in the assignment prompt. If a student prompts that in the AI, it will not know what happened last Tuesday in your class.

Setting AI Policies

Be explicit from day one. Students need to know the rules before the first assignment, not after a problem occurs. Some students will read no policy as “everything goes.” Others will assume nothing is allowed. Consider language along these lines:

“You may use AI tools (ChatGPT, Claude, Copilot, etc.) for brainstorming and exploring concepts. You may not submit AI-generated text, code, or analysis as your own work. If you use AI at any stage of an assignment, disclose the tool, the prompts you used, and how you modified the output. Unacknowledged use of AI will be treated as an academic integrity violation.”

You can start small and add more specifics in subsequent semesters as your experience grows.

Recognizing AI-Generated Work

Do not rely on AI detection software. These tools have significant false positive rates, and there have been notorious cases when people were flagged for using AI when they had not (Sun, 2020, Weber-Wulff, 2023). Instead, look for the following patterns:

• Uniform tone with no variation in voice, 
• Correct terminology without the small errors that students typically do, 
• Generic examples that do not reflect your course, 
• A sudden quality shift in a particular student’s work.

The goal is to evaluate whether students know the material, not to police them. Reserve the right to ask questions about submitted work. If a student cannot explain what they turned in, that tells you what you need to know.

Conclusions

AI tools are changing fast, and the specific capabilities and limitations described here will evolve. What will not change is the need for instructors to model critical evaluation. When we verify AI output before distributing it, when we show students how we caught an error, and when we are transparent about our own AI use, we teach something more durable than any single tool; we teach the habit of questioning what looks correct and confirming it before passing it on.

With the power of AI comes a bigger responsibility for us (Yee et al., 2025). We can be better teachers because we have tools that accelerate creating exercises, quizzes, and well-organized course materials. The work now is making sure we verify what these tools produce and teaching students to do the same. AI should make us better at what we do, not replace what we do.

Resources

U-M ITS AI Resources.

Digital Education Council (2025). Global AI Faculty Survey. 

Swiss Institute of Artificial Intelligence (2026). Cognitive Outsourcing in Education: Why AI’s Real Classroom Crisis Is Verification, Not Cheating.  

References

Chen, W., et al. (2025). Reasoning Models Don’t Always Say What They Think. Anthropic Research.

Yee, K., et al. (2025) Coach for the Approach: The Educator’s New Role in the Age of AI. FCTL Press. 

Devakumar, P. R. (2020). Covid 19 JHU Data Web Scrap and Cleaning (Version master) [Data set]. GitHub. 

JHU CSSE COVID-19 Data.

Jauhari, N. (2020). COVID-19 All Countries Data (Version 1) [Data set]. Kaggle. 

Hadra, M., et al. (2026). Evaluating the accuracy and reliability of AI content detectors in academic contexts. International Journal for Educational Integrity, 22(1), 4. 

Sun, Y., et al. (2026). Trusting AI to detect AI? A systematic evaluation of the reliability and robustness of current AIGC detection tools for student academic work. Computers & Education, Volume 249, Issue C.

Weber-Wulff, D., et al. (2023). Testing of detection tools for AI-generated text. International Journal for Educational Integrity, 19(1), 26. 

Topaz, M., et al.. (2026). Fabricated citations: an audit across 2· 5 million biomedical papers. The Lancet, 407(10541), 1779-1781.