What 210 Professors Actually Say About AI: A Language Analysis of Syllabus Policies

210 Professors Walk Into a Syllabus

There is a public spreadsheet floating around academia called the Syllabi Policies for Generative AI Repository. It contains 210 course-level AI policies from 181 institutions across 75 disciplines. Faculty members voluntarily submitted their own syllabus language, creating an unusually honest snapshot of how professors actually talk about AI when they sit down to write the rules.

Most analyses of these policies focus on stance: does the professor allow AI or ban it? We did something different. We read every single policy for language, tone, metaphor, and emotion. We tracked the specific words professors chose, the comparisons they reached for, and the anxieties (or excitement) bleeding through their sentences.

This is a complement to GradPilot's AI Policy Observatory, which tracks admissions-level AI policies at 174 universities. That project tells you what schools require of applicants. This analysis tells you what professors reveal about themselves.

The findings were far more interesting than we expected.

The Spectrum of Tone: From "Don't." to "No Restrictions"

The shortest policy in the dataset belongs to a Harvard course. It is one word long:

"Don't."

The second-shortest belongs to a course at UMass Boston. It is two words:

"No restrictions."

Between those two poles lies every shade of academic anxiety, optimism, pragmatism, and confusion you can imagine. But when you classify the tone of all 210 policies, a clear picture emerges: the average professor is neither panicking nor celebrating. They are trying to be reasonable.

Tone breakdown across 210 policies:

• Balanced/Pragmatic: 82%
• Supportive/Encouraging: 13%
• Cautious/Wary: 3%
• Hostile/Prohibitive: 1.4%

That last number is worth sitting with. Fewer than 2% of professors who chose to write and share an AI policy adopted a hostile tone. The narrative that faculty are uniformly terrified of AI does not hold up in their own words.

We also counted what we call "fear words" and "hope words" across the full corpus. Fear words include terms like integrity, plagiarism, violation, dishonest, cheat, bias, and penalty. Hope words include learning, assist, skill, develop, support, creative, and opportunity. The overall fear:hope ratio is 0.66:1. Hope words outnumber fear words by roughly 50%.

But that ratio is not evenly distributed. When you break it down by discipline, the differences are striking. Physical Sciences courses carry the highest fear ratio at 9.0x, meaning fear words outnumber hope words nine to one. Literature courses are close behind. At the other end, Research-focused courses show a fear ratio of just 0.1x, meaning hope words outnumber fear words ten to one. Computer Science is nearly as optimistic.

The professors who study technology are far more comfortable with it than the professors who don't. That is perhaps obvious. What is less obvious is just how dramatic the gap is.

View Harvard's admissions-level policy

The Metaphor Zoo: What Professors Compare AI To

Of the 210 policies, 167 (80%) refer to AI as a "tool." That is the default metaphor, and it is so pervasive that it barely registers as a metaphor anymore. But the professors who reach beyond "tool" for something more vivid tell us far more about how they actually understand this technology.

Here are the most memorable comparisons we found, quoted verbatim from the policies.

The E-Bike

A Georgia Tech biomedical engineering course (BMED2250) offers one of the most carefully constructed analogies in the dataset:

"Think of AI like an e-bike. If our goal is to only get somewhere faster, an e-bike might do the job. If our goal is to become a better cyclist, an e-bike can interfere with that happening. Long term, if you use an e-bike to do all your cycling you might end up in worse shape. Worse when the battery dies you will be stranded."

This is a professor who has genuinely thought about the pedagogical implications. The e-bike analogy captures something that "tool" does not: the distinction between using a technology to reach a destination and using it to build a capability. You can get to class faster on an e-bike, but you will not get faster yourself.

View Georgia Tech's admissions-level policy

Janet from The Good Place

A Wake Forest professor draws from television to make the point vivid:

"Imagine you lived in a world where everyone had immediate access to an anthropomorphized vessel of knowledge... Imagine, that is, we had a Janet. Yet this Janet had also just been rebooted, so she would sometimes give us a cactus when we asked for water."

The same professor extends the metaphor into a swimming lesson:

"If you wanted to learn to swim, you might ask Janet to explain it to you, to show you how to swim, and maybe even provide feedback on your stroke. But you wouldn't ask her to show up at the gym at 6 AM and swim your practice laps for you."

This captures the core tension better than almost any other policy we read. AI can explain, demonstrate, and evaluate. But it cannot do the learning for you, and it will occasionally hand you a cactus.

View Wake Forest's admissions-level policy

The Yelp Restaurant

Pierce College takes a consumer-awareness angle:

"AI answers are based on crowdsourced logic. If you've ever eaten at a 4.8-star Yelp restaurant and the food was bad, you know why this is problematic for accuracy."

This is a professor who understands that students already have intuitions about aggregated opinions being unreliable. The Yelp comparison makes hallucination tangible without using technical language.

The Pencil (With a Catch)

A version of this metaphor appears across multiple policies, including Pierce College and Clemson:

"AI is a tool, just like a pencil or a computer. However, unlike most tools you need to acknowledge using it."

The twist matters. A pencil does not generate content. A calculator does not write your proof. The "just a tool" frame breaks down precisely at the point where AI becomes interesting, and this professor notices.

The Ladder vs. the Crutch

Harvey Mudd College keeps it concise:

"When we choose to use it, we should be using it as a ladder and not a crutch."

Six words that distinguish between using AI to reach higher and using it to avoid standing on your own. The vertical vs. supportive metaphor is sharper than it first appears.

The New Wikipedia

Central Michigan University reaches for a comparison that older students and faculty will immediately understand:

"I think of it as the new Wikipedia -- a great place to start but you, as the author, are responsible for ensuring that the information and outputs are appropriate."

This is strategically useful because it normalizes AI use (everyone uses Wikipedia) while preserving accountability (but you still need to verify).

The Full Menagerie

Beyond these featured metaphors, we tracked every comparison across all 210 policies:

• Substitute/replacement: 20 policies
• Assistant: 19 policies
• Collaborator: 18 policies
• Tutor: 16 policies
• Spell-checker: 8 policies
• Calculator: 4 policies
• Weapon: 2 policies
• Pill: 1 policy (Lyon College)
• Slide rule: 1 policy

The gap between "collaborator" (18) and "weapon" (2) tells you everything about the overall professional disposition. Professors may be cautious, but they are not, as a group, combative.

Fear Words vs. Hope Words: The Emotional Register

The raw numbers across all 210 policies:

• Total fear-word occurrences: 1,228
• Total hope-word occurrences: 1,851

That is a meaningful gap. Professors are not writing these policies from a place of dread. They are, on average, cautiously optimistic. But the averages obscure real variation.

Top fear words by frequency:

• "integrity" -- 170 occurrences
• "plagiarism" -- 117
• "bias" -- 82
• "violation" -- 79
• "dishonest" -- 68

Top hope words by frequency:

• "learning" -- 225 occurrences
• "assist" -- 140
• "skill" -- 114
• "develop" -- 102
• "support" -- 81

The most common bigram in the entire dataset is "academic integrity," appearing 141 times. This phrase has become the gravitational center of AI policy language. Whether a professor permits AI or restricts it, they almost always frame the conversation through the lens of academic integrity. It has become the shared vocabulary, the common ground that both sides of the debate stand on.

The extremes are instructive.

The most fear-heavy policy belongs to Bethune-Cookman University, with a ratio of 21:0. Twenty-one fear words. Zero hope words. The entire document reads as a warning.

Michigan State's SPN320 course is close behind at 13:0. Another policy composed entirely in the language of prohibition and consequence.

At the other end, Fairfield University's BI1151 course policy contains 18 hope words and zero fear words. It reads like an invitation. Northwestern Law hits 0:12 in the same direction.

These extremes are rare. The vast majority cluster around the balanced middle. But they reveal the full emotional range available to a professor sitting down to write about AI, and the choices they make in that moment say more than any survey.

View more restrictive policies

The 15 Policies That Were Written By AI

Here is an irony that no one in the dataset seems embarrassed about: 15 of the 210 policies (7%) were written, at least in part, by AI. The professors who govern student AI use are themselves AI users.

Some are transparent about it. Some are remarkably casual. Some contain revealing typos.

Santa Barbara City College's entire disclosure is a single sentence:

"This was the policy created by CHATGPT"

That is the whole disclosure. No elaboration. No justification. Just a professor saying: I asked ChatGPT to write my AI policy, and here it is.

The University of Missouri preserved a notable typo:

"This policy was edited for clarity using ChatGPTo"

"ChatGPTo." Not ChatGPT-4o. Not ChatGPT. ChatGPTo. The typo survived the editing process, which raises interesting questions about how carefully the AI-assisted editing was reviewed.

Western New England University went all in:

"Several GenAI models were used in the creation of this document, including but not limed to ChatGPT, Claude, Co-pilot, and Gemini."

"Not limed to." Another typo that made it through. But the substantive point is more interesting: this professor used four different AI models to write a policy about AI use. That is a level of engagement that suggests genuine curiosity, not reluctant compliance.

Clemson offers possibly the most honest attribution statement in any academic document:

"I also likely saw something on Twitter that prompted me writing this. I can't remember."

That is a professor acknowledging the actual messiness of intellectual influence, something most policy documents would never admit.

And then there is CSU Fullerton's Criminal Justice professor, who turned the whole thing into a game:

"Can you spot the sentences it wrote?"

Seven percent of AI policies were drafted with AI assistance. The significance is not that professors are hypocrites. It is that AI has already been normalized as a writing tool in the very community that is still debating whether students should use it as a writing tool. The policy-writers have answered the question with their behavior, even as their policies hedge.

What Professors Actually Require: The Citation Specificity Ladder

When professors do permit AI use, they diverge wildly on what they require students to document. We identified seven tiers of citation specificity, forming a ladder from "say nothing" to "prove everything."

More than a third of professors who wrote AI policies did not require any citation at all. Among the 63% who do require acknowledgment, most stop at "tell me what tool you used." The deep end of the pool, where students must produce transcripts and screenshots, is populated by only a handful of courses.

The most demanding policy belongs to Western New England University, which requires students to submit the full chat dialogue (highlighted), attach it as an appendix, write a reflection on the AI's contribution, and, if they did not use AI, submit a non-use declaration. Proving you did not do something is a distinctly academic requirement.

The most creative requirement goes to Kennesaw State University, which asks students to include carbon, water, and electricity impact statements for their AI usage. This is the only policy in the dataset that frames AI use as an environmental question rather than an integrity question.

The ladder matters for students because it reveals how much overhead AI use actually creates. At the top, using AI arguably takes more work than not using it. Some professors may have designed it that way on purpose.

Explore the disclosure landscape | How to write an AI disclosure

By Discipline: Who's Scared and Who's Excited

The 210 policies span 75 disciplines, but clustering them into broader categories reveals a pattern that is both predictable and ironic.

Most fear-heavy disciplines:

• Physical Sciences (9.0x fear:hope ratio)
• Literature
• Foreign Languages

Most hope-heavy disciplines:

• Research (0.1x fear:hope ratio)
• Computer Science
• Education

The perfectly balanced discipline: Writing courses, the largest single group in the dataset at 41 policies, land at an almost perfect 0.9:1 fear:hope ratio. The people who teach writing for a living are neither panicking about AI nor celebrating it. They are, characteristically, choosing their words carefully.

Engineering presents an interesting split. Some engineering courses treat AI as a professional tool that students need to learn (the "you'll use this at Boeing" framing). Others treat it as a learning obstacle (the "you need to derive this yourself" framing). The same discipline, arguing with itself.

The deepest irony belongs to the gap between Computer Science and Literature. CS professors, who understand the technology most intimately, are among the most permissive. Literature professors, whose discipline depends on close reading and original interpretation, are among the most restrictive. The people who build the tools trust them more than the people who analyze texts for a living.

This maps, roughly, onto a divide between disciplines that value product and disciplines that value process. If the goal of your course is to produce a working program, AI is an accelerant. If the goal is to develop a student's capacity for original thought, AI is a short circuit. The policies reflect that fundamental pedagogical difference more than any ideological position on technology.

View law school AI policies | View business school AI policies

Methodology and Data Source

Data source: Syllabi Policies for Generative AI Repository, a publicly maintained spreadsheet of 210 voluntarily submitted course-level AI policies from 181 institutions across 75 disciplines.

Analysis method: GradPilot performed language analysis across all 210 policy texts, tracking word frequency, bigrams, metaphor usage, tone classification, and fear/hope word ratios. Tone classification was based on a composite assessment of word choice, framing, and stated disposition toward student AI use.

Limitations: This is a self-selected sample. Professors who submit their policies to a public repository may skew toward those who have thought more carefully about AI, which likely biases the sample toward more balanced and permissive policies. This dataset should not be read as representative of all higher education AI policies. It is a window into what engaged, thoughtful faculty are writing, which is valuable on its own terms but not generalizable.

View our full methodology