The Texas threat vs. the Oregon State suggestion

Imagine a single line in a syllabus: “Students will learn to dismantle stereotypes about who belongs in computer science.”

In Texas, that sentence is a liability. In late 2025, Texas State administrators handed faculty an AI prompt designed to hunt for “ideological” wording, targeting words like dismantling and decolonizing. The system flagged 280 courses, forcing faculty to rewrite them by a strict deadline or risk cancellation. Administrators called it “consistent, evidence-based criteria”. Critics called it a project to control education. Tellingly, the search results shifted depending on how the prompt was phrased, exposing the supposed consistency as a facade.

Feed that exact sentence to a different scanner. Researchers at Oregon State built an “inclusivity detector” that reviews course materials for subtly alienating phrasing. Tested on a massive online computer science course, the goal was assistance, not enforcement. The AI flagged potential issues to help the teaching team; nobody was forced to act.

Both tools read the same text. The aftermath differs: one flag arrives with a deadline and a threat, the other as a take-it-or-leave-it suggestion. The act of reading can in both, on itself, be its own ethical problem depending on who gets to flip the switch and who gets to act on its outcomes.

Scale removes the safety valve of human limits

Before AI, human bandwidth acted as a natural safety valve: you couldn’t be punished for a sentence nobody had the time to scrutinize.

Scale removes that valve. Now, everything can be watched constantly, and no one has to explain why a specific course was singled out.

Scanning everything does have an upside. A single automated pass can catch a genuine accessibility gap an overworked reviewer might miss for years. Treating every course identically is, in one sense, fairer. But sameness bought this way relies on a vendor’s standardized wordlist, not on field experts agreeing on what their own language means. It’s a thinner, fragile kind of fairness.

The algorithm didn’t seize power. The people behind it just stopped having to explain themselves.

A traditional syllabus review involved a conversation: someone asked why you wrote something, and trusted your answer unless given a reason not to. An AI scan replaces that human question with an automated flag. It’s wildly efficient, but the act of asking is what built trust between teachers and administrators. Removing the friction of conversation erodes that trust.

The rules hold the power, not the model

Two recent pieces put this into perspective. A New Yorker column warns that AI could flatten the unique quirks of different educational styles into a bland, inoffensive average. A De Correspondent column argues that an AI model’s power isn’t a property of the code itself; it stems entirely from the rules, habits, and bureaucracies built around it.

If true, the Texas tool’s chilling effect never came from its technical capabilities. It came from a power structure where nobody was positioned to demand an explanation.

UNESCO recently published a discussion paper proposing a different path: AI used in education should be required to prove its benefit before it gets deployed, not just after someone complains. This concept (deliberative governance) is the missing piece in both the punitive Texas model and the helpful Oregon State model.

The urgent question in education isn’t whether AI should read everything.

It’s who gets to ask it why, and what happens if they don’t like the answer.

References and further reading

Bryant, P. (2026, February 12). The Mirror University 7: Trust is the only currency that matters in higher education. Post Digital Learning. https://peterbryant.org/the-mirror-university-7-trust-is-the-only-currency-that-matters-in-higher-education-rebuilding-a-culture-of-trust-in-an-era-of-distrust/

Chatterjee, A., Choudhuri, R., Sarkar, M., Chattopadhyay, S., Liu, D., Hedaoo, S., Burnett, M., & Sarma, A. (2024). Debugging for inclusivity in online CS courseware: Does it work? In Proceedings of the 2024 ACM Conference on International Computing Education Research (Vol. 1, pp. 419-433). Association for Computing Machinery. https://doi.org/10.1145/3632620.3671117

Kang, J. C. (2026, October 17). Eight predictions for the future of higher education. The New Yorker. https://www.newyorker.com/news/fault-lines/eight-predictions-for-the-future-of-higher-education

Martijn, M. (2026, October 15). Een AI-model kan nog zo krachtig zijn, macht krijgt het pas als wij die geven [An AI model can be as powerful as it likes, it only gets power when we give it]. De Correspondent. https://decorrespondent.nl/17322/een-ai-model-kan-nog-zo-krachtig-zijn-macht-krijgt-het-pas-als-wij-die-geven/dd440c39-b86a-0bd3-3632-468120fcd708

The Texas Tribune. (2025, December 15). Texas universities deploy AI tools to review and rewrite how some courses discuss race and gender. https://www.texastribune.org/2025/12/15/texas-universities-ai-course-audits/

UNESCO. (2026). Sustaining education as a common good in the age of AI: The case for deliberative governance [Discussion paper]. https://www.unesco.org/en/articles/education-ministers-call-education-remain-common-good-age-ai-unescos-digital-learning-week?hub=722