AI & Open
Educational Resources

It invents things

Citations and quotes that never existed, written as confidently as the true ones.

It carries its patterns

Including the skews in the text it learned from, which show up unevenly in what it produces.

The concern

It reflects and can amplify the skews in its training data, and it hurts some groups more than others. Identical writing gets judged differently by dialect, and speakers of African American English are steered toward worse outcomes.

What we’d add

A model’s bias can be measured and tested before anyone relies on it, so a department can check a tool and turn it down. The available fixes still go only partway. No one runs the same checks on the human judgment it would replace, which is biased too.

The concern

The systems are trained and moderated by workers, many of them in the global south, paid very little to label data and screen disturbing content. They work under constant monitoring and suffer lasting mental-health harm from what they screen.

What we’d add

The accounts here are accurate. Companies rarely name who labels and moderates their systems, or under what terms, which leaves the conditions hard to check or improve.

The concern

Employers already use AI to argue for doing more with fewer people. In universities, administrators use it to push heavier workloads and more contingent jobs.

What we’d add

Early-career workers in the most AI-exposed occupations have already seen about a 16 percent relative drop in employment, concentrated in the jobs AI can do on its own. How far it spreads depends on the staffing and budget calls that managers make.

The concern

A small number of firms own the chips and the cloud that AI runs on, and they set the terms and prices. One company makes about 90 percent of the AI chips, and three run nearly two-thirds of the cloud.

What we’d add

Open-weight models have nearly caught up to the closed ones, so anyone can download and run a capable model. The hardware underneath is far more concentrated. One chipmaker and three cloud firms supply most of it.

The concern

AI uses a lot of water and energy, more every year, and the companies disclose little about it. Many see it as an environmental disaster.

What we’d add

• In 2023, U.S. data centers used about 17 billion gallons of cooling water, less than American golf courses use in two weeks and well under a single day of U.S. farm irrigation.
• In a dry region, cooling a datacenter with water can strain a tight supply, and using less water there means burning more electricity.
• Electricity use is a bigger concern than water. Data centers use a few percent of the U.S. grid, and companies keep building more.
• Often people who raise water use are worried about whether AI is worth doing at all.

The concern

A lot of open and public writing was used for training without anyone being asked. Many see it as theft, with companies profiting from work they took for free.

What we’d add

In 2025, courts ruled that the training itself is fair use, even for copyrighted work, so most of it is legal. Anthropic still paid 1.5 billion dollars for downloading the books from pirate sites, which the training ruling did not cover. Consent and payment are being worked out alongside copyright now, through licensing deals and Creative Commons’ new preference signals, which let creators say how their work may be used.

The concern

Writing is worth something because a person is behind it, trying to say something they mean. Machine text can read just as well with nothing behind it. As more of it fills the web, human writing gets harder to find.

What we’d add

What is writing for? We don’t really ask. If you only need the information, a machine serves you about as well as a person. When you want to hear from a particular person, you get something else from a machine, even with the same words. We see it most with students. We ask them to write because writing is how they think, and we use the essay to check the thinking happened. A student can now hand in the essay and skip the thinking. There’s no neat answer here, which is why it’s on the agenda for the discussion.

The concern

Students can now produce essays in seconds, so our usual ways of assessing writing stop working. Writing is also where a lot of thinking happens.

What we’d add

Detectors and bans don’t fix it. They mis-flag honest work, fall hardest on non-native writers, and have wrongly accused students of cheating. So we rethink how we assess.

Adopt, with reasons

Use it where it helps, with someone accountable for the result, and say why.

Refuse, with reasons

Decline it where it doesn’t fit, and say why. The GC guidance leaves room for exactly that.