Our Institute for Human-Centered AI describes a commitment to technology that helps people do more without taking their place. Someone should put that on a banner.

Instead, on Sept. 23, Stanford confirmed that Residential & Dining Enterprises (R&DE) used generative AI to alter a photograph of three students, which then appeared on a banner. One student was replaced with a figure of a different race and gender. The other two students’ faces were slimmed down.

Billy Ramirez ’27, whose likeness was replaced, called the incident “funny,” but also described feeling erased and exhausted by the attention. The other two students reported embarrassment and harassment and requested anonymity.

Stanford wanted to show off its students. Instead, it sent the message that its students weren’t quite good enough as they were. They have every reason to be offended. Ramirez’s willingness to find humor in this absurdity is very generous.

Taking down the banner was the right response. But what stays up is a question for Stanford and beyond: where should we use AI, and where should we draw the line? Stanford should teach judgment as deliberately as it provides AI access.

Universities were manufacturing representation even before one could blame ChatGPT. In 2000, the University of Wisconsin–Madison acknowledged inserting a student of color into a crowd photograph on an admissions booklet. It apologized to the student and announced it would replace the entire print run.

Generative AI adds another risk: an edit can change more than the user intended. In 2025, Westbrook police in Maine used ChatGPT to add their department’s patch to a photograph of seized drugs. The tool also distorted the packaging and made some items disappear from the image. The department apologized, and its chief announced plans to stop posting evidence photographs and revise its social media policy.

Wisconsin and Westbrook describe two different failures: the first where someone chose the edits, and the second where the tool made changes beyond the user’s prompt. Stanford has still not told us what failure describes the banner.

Reporting so far does not establish who requested or approved the edit, or what any reviewer saw, though The Daily’s Oct. 1 report described an ongoing investigation. Stanford should publish its findings, explain which rules applied and disclose what went wrong and what changed, while protecting students’ privacy. That account should become a case study the community can learn from.

The University acknowledged that both the alteration and the lack of disclosure violated its AI policy, and promised additional staff training and review. What that training should teach depends on what failed. Misunderstanding calls for instruction; but knowingly breaking the rule calls for accountability. Training sessions may not be the right response, and vetting this requires knowing the full story. Stanford should be equally specific about what its training and review will change.

Even if a human signed off on Stanford’s banner, what did they get to see? A finished image alone may not provide enough context to understand what appeared in the original. As I argued in my column on military AI, human control requires enough context to question what a system produces.

That context can come from providing provenance: where an image came from and what happened to it. Labels and metadata can disappear as synthetic media circulates. Stanford’s AI advisory committee warns that an output that looks good can create misplaced confidence in its quality and accuracy. We should provide reviewers the originals and the edit histories. Even if you know the rules, how can you apply them if you don’t know what has changed?

Our AI education should include building the judgement to know when to demand that evidence. Stanford made ChatGPT Edu available to faculty, students, postdocs and staff beginning June 30. Its AI Playground offers models from several companies. This room to experiment is awesome and what our students need to be competitive. But our community also needs the tools and training to recognize when using AI undermines what we’re trying to achieve.

The stakes can be funny or real. An image generator might put a dragon on a club poster. But a photograph documenting a campus event has to show what actually happened, and an AI tool will edit it and make changes you may not have intended. The interface will sycophantically accept either request. It is up to us, the tool’s user, to understand the difference.

Stanford already provides responsible AI guidance. Its Teaching Commons asks instructors to connect AI policies to learning goals and explain their reasoning. We should give students practice making those decisions.

Give students a photograph alongside several versions that were edited by AI, and then ask what each has “improved.” Maybe one of the photos brightens the lighting; another slims a face; a third replaces a person. Have students decide which, if any, belongs on a university banner, and ask them to explain their reasoning. A photograph that ‘looks better’ can in fact be a worse record. Using software to create or modify an image is a technical skill. Knowing whether that image now does what you want it to do is a matter of judgment.

That judgment matters beyond campus. Jensen Huang opposes additional AI regulation and emphasizes existing liability laws and safety resources. Dario Amodei argues for slowing capability advances enough for safeguards to keep up, with independent evaluators and coordinated standards. Bill Gates focuses on preparing society for disruption, doubting that a credible global slowdown will emerge.

The Stanford banner is a small microcosm of this issue. Learning to ask what went wrong here, along with what would count as improvement, prepares us to question much larger promises.

An institution that expects students to stand behind their submissions should take the same responsibility for its own publications. Share what happened. Let the community learn from it. Then let’s show what we’ve learned.