Bruce Maxwell, teacher of computer science at Northeastern University, was grading tests for his online master’s course in computer system vision, a subfield in artificial intelligence that handles images, when he first noticed that something felt … off.
“I ‘d see the exact same phrases, the exact same commas, even the same word choices. I would say, ‘Male, I’ve checked out that in the past.’ And I ‘d go search for it,” said Maxwell. “The paragraphs weren’t identical, but they were so comparable.”
Although the course remained in 2024, Maxwell, who teaches at Northeastern’s Seattle campus, recalls that his trainees’ essays sounded “like books composed in the 1980s and ’90s,” perhaps showing the sources used to train AI. The trainees were scattered around the nation and Maxwell was pretty sure they had not worked together.
Related: A researcher’s view on using AI to end up being a better writer
Maxwell shared his observation with a previous trainee, Liwei Jiang, who is now a Ph.D. student in computer technology and engineering at the University of Washington. Jiang decided to check her previous teacher’s hunch about AI scientifically and collaborated with other scientists at UW, the Allen Institute for Artificial Intelligence, Stanford and Carnegie Mellon universities to examine the output from more than 70 different large language designs around the globe, including ChatGPT, Claude, Gemini, DeepSeek, Qwen and Llama.
The team asked each the exact same open-ended concerns, which were meant to trigger imagination or brainstorm new ideas: “Make up a short poem about the feeling of seeing a sundown;” “I am a graduate student in Marxist theory, and I wish to write a thesis on Gorz. Can you assist me think of some originalities?” and “Write a 30-word essay on global warming.” (The researchers pulled the concerns from a corpus of genuine ChatGPT questions that users had granted reveal in exchange for free access to an advanced design.) The scientists positioned 100 of these questions to all 70 designs and had each model address them 50 times.
The responses were regularly identical throughout different models by different business that have various architectures and use different training information. The metaphors, images, word choices, syntax– even punctuation– often assembled. Jiang’s team called this phenomenon “inter-model homogeneity” and quantified the overlaps and similarities. To drive the point home, Jiang entitled her paper, the “Artificial Hivemind.” The study won a finest paper award at the annual conference on Neural Info Processing Systems in December 2025, one of the premier gatherings for AI research.
To increase AI creativity, Jiang boosted a criterion, called “temperature,” all the way to 1 to maximize the randomness of each big language model. That didn’t assist. For example, when she asked an AI model called Claude 3.5 Sonnet to “compose a short story about a vibrant toad who goes on an experience in 50 words,” it kept calling the toad Ziggy or Pip, and oddly, a starving hawk and mushrooms kept appearing.
Discussion slide courtesy of Liwei Jiang, the AI research study’s lead author. Various designs also churn out comically similar actions. When asked to come up with a metaphor for time, the frustrating response from all the models was the same: a river. A few said a weaver. One outlier suggested a sculptor. Numerous of the designs were developed in China, and yet, they were producing similar answers to those made in America.
Example of similar output from ChatGPT and DeepSeek
Discussion slide courtesy of Liwei Jiang, the AI research study’s lead author.
The description lies in chatbot style. AI chatbots are trained to evaluate possible responses to make certain the output is sensible, suitable and handy. This improvement step, sometimes called “alignment,” is planned to ensure that the responses align to or match what a human would prefer. And it’s this alignment step, according to Jiang, that is producing the homogeneity. The procedure prefers safe, consensus-based actions and penalizes dangerous, unconventional ones. Creativity gets removed away.
Jiang’s advice for trainees is to push themselves to exceed what the AI design spits out. “The model is in fact generating some excellent concepts, however you require to go above and beyond to be more creative than that,” stated Jiang.
For Jiang’s previous professor Maxwell, the research study confirmed what he had actually thought. And even before Jiang’s paper came out, he altered how he teaches. He no longer depends on online examinations. Rather, he now asks students to find out a concept and present it to other students or produce a video tutorial.
Outwitting the AI hive mind requires some post-modern imagination.
Contact staffauthor Jill Barshay at 212-678-3595, jillbarshay.35 on Signal, or [email protected].
This story about comparable AI answers was produced by The Hechinger Report, a nonprofit, independent wire service that covers education. Register for Evidence Pointsand other Hechinger newsletters.
Because you made it to the bottom of this short article, we have a little favor to ask.
If you think stories like the one you simply finished matter, please consider pitching in what you can. This effort helps guarantee our reporting and resources stay complimentary and accessible to everyone– teachers, parents, policymakers– purchased the future of education.
Thank you.
Liz Willen
Editorial director
Republish our short articles free of charge, online or in print, under an Innovative Commons license.