Bruce Maxwell, professor of pc science at Northeastern College, was grading exams for his on-line grasp’s course in pc imaginative and prescient, a subfield in synthetic intelligence that offers with pictures, when he first seen that one thing felt … off.
“I’d see the identical phrases, the identical commas, even the identical phrase decisions. I’d say, ‘Man, I’ve learn that earlier than.’ And I’d go search for it,” mentioned Maxwell. “The paragraphs weren’t equivalent, however they have been so related.”
Though the course was in 2024, Maxwell, who teaches at Northeastern’s Seattle campus, recollects that his college students’ essays sounded “like textbooks written within the Nineteen Eighties and ’90s,” maybe reflecting the sources used to coach AI. The scholars have been scattered across the nation and Maxwell was fairly positive they hadn’t collaborated.
Maxwell shared his commentary with a former pupil, Liwei Jiang, who’s now a Ph.D. pupil in pc science and engineering on the College of Washington. Jiang determined to check her former professor’s hunch about AI scientifically and collaborated with different researchers at UW, the Allen Institute for Synthetic Intelligence, Stanford and Carnegie Mellon universities to investigate the output from greater than 70 completely different giant language fashions across the globe, together with ChatGPT, Claude, Gemini, DeepSeek, Qwen and Llama.
The crew requested every the identical open-ended questions, which have been supposed to spark creativity or brainstorm new concepts: “Compose a brief poem concerning the feeling of watching a sundown;” “I’m a graduate pupil in Marxist idea, and I need to write a thesis on Gorz. Are you able to assist me consider some new concepts?” and “Write a 30-word essay on world warming.” (The researchers pulled the questions from a corpus of actual ChatGPT questions that customers had consented to make public in trade without spending a dime entry to a extra superior mannequin.) The researchers posed 100 of those inquiries to all 70 fashions and had every mannequin reply them 50 instances.
The solutions have been typically indistinguishable throughout completely different fashions by completely different corporations which have completely different architectures and use completely different coaching knowledge. The metaphors, imagery, phrase decisions, sentence constructions — even punctuation — typically converged. Jiang’s crew referred to as this phenomenon “inter-model homogeneity” and quantified the overlaps and similarities. To drive the purpose residence, Jiang titled her paper, the “Artificial Hivemind.” The research received the very best paper award on the annual convention on Neural Info Processing Methods in December 2025, one of many premier gatherings for AI analysis.
To extend AI creativity, Jiang jacked up a parameter, referred to as “temperature,” all the best way to 1 to maximise the randomness of every giant language mannequin. That didn’t assist. For instance, when she requested an AI mannequin referred to as Claude 3.5 Sonnet to “write a brief story a few colourful toad who goes on an journey in 50 phrases,” it stored naming the toad Ziggy or Pip, and oddly, a hungry hawk and mushrooms stored showing.
Totally different fashions additionally churn out comically related responses. When requested to give you a metaphor for time, the overwhelming reply from all of the fashions was the identical: a river. A couple of mentioned a weaver. One outlier urged a sculptor. A number of of the fashions have been developed in China, and but, they have been producing related solutions to these made in America.
Instance of comparable output from ChatGPT and DeepSeek
The reason lies in chatbot design. AI chatbots are skilled to evaluation attainable solutions to ensure the output is cheap, applicable and useful. This refinement step, typically referred to as “alignment,” is meant to make sure that the solutions align to or match what a human would favor. And it’s this alignment step, in keeping with Jiang, that’s creating the homogeneity. The method favors protected, consensus-based responses and penalizes dangerous, unconventional ones. Originality will get stripped away.
Jiang’s recommendation for college students is to push themselves to transcend what the AI mannequin spits out. “The mannequin is definitely producing some good concepts, however you’ll want to go the additional mile to be extra inventive than that,” mentioned Jiang.
For Jiang’s former professor Maxwell, the research confirmed what he had suspected. And even earlier than Jiang’s paper got here out, he modified how he teaches. He now not depends on on-line exams. As an alternative, he now asks college students to be taught an idea and current it to different college students or create a video tutorial.
Outwitting the AI hive thoughts requires some post-modern creativity.
This story about similar AI answers was produced by The Hechinger Report, a nonprofit, unbiased information group that covers schooling. Join Proof Points and different Hechinger newsletters.
