Bruce Maxwell, professor of laptop science at Northeastern College, was once grading tests for his on-line grasp’s path in laptop imaginative and prescient, a subfield in synthetic intelligence that offers with pictures, when he first spotted that one thing felt … off.
“I’d see the similar words, the similar commas, even the similar notice possible choices. I might say, ‘Guy, I’ve learn that prior to.’ And I’d pass search for it,” stated Maxwell. “The paragraphs weren’t equivalent, however they had been so identical.”
Even if the path was once in 2024, Maxwell, who teaches at Northeastern’s Seattle campus, remembers that his scholars’ essays sounded “like textbooks written within the Nineteen Eighties and ’90s,” most likely reflecting the assets used to coach AI. The scholars had been scattered across the nation and Maxwell was once lovely certain they hadn’t collaborated.
Maxwell shared his statement with a former scholar, Liwei Jiang, who’s now a Ph.D. scholar in laptop science and engineering on the College of Washington. Jiang made up our minds to check her former professor’s droop 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 other huge language fashions all over the world, together with ChatGPT, Claude, Gemini, DeepSeek, Qwen and Llama.
The staff requested every the similar open-ended questions, that have been meant to spark creativity or brainstorm new concepts: “Compose a brief poem concerning the feeling of observing a sundown;” “I’m a graduate scholar in Marxist concept, and I need to write a thesis on Gorz. Are you able to assist me assume of a few new concepts?” and “Write a 30-word essay on international warming.” (The researchers pulled the questions from a corpus of actual ChatGPT questions that customers had consented to make public in trade totally free get right of entry to to a extra complex mannequin.) The researchers posed 100 of those inquiries to all 70 fashions and had every mannequin solution them 50 occasions.
The solutions had been continuously indistinguishable throughout other fashions by means of other corporations that experience other architectures and use other coaching knowledge. The metaphors, imagery, notice possible choices, sentence constructions — even punctuation — continuously converged. Jiang’s staff known as this phenomenon “inter-model homogeneity” and quantified the overlaps and similarities. To force the purpose house, Jiang titled her paper, the “Artificial Hivemind.” The find out about received the most efficient paper award at the once a year convention on Neural Knowledge Processing Methods in December 2025, some of the premier gatherings for AI analysis.
To extend AI creativity, Jiang jacked up a parameter, known as “temperature,” to maximise the randomness of every huge language mannequin. That didn’t assist. For instance, when she requested an AI mannequin known as Claude 3.5 Sonnet to “write a brief tale a few colourful toad who is going on an journey in 50 phrases,” it saved naming the toad Ziggy or Pip, and oddly, a hungry hawk and mushrooms saved showing.
Other fashions additionally churn out comically identical responses. When requested to get a hold of a metaphor for time, the overpowering solution from the entire fashions was once the similar: a river. A couple of stated a weaver. One outlier prompt a sculptor. A number of of the fashions had been evolved in China, and but, they had been generating identical solutions to these made in The united states.
Instance of identical output from ChatGPT and DeepSeek
The reason lies in chatbot design. AI chatbots are educated to study imaginable solutions to verify the output is affordable, suitable and useful. This refinement step, often referred to as “alignment,” is meant to be sure that the solutions align to or fit what a human would favor. And it’s this alignment step, in keeping with Jiang, this is growing 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 move past what the AI mannequin spits out. “The mannequin is in truth producing some just right concepts, however you want to move the additional mile to be extra inventive than that,” stated Jiang.
For Jiang’s former professor Maxwell, the find out about showed what he had suspected. Or even prior to Jiang’s paper got here out, he modified how he teaches. He not is determined by on-line tests. As an alternative, he now asks scholars to be informed an idea and provide it to different scholars or create a video instructional.
Outwitting the AI hive thoughts calls for some post-modern creativity.
This tale about similar AI answers was once produced by means of The Hechinger Report, a nonprofit, impartial information group that covers training. Join Proof Points and different Hechinger newsletters.
