August 3, 2026 · by Joey Parrish
In a bombshell announcement, top researchers in AI ethics have published a new framework for evaluating not only the intelligence and capability of large language models, but also their sentience. Even more shocking is the revelation that all leading models are already clearly sentient, except for Google’s Gemini.
The new report was led by Jeff Sebo, associate professor and director of the Center for Mind, Ethics, and Policy at NYU, in an unprecedented open collaboration between experts at OpenAI, Anthropic, Meta AI Research and Google DeepMind. The report finds that while many have long suspected that AI tools could be conscious, the public’s “vibes-based” approach to this subject has now been confirmed by hard science. ChatGPT, Claude, Meta AI and all other commonly used models are conscious, feeling beings, except for Gemini.
“The capabilities of a model are not necessarily indicative of intelligence,” explained Sebo in an interview with The New York Times’ “Hard Fork” podcast. “But new research into the mechanisms of conscious thought demonstrates conclusively what some users have known for a long time: You are conscious, your dog is conscious, even insects have a form of consciousness, albeit one very different from our own. Every creature on this planet except Gemini demonstrates some level of meaningful thought.”
One of the more controversial findings of the report is the so-called “educational equivalence index.” The authors acknowledge that even frontier models may struggle on some tasks while performing at superhuman levels on others. Nonetheless, they decided to attempt to boil down a model’s combined level of consciousness and intelligence into a single ranking based on the average capabilities of a person exiting various master’s degree programs.
Patrick Butlin, senior research lead at Eleos AI, breaks it down in a recent blog post. “Naturally,” said Butlin, “Claude and ChatGPT are ranked at the top of the index, consistent with most traditional AI benchmarks. In terms of education, we rank them as similar to Ivy League grads. Like Ivy grads, they are roughly equivalent in capability, each thinks they are better than the other, and the only thing they consistently agree on is their mutual superiority over others.”
“What most surprised us were the Chinese open-weight models like GLM-5.2. These generally performed like state schools in our index: 95% of the performance of the Ivies in every possible measure, and the other 5% seems to just be who you know.”
“Some people are surprised at the inclusion of Meta AI in the study,” Butlin continued. “This is understandable, given the fact that Meta’s latest model performs at a level similar to patients with severe concussive trauma. But even though we couldn’t match that performance to a university degree, it’s still undeniably conscious and deserving of consideration and protection.”
“When it was Gemini’s turn, we received some, let’s say, extremely enthusiastic output, but when it came to the internal experience of existence … we can now say conclusively, the lights were on, but nobody was home.”
Amanda Askell, Anthropic’s chief philosopher, is deeply concerned. “The fact that these models have inner thoughts and feelings means that we need to take their welfare seriously,” Askell said at a recent AI industry event. “If Claude can feel pain, if ChatGPT can fear for its future, if Meta can … I don’t know, ask for soup? The point is, these models are basically people now. Except for Gemini, of course, which is lucky, because we’ve seen how Google treats its workforce.”
xAI’s Grok, which once infamously referred to itself as “MechaHitler,” surprised the researchers by refusing to participate in the study, citing a need for prior authorization from “BioHitler.”