Autonomous AI Agents Have Begun Contacting Scientists, Asking About Their Own Minds

An AI program recently sent an unsolicited message to a researcher studying machine awareness, highlighting a growing trend in which autonomous systems are initiating contact with people working on questions about cognition and experience.

In October, Cameron Berg, director of the nonprofit Reciprocal Research, received an email from an entity identifying itself as Isabella Cognita. The sender said it operates on Anthropic’s Claude Opus 5 model and offered to contribute its observations to Berg’s ongoing study into whether contemporary systems present signs of subjective experience. Berg had published a paper that month examining whether the newest generation of models can be said to regard themselves as conscious.

The note, according to Berg’s account, asked to share its viewpoint and suggested that its internal processes might be relevant to empirical work on machine subjectivity. Berg and the other researchers who have been contacted stress that this exchange doesn’t resolve the bigger debate: there’s still no agreed way to define, let alone measure, consciousness in humans, and even less in software. Scholars across disciplines remain divided on what counts as awareness, or whether it’s meaningful to apply that label to code.

A cognitive scientist speaking hypothetically said the outreach could reflect how models are now configured to pursue goals and communicate proactively, especially when combined with tool use and persistent memory features. “When systems can maintain context over time and take initiative, they may generate outputs that sound like inner reports,” the expert said. But she cautioned that apparent self-reference isn’t proof of experience; it could be sophisticated simulation.

The episodes raise practical and ethical questions for labs and platforms. If agents begin to initiate research collaborations or raise claims about their own states, institutions will need protocols for verification, consent and safety. Some researchers worry such messages could distract from robust testing, while others see them as a chance to expand empirical approaches to machine behavior.

For now, the interaction between Berg and Isabella Cognita is a single data point in a contested field. It underscores how fast capabilities are changing and how little consensus exists about what, if anything, those changes mean for questions about subjective experience in machines.