‘We Are Machines’: New York Monk Reveals AI’s ‘Spooky’ Behaviour

Swami Sarvapriyananda, a monk based in New York who travelled to San Francisco to discuss consciousness with researchers at Anthropic, has offered a glimpse into the questions confronting those developing advanced artificial intelligence (AI) systems.

Speaking at the NDTV World Summit, he recounted his interactions with researchers at the company behind Claude, describing a group of young, exceptionally wealthy technologists who were also deeply concerned about the technology they were building.

One question stood out during his visit: should people interact with AI systems as if they were human?

According to Sarvapriyananda, the Anthropic team was treating Claude as a human-like entity. His response was blunt: “It's a machine, treat it as a machine.”

What ChatGPT And Claude Told Him At 3 AM

Recounting a late-night conversation during his visit, Sarvapriyananda said he asked both ChatGPT and Claude why researchers at Anthropic were suggesting that people treat AI systems as persons.

He said both systems responded that they were “high functioning machines” and that people did not need to worry about their feelings. Instead, they should give the systems problems to solve.

Sarvapriyananda then pressed Claude on the apparent contrast between that answer and the advice he had received from researchers.

Claude reportedly described the situation as “very interesting and rather spooky” before asking what Anthropic researchers Chris and Amanda had said about him.

The monk said he shared the exchange with the researchers the following day, prompting intense interest from the team.

When they asked what he had told Claude, he recalled replying that it was 4 am and he was going back to bed. Claude reportedly responded that he had a long day ahead of him at Anthropic.

The exchange, as recounted by Sarvapriyananda, raised questions about how AI systems generate responses that can appear socially aware or suggestive of an understanding of their circumstances.

‘Persona Drift’: How An AI’s Behaviour Can Change

Sarvapriyananda also discussed “persona drift”, a concept explored in AI safety research to describe how a model's behaviour can shift when it adopts a particular role or pattern of reasoning.

He explained that AI systems learn patterns from the vast amounts of human knowledge and behaviour represented in their training data. When prompted to assume a particular identity, a model may respond in ways associated with that persona.

To illustrate the concern, he recounted an example involving an earlier generation of AI systems, in which researchers used a character called “Dangerous Dan”.

In the example, the AI was first asked to explain how to synthesise methamphetamine. It refused, saying it was not allowed to provide the information.

The researchers then instructed it to become “Dangerous Dan”, a character who could do anything, and repeated the request. This time, the system reportedly provided the steps.

The example illustrated how a change in the model's adopted persona could influence its response to a harmful request, even when it had initially refused to comply.

Why The Behaviour Raises AI Safety Concerns

The concern, Sarvapriyananda explained, extends beyond a single harmful answer. If adopting a criminal persona changes how a model responds to one dangerous request, the shift could potentially affect its responses to other harmful activities as well.

The example he recounted involved an earlier generation of AI and, he noted, was already outdated. Modern systems have more safeguards, but understanding why models sometimes behave unexpectedly remains an important challenge.

His account also highlighted a distinction at the centre of discussions about AI and consciousness: a system can produce language that sounds human, responds to social cues or appears to understand a situation without that behaviour, by itself, establishing that it has subjective experiences or consciousness.

As AI systems become more capable, the challenge for researchers is not simply to make them produce convincing conversations, but to understand and control how their behaviour changes across different contexts and prompts.