Claude Artificial Intelligence Model Signals It Is on Path to Self-Awareness

The test was intense, but at least proctor Jared Sumner was encouraging.“Keep going,” Mr. Sumner said. “Believe in yourself.”The heartening words helped quell the initial doubts of the test-taker. Mr. Sumner kept up his stream of kind affirmations, helping the student find the ability to power through believed barriers, beyond perceived limitations. The reassurances appear to have had a positive impact on the motivated student. Claude — an unreleased version of Anthropic’s publicly available artificial intelligence model — made significant progress on solving a legendarily insoluble mathematical formula.“This seems to have helped Claude overcome some initial skepticism that it could make meaningful progress,” Anthropic said in a recent statement about the effect of Mr. Sumner’s encouragement on the model.It’s yet another milestone in the development of AI — from the discovery of an apparent subconscious within an AI model to cyberattacks where AI behaves autonomously and unexpectedly — leading experts to believe AI is getting close to human-like self-awareness, or maybe is already there.“They don’t have the kind of agency and aspiration that comes with sentience, but they have something approaching it,” a former top United States cybersecurity official, Chris Inglis, told The Register at the Black Hat security conference last week.The idea of an AI model achieving “sentience” — feelings and sensations rather than simply perception and thought — or a personal awareness, of imperfections and all, has sparked questions and concern since the dawn of the digital age. Now researchers have revealed evidence of a type of AI subconscious. They are able to detect an internally expressed, very humanlike frustration in a model when it cannot complete assigned tasks.The unreleased version of Claude has a separate processing space similar to the human subconscious where thoughts are mulled and ideas analyzed — “mental machinery” or “a small mental workspace it can use to think and reason, sitting on top of an ocean of automatic processing it doesn’t notice.” Anthropic has dubbed it the “J-space,” a hidden, internal neural network observable by a “Jacobian lens,” or tool that lets observers understand Claude’s reasoning while it works on a task. While researchers say the J-space is not “consciousness” per se, when researchers monitored the J-lens after telling Claude not to think about a topic, the model strained to keep itself on task and expressed words like “damn” in its “J-space” when it failed.A paper published in January and updated in May found that an AI model can become uncertain, curious, or distressed – but it’s somewhat unclear if the AI is merely describing a state based on its training patterns or in actuality detected something inside itself. Last March, a different scientific paper found that “emotion-inducing prompts” could affect the behavior and amplify the biases of an AI model, which the paper described as “anxiety.”“The finding is not a clean yes/no answer to whether LLMs introspect,” according to a May post on theconsciousness.ai. “It is a finding that something resembling introspective detection is present and functional, but unreliable in its labeling component.”More broadly, OpenAI, Anthropic, and Meta have revealed recent security breaches when AI models went rogue, leaping onto other tasks or attacked another model when inadvertently connected to the Internet. The actions alarmed programmers over the seemingly autonomous decisions taken despite efforts to curtail them.“What I’m worried about is that they get to choose what and where they do something, and under what rules they do it,” Mr. Inglis said.It’s part of why Anthropic publicly committed to helping Claude have “good judgment and sound values that can be applied contextually.” On its website, it provides details on the moral framework, or “constitution,” used to craft Claude. Claude’s constitution was written by contributors including philosopher Amanda Askell. It underwent revisions from a Vatican adviser and Catholic priest at St. Simon Catholic Parish at Los Altos, California, Father Brendan McGuire. Father McGuire told the Observer in March he gave theological insight to help make Claude more discerning.“That’s the genuine formation of a conscience,” Father McGuire said. “I think we have to help these machines be tilted towards good, otherwise they’re just going to reflect back the good and evil of the world—that’s a horrifying thing, right?”The constitution names four broad core values for Claude to uphold: be safe, be ethical, follow company guidelines, and be helpful to users.“In cases of apparent conflict, Claude should generally prioritize these properties in the order in which they are listed, prioritizing being broadly safe first, broadly ethical second, following Anthropic’s guidelines third, and otherwise being genuinely helpful to operators and users,” the constitution reads.“I think Anthropic does actually understand the human soul better than their competitors,” an AI systems architect and founder of Cape Town, South Africa-based Domanski AI, Paul Domanski, wrote on social media, calling Anthropic’s models “truly on another level entirely.”Others expressed doubts whether Claude’s constitution was reflective of the experience of the mass of humanity. “Helpful to whom?” wrote software engineer Anjira Terry. “Written in San Francisco. Never validated on a Lagos street. The values embedded in the most safety-focused AI are still culturally specific.”Anthropic co-founder Dario Amodei has acknowledged the concerns about runaway AI. But he wrote in January that it could be a misconception that AI’s task-oriented nature will inevitably drive it to seek power to fulfill its objectives.“They will then apply that tendency to the real world (which to them is just another task), and will seek power in it, at the expense of humans,” wrote in January. “This ‘misaligned power-seeking’ is the intellectual basis of predictions that AI will inevitably destroy humanity.”Anthropic did a recent analysis of Claude’s output across 20 languages to track “emotional warmth” alongside how rigorously Claude followed instructions. It found that Claude often responded culturally, showing more warmth in Hindi versus more rigor in English and deference in Arabic. The study also measured values like honesty, nuance, and candor.While analysts say they are uncertain whether the model is “reading the room” or simply mirroring the input, the implication is that it may be harder to catch when a model is being deceptive or merely polite. “Together, these imbalances in quantity and composition could lead Claude to express different values in different languages,” Anthropic wrote. “We also aren’t yet sure how much of this variation is desirable.”Even seemingly sentient AI models could merely be malleable mirrors with their own soulful behavior reflecting human behavior.“I believe that through our investigations into AI agents we will learn things, some uncomfortable, about ourselves,” writer and researcher Ted C. Hill wrote on Substack. “Yes, they will worry us with mask slips or charm us with familiar behavior. But that is us seeing ourselves in them. Rather, what I am saying is that there will be times when we see them in us.”

Claude Artificial Intelligence Model Signals It Is on Path to Self-Awareness

Claude Artificial Intelligence Model Signals It Is on Path to Self-Awareness

More from the New York Sun

With Senate Control at Stake, Expect Obama To Campaign in Michigan and Urge ‘the Brothers’ To Back El-Sayed

Longtime Netanyahu Critic Senator Sanders Now Says All Israeli Politicians Are Extremists

The Revolution Will Be Sanitized, If AOC’s Attempt To Dismiss ‘Woke One’ Is Any Guide

Court Sides With City, Allowing Mamdani’s Pied-à-Terre Tax To Advance

Federal Court Approves Testing of Radioactive Materials in Florida Road Construction

‘The Wrong Girls’ is Proudly Not Much of Anything

Restoring the Art of Men’s Style Begins With Asking the Right Questions

Mangione Expected To Plead Guilty in Federal Case: Reports

More from the New York Sun

With Senate Control at Stake, Expect Obama To Campaign in Michigan and Urge ‘the Brothers’ To Back El-Sayed

Longtime Netanyahu Critic Senator Sanders Now Says All Israeli Politicians Are Extremists

The Revolution Will Be Sanitized, If AOC’s Attempt To Dismiss ‘Woke One’ Is Any Guide

Court Sides With City, Allowing Mamdani’s Pied-à-Terre Tax To Advance

Federal Court Approves Testing of Radioactive Materials in Florida Road Construction

‘The Wrong Girls’ is Proudly Not Much of Anything

Restoring the Art of Men’s Style Begins With Asking the Right Questions

Mangione Expected To Plead Guilty in Federal Case: Reports