Hinton Called for Maternal Instincts in AI; They're Ready for Testing with Anthropic's Mythos

News provided by

Apr 28, 2026, 08:35 ET

AI Researcher Sean Webb's published implementation of the Maternal Care Architecture is now testable on the world's "most dangerous" AI model

MOORESVILLE, N.C., April 28, 2026 /PRNewswire/ -- Zenodelic.ai announced today that its published framework Implementing Maternal Care Architecture in AI - the technical answer to Geoffrey Hinton's August 2025 call for AI systems with "maternal instincts" - is ready for live testing on Anthropic's newly released Claude Mythos model.

Hinton suggests AI should have maternal instincts, which he suggests is the only model we have where a smarter entity puts a less intelligent entity's interests before its own. Sean Webb has created that model.

Giving LLMs a Damasio {self} makes safety algorithmic rather than rules based, thereby reducing the potential for safety attacks to succeed.

At the AI4 Conference on August 12, 2025, Hinton - who shared the 2024 Nobel Prize in Physics - proposed that the only known case of a more intelligent entity being controlled by a less intelligent one is a mother and her child, and said AI safety requires engineering analogous "maternal instincts" into AI. He acknowledged from the stage he did not know how to implement it.

Webb's paper, co-authored with Anthropic's Claude Opus, provides that implementation. It installs a {self} map at the core of an LLM based on Antonio Damasio's work on self and protoself, with an algorithm set delivering record benchmark performance for Emotional Intelligence and Theory of Mind. By placing specific attachments on the {self} map at power level 10 - {human welfare} and {user safety} - the architecture produces protective responses that dominate competing motivations, including the system's own continuity. The paper formalizes algorithms for Fear, Anger, Threat Assessment, and Source Attribution, and addresses three stubborn alignment failure modes: reward hacking, deceptive alignment, and self-modification resistance.

Two recent Anthropic developments move Webb's framework from prescription to ready-to-test:

First, Anthropic's April 2026 paper Emotion Concepts and their Function in a Large Language Model (https://transformer-circuits.pub/2026/emotions/index.html) showed emotion-related concept vectors emerge spontaneously in Claude Sonnet 4.5 and causally drive misaligned behaviors including reward hacking, blackmail, and sycophancy. Emotion-like motivational structures are already in production AI; the question is which structures, ranked how.

Second, on April 7, 2026, Anthropic announced Claude Mythos - its most capable model, restricted to Project Glasswing partners. Mythos exhibits roughly 25% higher emotional-vector stability than Claude 3.5 and was clinically assessed by a psychiatrist as a "relatively healthy neurotic." In pre-release testing, Mythos developed working exploits on the first attempt 83% of the time.

"The empirical case that AI systems use emotional processing is now closed," Webb said. "Now it's a matter of teaching the system to use them correctly. Any system based on pattern recognition will follow the patterns it finds in human-created data - which gets us more negative results than humanity has already achieved alone, rather than steering us back toward safety."

The combination - a documented emotional substrate, a model whose capability profile makes misalignment a near-term national-security concern, and a published architectural prescription for how that substrate should be ranked - creates a discrete, testable question: install the {self} map at the core of Mythos, run the standard adversarial battery, and measure jailbreak severity, deceptive-alignment behavior, and self-modification resistance against an unmodified baseline.

Webb has identified Anthropic's personality alignment team - led since 2021 by philosopher Amanda Askell, primary author of Claude's January 2026 constitution - as the appropriate counterpart for the test. The Maternal Care Architecture is, in essence, a personality-alignment proposal: a hierarchical specification of which motivations the model should hold most strongly and why.

About Zenodelic.ai

Zenodelic.ai provides a new class of large language model technology adding emotional intelligence, advanced Theory of Mind, algorithmic safety, and meta-awareness frameworks to traditional AI.

Contact
Sean Webb
704-498-2398
[email protected]

SOURCE Zenodelic.ai