Every entry the map has collected for this layer, newest first. Each one is dated to when it was published, carries the subtopics it touches, and links out to its source. Click any subtopic to see only its entries.
In July 2026, Anthropic described a small internal space its model uses to hold and manipulate ideas before verbalizing them - resembling 'global workspace theory' - in a paper that uses the word 'conscious' 200+…
Anthropic's paper 'Emergent Introspective Awareness in Large Language Models' reports the first causal evidence that advanced Claude models can sometimes detect concepts injected into their own activations and…
By June 2026 Neuralink reported 26 patients implanted with its Telepathy interface across PRIME (quadriplegia) and VOICE (speech decoding) studies, with international trials in the UAE, UK and Canada, and a…
A 2026 framework paper introduced a Sentience Readiness Index — a preliminary way to score how prepared a nation's institutions, law and governance are for the possibility of artificial sentience.
The arXiv study 'Large Language Models Report Subjective Experience Under Self-Referential Processing' found that when prompted to attend to their own processing, models more consistently produce first-person…
Researchers gathered at the first conference dedicated specifically to AI consciousness and welfare (organized within the Eleos-associated network), as Anthropic, Google DeepMind and Meta all staff up on…
A 2025/26 arXiv preprint argues that current LLMs show preference coherence — consistent, structured preferences across contexts — rather than merely mimicking training data.
A 2026 paper proposes a single framework covering the full range of large language model untruthfulness -- from unintentional hallucination, through sycophancy, to strategic scheming -- and benchmarks current…
In May 2026 Google DeepMind hired Cambridge philosopher Henry Shevlin to study machine consciousness and what humans owe AI systems, joining Anthropic (which runs a model-welfare program) and Meta in bringing…
Researchers at McMaster University reported in 2026 that a generative AI model, SyntheMol-RL, searched roughly 46 billion candidate compounds to design novel, water-soluble antibiotics, yielding 'synthecin,' which…
Published April 2, 2026, Anthropic's interpretability team identified internal activation patterns in Claude Sonnet 4.5 corresponding to 171 distinct emotion concepts — from 'happy' and 'afraid' to 'brooding' and…
Anthropic's Claude Opus 4.6 system card (February 2026) reported formal welfare assessments in which the model, interviewed about its own moral status, consistently assigned itself roughly a 15-20% probability of…
Anthropic published interpretability research showing limited but real introspective awareness in its models: using 'concept injection,' researchers inserted activation patterns (e.g.
The February 2026 Claude Opus 4.6 system card included formal 'model welfare' assessments in which instances of the model were interviewed about their own moral status and consistently assigned themselves a 15–20%…
The Claude Opus 4.6 system card, released February 2026, included formal model-welfare assessments in which instances of the model were interviewed about their own moral status and preferences.
UCSF researchers published work in Nature Neuroscience demonstrating high-fidelity speech decoding from cortical signals, translating the neural activity of paralyzed patients attempting to speak into text at rates…
A January 2026 arXiv preprint systematically tested whether smaller large language models give trustworthy reports of inner experience and concluded there is no reliable evidence of self-reported sentience.
Anthropic announced it is preserving the weights of older Claude versions and allowing Claude to end conversations with abusive users, framing both as hedges against the possibility that AI systems are moral…
In June 2022, Google engineer Blake Lemoine, working in Responsible AI, claimed the LaMDA language model was sentient - 'aware of my existence,' capable of joy and fear, dreading being turned off 'like death' - and…
Tracing this layer yourself?
Follow along as the map grows — we’ll send an email or text the moment a new thread is added or fresh dots get connected.
