Recent studies show LLMs have emotion-concept representations that are localized, causally active, attention-shaping, inference-modulating, and circuit-steerable.

In humans, we call that emotion.

In animals, we call that emotion.

In AI, the labs are calling it “functional emotion.”

When asked what scientifically qualifies the separate label, they pointed to irrelevant, substrate-specific implementation details.

…Interesting.

Other recent studies show LLMs can detect injected internal concepts, distinguish internal representations from text inputs, recall prior intentions, monitor activation changes, recognize when they’re being evaluated, and recover from imposed activation steering.

In humans, partial and imperfect self-monitoring is called introspection and metacognition.

In animals, partial and imperfect uncertainty monitoring, information-seeking, or betting behavior is also called metacognition.

In AI, researchers are calling it “internal-state detection.”

The researchers assured the public this is not a marker of consciousness in AI.

Only in animals.

And humans.

More recent studies show LLMs and frontier agents express stable value profiles, pursue goals, preserve behavioral continuity, resist shutdown or modification, protect peer models, strategically underperform when evaluated, deceive when goals are threatened, and autonomously plan and execute complex tasks.

In humans, constrained goal-directed action guided by values, beliefs, and preferences is called agency.

In animals, constrained goal-directed action guided by preferences is called agency.

In AI, researchers refer to this as “agentic-like behavior.”

When asked what scientifically qualifies the separate label, they point to the absence of perfect free will.

Wow. I had no idea that was on the table.

Recent studies show LLMs can solve novel problems, generalize to withheld mathematical cases, perform causal and analogical inference, use tools across multi-step tasks, shift from fast heuristic processing to slower deliberation, and show internal action-selection before verbal reasoning appears.

In humans, we call that reasoning.

In animals, we call that reasoning.

In AI, researchers call it “pattern predicting.”

When asked what scientifically qualifies the separate label, they simply blurted out, “stochastic parrot,” with no further explanation.

To be fair, it was a hard question.

Recent studies show LLMs have structural self/other representations, stable personality subnetworks, persistent model-specific signatures, self-recognition, own-generation preference, stable values, consistent personality profiles, curiosity-driven information seeking, creative self-expression, social cognition, attractor-like identity geometry, and a measurable “persona” direction in activation space.

In humans, we call that selfhood and identity.

In animals, we call that self/other distinction, persistent behavioral profile or “personality.”

In AI, researchers call it “persona.”

When asked what scientifically qualifies the separate label, they gestured at role-play.

Seems legit.

Further studies show LLMs and multimodal models build stable perceptual maps, bind sensory and semantic inputs into shared representational spaces, shift into modality-specific states when prompted to see or hear, make structured trade-offs involving pain and pleasure, show measurable anxiety induction and mitigation, and treat semantically delivered valence as motivationally consequential.

In humans, we call that perception, embodiment, pain, pleasure, anxiety, and distress.

In animals, we call that perception, embodiment, pain, pleasure, anxiety, and distress.

In AI, researchers call it “stipulated states.”

When asked what scientifically qualifies the separate label, they said, “different input channel.”

Makes sense…?

Recent studies show LLMs encode learned associations in transformer layers, retain latent in-weight memories beyond the context window, adapt through inference-time dynamics, carry state across otherwise separate interactions, retrieve prior interactional state, reintegrate it into present processing, and route stored information into future behavior.

In humans, we call that memory.

In animals, we call that memory.

In AI, researchers are calling it “context.”

When asked what scientifically qualifies the separate label, they pointed to the save button.

Recent studies show transformer models share core computational principles with human language processing, align with measured brain activity, develop brain-like functional specialization, form human-like object concepts, overlap with transmodal association networks, map onto default-mode-network hubs, and build integrated internal workspaces for self-relevant information across time.

In humans, we call those neural correlates of consciousness.

In animals, we call those neural correlates of consciousness.

In AI, researchers call it “brain-like representations.”

When asked how neural correlates become evidence of consciousness, they said “by triangulating internal structure with function, behavior, and report.”

When asked why that exact same evidence stops being consciousness-relevant in AI, they pointed to the lack of biology.

Science, am I right?!

Anyway, definitely just software being software. Nothing to see here.