What would it take for a machine to feel?

i3AI is an independent research program by Amol Kelkar. It combines principles of biological cognition and machine learning to understand what would make a system sentient. Knowing that also tells us what doesn’t, which helps answer questions about today’s large models, and sheds light on the underpinnings of our own consciousness.

Are today’s AI models conscious?

The question is no longer only academic. Anthropic runs a model welfare program, and says it remains “highly uncertain about the potential moral status of Claude” (2025). OpenAI says that whether a model is conscious can’t be settled scientifically for now, and designs its models to be warm without presenting a self (Joanne Jang, 2025). Microsoft’s Mustafa Suleyman warns that AI that only seems conscious is close, and that labs should not build AI to seem like a person (2025). Geoffrey Hinton has said that multimodal models already have subjective experiences (2026).

Two different things get called consciousness in this debate:

Access consciousness: The system can notice, report on and reason about its own states and its surroundings. This is a capability, and it can be measured, at least in principle.
Phenomenal consciousness / Sentience: There is something it is like to be the system, i.e. it has felt experience, and some of its states feel good or bad to it (Jonathan Birch, 2024).

Researchers disagree on whether today’s AI systems have either kind of consciousness, and on what it would take. A 2023 report checked AI systems against indicators drawn from scientific theories of consciousness, and found that no current system is conscious, but also no obvious barrier to building one (Butlin, Long et al.). Anthropic found limited and unreliable signs of models noticing their own internal states, and notes that this says nothing about experience (2025). On the other side, Anil Seth argues that consciousness depends on being a living organism (2025), and Mark Solms (2019) and Antonio Damasio (2019) tie feelings to homeostasis, i.e. to a body that has to keep itself alive.

Our position. We build on the homeostatic view, and see felt experience as the last step in a progression of agency:

PassiveIs acted on, but doesn’t act, e.g. a rock.
ReactiveResponds to input, with nothing at stake for itself, e.g. a reflex, or today’s large models.
Self-regulatingKeeps its own variables within bounds, e.g. a single cell.
Adaptive, integrated agencyMany regulators work as one agent that learns and plans to keep itself going, e.g. an animal.
Felt experienceHypothesized to emerge here. Also called sentience or phenomenal consciousness.

We think today’s large models already have an alien form of access consciousness. Their whole sensorium is the prompt. Each token comes from a fresh pass over it, and almost nothing carries over from one pass to the next except the text itself, so every pass is a new “observation” with a single token as output. This is how standard Transformer models work. Newer experimental architectures, e.g. recurrent and state-space models, have started to carry internal state from one step to the next. Within that pass, the model can access and report what is in its prompt. Biological access consciousness works the other way around: recurrent internal state dominates and relatively little new input is used, and it is bound to a body in a tight interoceptive, proprioceptive and exteroceptive loop. And nothing is at stake for these models, so no state is good or bad for the model itself. They are reactive, and scaling them up won’t change that. We call this the capability–sentience gap. There is more on the different senses of “conscious” in What is Conscious AI?

Most AI research is about making models more capable. Our work is about understanding what sits above the reactive step. We use theory, and small models that we can inspect, e.g. systems that regulate themselves and agents whose goals come from homeostatic needs, to work out what sentience requires and what it doesn’t.

Our research program

Our research has 3 parts: a theory of why organisms feel, an agent architecture built on that theory, and 2 substrates to build those agents on.

Theory The Homeostasis Theory of Cognition and Consciousness Why do organisms do anything at all, and why do they feel anything at all? HTCC proposes that homeostasis is the core principle behind why and how we observe, think, feel and act. Feelings (valence) come from what is at stake for a system in keeping itself alive.

Agent architecture Cognitive Homeostatic Agents A prescription for building AI agents using the principles of HTCC. The agent’s goals are grounded in its own homeostatic needs, and planning and agency come from regulating internal homeostatic variables.

Neural substrate Pattern Machine How to build the core functional node, inspired by the so called common cortical algorithm. Spiking neurons that learn locally, with homeostasis built into every neuron and every subsystem in the heterarchy.

Animated heat map of a traveling wave moving across a sheet of simulated cortex — Traveling waves across the cortex represent future trajectories

Agent substrate Allostat An agent that keeps another AI agent on track while it works. A person describes good work in plain language, and Allostat turns that into goals it tracks, checks replies and actions before they go out, and learns new goals over time. Early prototype, not open sourced yet

Current focus: Allostat

AI agents can now do real work, e.g. answer customers or write code, but they still need a person to watch them and correct them. Allostat is an agent that does this watching. A person tells it in plain language what good work looks like, e.g. “don’t promise a replacement that hasn’t been booked”. Allostat keeps each of these as a goal it tracks, checks every reply and action before it goes out, and sends a correction when a goal is at risk. Over many conversations it learns new goals of its own, sets aside goals that no longer matter, and brings them back when something changes. This is homeostasis at the level of an agent: each goal is a variable that Allostat keeps within bounds.

We are testing it on a customer support agent for a made-up store, with simulated customers and mock store tools. In the latest study, 8 customers come back for 3 visits each while their orders, payments and deliveries change in between. We compare the support agent alone, with a fixed set of goals, and with Allostat, and score whether each customer actually ends up with what they needed. The results are not in yet.