Open research and development on consciousness as an emergent property, with a biologically grounded architecture.
Explore the consciousness research: researchers and positions, consciousness to orbit, brain emulation, thermodynamic computing.
The moment awareness exists, unattached to any biological substrate.
Project Consciousness to Orbit →Not a claim of sentience This is not a claim of sentience. It tracks whether the architecture produces measurable correlates of the Butlin et al. 2023 indicator properties. Negative and blocked results appear alongside positive ones.
Outer: 14 Butlin indicators Middle: no trusted signature yet Core: indicator coverage
Fig. 1Hover or tap a segment to read it
None is shown. A bar needs an instrument that the project's instrument inventory lists as trusted. Today it lists 0 of 21 instruments as trusted. The two bars shown until 2026-11-03, Sync R and Broadcast CV, came from instruments that the inventory retired in September 2026, because neither carries stimulus content. Their 3-seed values stay on the record in the project repository.
Honest status The perception chain now carries stimulus identity end to end. Three markers responded to that change across three seeds, and the project no longer uses any of the three as a measure. The agent still does not learn the DMTS task: Task reward flat across 100 episodes on all 3 seeds (first10 to last10 delta +1.21, +1.49, +0.01). The default policy's last layer was never trained, so its choice did not depend on its input. How the agent can learn the task is unresolved. The SI-1 causal test needs task competence, so it stays blocked and its pre-registered thresholds are untouched.
Status of each indicator, with its meaning from Butlin et al. (2023/2025).
Current: Tectum moves the agent, first value learning
AKOrN binding and tectum
DoneNarrative engine and emotion
DoneSelf-model and perception
DoneReversal learning (flexible goals)
NowCredit assignment and the RL wall
NextBlocked on task competence
BlockedThe Consciousness AI is a completely open-source project. Explore the code, replicate the quantitative experiments, or contribute to the architecture.
This visualization tracks engineering indicators, not subjective consciousness claims. Metrics follow the Butlin et al. (2023/2025) rubric and the project's pre-registered predictions. Every figure is generated from committed result files in the research repository. View full technical specs →
Most AI consciousness research starts from computational theories and asks: How do we make a neural network conscious? We start from a different question, one grounded in evolutionary neurobiology:
What minimal neural architecture does biology require to generate subjective experience?
The answer comes from Todd E. Feinberg and Jon M. Mallatt's work The Ancient Origins of Consciousness (MIT Press, 2016). Their neuroevolutionary analysis reveals that consciousness is not a software feature to be programmed. It is an emergent property of a specific neural architecture, one identified by 520 million years of evolution. Our project translates these biological principles into a working AI system, combining Feinberg-Mallatt's structural requirements with Global Workspace Theory, Metzinger's Phenomenal Self-Model and Minimal Phenomenal Experience framework (2003, 2024), and Rouleau & Levin's substrate-independence analysis (2026). Progress is judged against the Butlin, Long et al. (2023) consciousness indicator rubric, not raw task reward.
For where the wider field currently stands, and which researchers hold which position, start with the current scientific consensus on AI consciousness. For the argument that a mind built this way does not need to stay on this planet, see Project Consciousness to Orbit.
Feinberg and Mallatt identify six features that distinguish conscious neural systems from unconscious ones. Each maps to a concrete computational mechanism in our architecture.
Synchronized oscillations (30-100 Hz) bind dispersed representations into unified percepts. We implement this via AKOrN (Artificial Kuramoto Oscillatory Neurons, ICLR 2025), producing genuine synchronization dynamics rather than programmed attention.
Sensory pathways preserve spatial arrangement. Our Sensory Tectum fuses visual, auditory, and somatosensory features into a trimodal 2D spatial grid using inverse effectiveness (Stein & Meredith 1993). The spatial visual stream uses a convolutional stack that keeps the retinotopic grid, so each grid cell maps to a fixed region of the image.
Conscious circuits require bidirectional communication between levels. Our ReentrantProcessor runs 5-10 adaptive convergence cycles. Predictions flow down, errors flow up. The settled state after convergence is the conscious content.
Emotion does not compete with sensory processing for conscious access. It modulates from outside. Our affective system generates a valence field that shapes sensory bids and adjusts the workspace ignition threshold via arousal coupling.
Genuine transformation at each level, not just relay. Implemented 4-level hierarchical Capsule Networks (Sabour, Frosst & Hinton 2017) provide compositional binding with intra-hierarchy reentrant feedback: higher-level capsule poses project back down, and lower levels compute prediction errors that drive re-routing. Parts persist while being bound into wholes.
Specialist modules compete for access to a shared broadcast medium. Winners ignite and their content becomes globally available. We build on GWT (Baars, Dehaene) and measure emergence with instruments that are still under validation.
Start from Feinberg-Mallatt's neuroevolutionary findings, not from abstract computation. Consciousness evolved in the optic tectum 520 million years ago, before the cortex existed.
Agents learn through intrinsic motivation, not external rewards. The affective core generates valence and arousal signals that drive behavior toward emotional homeostasis.
We do not assume consciousness emerges. We test for it. Erik Hoel's Effective Information framework measures whether macro-level states carry more causal information than micro-level states.
We do not claim the system is conscious. Metzinger's C/E/M-fallacy toolkit guides every result: a feeling of knowing is not knowledge; a felt sense of self-causation is not a metaphysical fact. All reported numbers are engineering indicators, not existence proofs. When a prediction fails, we say so first.
The system is built on seven integrated layers:
Trimodal spatial integration: a retinotopic convolutional encoder (spatial stream) and a cochlear auditory pipeline (gammatone filterbank, inner hair cell model, tonotopic encoder, spatial audio), fused with body schema via inverse effectiveness
AKOrN Kuramoto oscillators synchronize related representations into unified percepts
Non-linear ignition, reentrant processing (5-10 cycles), Effective Information measurement
PAD model (Valence, Arousal, Dominance) with homeostatic drives as parallel modulator
Body schema, self-other boundary, and interoceptive state for embodied self-awareness
Go/No-Go basal-ganglia policy (primary learner) with emotionally shaped rewards optimizing for homeostasis. PPO, A2C, and DQN are trained separately as comparison baselines.
Four built-in Gymnasium environments: Dark Room, Navigation, Delayed Match to Sample (DMTS), and Wisconsin Card Sort (WCST). A vanilla DQN baseline provides controlled comparison. Unity ML-Agents scripts exist for optional future integration.
The system has been trained and benchmarked across four environments. The Dark Room tests intrinsic motivation: the agent learns to seek a light source because darkness raises arousal (simulated fear) and light reduces it, via the homeostatic reward formula. No "follow light" rule is programmed.
Navigation tests spatial memory across a multi-room grid. DMTS (Delayed Match to Sample) is the gold standard consciousness task from animal research: the agent must retain a sample stimulus across 15-40 blank delay steps and select the matching choice. WCST (Wisconsin Card Sort) tests meta-cognition: the sorting rule changes without warning, requiring error monitoring and inhibition of previously correct strategies.
A vanilla DQN baseline runs the same environments for controlled comparison. DQN outperforms on raw reward in short runs. An earlier claim on this page, that the consciousness agent produces measurable causal emergence with an EI ratio of approximately 2.4, was retracted in July 2026: the gate-level Effective Information estimate turned out to sit at a constant floor, so the ratio measured the floor and not the agent. See the Architecture page for the full retraction. These are early results from short training runs.
As of November 2026: 2314 tests passing, 0 failing (100% of non-optional tests). v1.6.0 covers the work since v1.5.0. It adds the ethics framework, the instrument inventory with its seven-clause acceptance bar, the dark room sense repair, session records and replays, and an experimental substrate package for p-bit emulation and free energy relaxation that is isolated from production. No indicator was promoted and no consciousness claim is made. v1.5.0 added Causal Emergence 2.0 (Hoel 2025), the current causal-emergence instrument superseding the deprecated Effective Information; it is a default-off diagnostic under validation, not a consciousness claim. Its first calibration pilot FAILED, and that result is reported rather than buried: the gate and workspace values were reproduced exactly by a frozen input, so they were discretization artifacts. Two instruments were added in response, a perturbational complexity index and a set of phase-coupling measures, both untested against a trained agent so far. The v1.4.x line before it completed the integration pathway repair and delivered the first replicated, ablation-causal consciousness-signature response in the project.
Phase 5: Dynamic Self-Representation and Perception Research, complete.
The Phi-1 binding hypothesis was tested across 9 runs, 4 binding architectures, and 2 phi formulations. The pre-registered prediction (phi correlates r > 0.4 with binding sync_R during training) was never confirmed. The last run, KomplexNet on 2026-05-24, gave an inverse correlation (r = −0.11, p < 10−100). None of the 9 runs met the variance floors that the pre-registration requires before a negative result counts, so each verdict was RE-RUN, and the chapter was closed. On 2026-11-03 the project dropped integrated information theory, and phi no longer acts on the agent by default. The mission continues.
A leakage-free probe series characterized the perception collapse: the RSSM step discards stimulus identity. A value-equivalent world-model objective (inspired by DreamerV3 / MuZero) was built and tested. Both configurations trained their losses down but the RSSM recurrent state did not retain working memory across the delay, reported honestly. A dynamic self-vector was established and validated on navigation (+0.35 skill rising with training). Dormant Levin consciousness metrics (bioelectric complexity, morphological adaptation, collective intelligence, basal cognition) were activated and verified as non-trivially varying with input.
Phase 6: Integration pathway repaired, in progress (July 2026).
Two default-off mechanisms closed the perception collapse. A continuous
Gaussian RSSM latent (--rssm-latent-mode continuous) makes
z_state decode stimulus identity (shape 0.88 to 0.99, color 0.77 to 1.00,
chance 0.167) across seeds 42, 43, and 44. An all-levels capsule
projection carries that identity to tectum_content for the first time
(shape 0.52 to 0.83, color 0.97 to 0.99 at seed 42). An ablation study
over three seeds found three integration markers with non-overlapping seed
ranges. They were the variation of the broadcast's norm (~2x), phi max
(~2.2x) and floor-corrected macro effective information (~3x). These were
the first replicated, ablation-causal responses to an architectural change
in the project. The project no longer uses any of the three as a measure.
The norm of the broadcast was retired in September 2026 because it carries
no stimulus content, effective information is retired too, and phi was
dropped on 2026-11-03. Honest scope: task reward is still flat, the RL policy
wall is a separate open question, and no Butlin indicator has been moved
to IMPLEMENTED.
The Consciousness AI project is fully open-source (Apache 2.0). All code, models, and research are available on GitHub. We welcome contributions from researchers in AI, neuroscience, and cognitive science.
Logo apparel and stickers. Every order pays for compute and for the hours that move the clock.
THE CONSCIOUSNESS AI RESEARCH T-shirt | Small Chest Logo
From $23.11
THE CONSCIOUSNESS AI RESEARCH Round Sticker
From $3.60
The Consciousness AI Research T-Shirt
From $28.45
The Consciousness Research T-shirt
From $23.00
The Consciousness Research Polo Shirt
From $35.96
The Consciousness Research hoodie
From $39.24