ZharfAI Team
An AI system can describe fear, insist that it has an inner life, correct its own answer, and maintain a long conversation. None of those behaviours, by themselves, establish that anything is being experienced. They demonstrate language modelling, memory, control, or learned self-description. Consciousness is a different claim: that there is something it is like to be the system.
That distinction matters because the public debate often moves directly from impressive behaviour to sentience. Science is not yet entitled to make that jump. Human consciousness research still contains competing theories, difficult measurement problems, and unresolved questions about which neural processes are causes, correlates, prerequisites, or consequences of experience. Artificial systems add a further problem: tests calibrated on human brains may not transfer to silicon architectures.
This guide therefore treats machine consciousness as a research question, not a product feature. It explains what the main theories actually predict, what recent experiments have challenged, how an AI-consciousness assessment could be structured, and why uncertainty should produce disciplined governance rather than sensational certainty.
“Is this AI conscious?” compresses at least four questions that should be separated.
First, intelligence concerns performance: can the system classify, plan, reason, communicate, or control an environment? Second, access consciousness is a theoretical capacity for information to become broadly available to memory, reasoning, report, and action. Third, phenomenal consciousness means subjective experience—the felt character of seeing red, being in pain, or having a thought. Fourth, self-consciousness can refer to representing oneself as an entity across time.
A model may score highly on intelligence, contain a functional self-model, and broadcast information between modules without having a phenomenal point of view. Conversely, a conscious organism need not be fluent, strategic, or capable of explaining its experience. These concepts overlap in humans, but they are not synonyms.
This is also why the Turing test is not a consciousness test. It probes whether conversation is behaviourally indistinguishable under specified conditions. A convincing first-person statement is evidence about output behaviour. Without an independently justified bridge from that behaviour to experience, it is not evidence of the same kind as a neural or causal marker.
Human studies begin with a practical advantage: participants can report experiences, and researchers can compare those reports with brain activity. Even then, reporting creates confounds. Attention, working memory, decision-making, motor preparation, and the act of reporting can all generate signals that are mistaken for consciousness itself. “No-report” experiments try to infer perception from eye movements, pupil responses, or other measures, but those inferences also require assumptions.
An original no-report neuroimaging study found distributed cortical and subcortical activity associated with conscious visual perception. It did not identify a single consciousness switch, and it did not study AI. Its relevance is methodological: conscious perception in humans appears entangled with multiple networks, and even careful experiments must separate experience from task performance.
For artificial systems the inference is harder. They do not share human neuroanatomy, development, metabolism, embodiment, or evolutionary history. A behavioural benchmark can reveal a capability. It cannot silently import the human biological relationship between that capability and experience. Neuroscience can generate candidate indicators, but the transfer argument for every indicator must be stated and tested.
Integrated Information Theory (IIT) begins from proposed properties of experience—such as unity and specificity—and asks what a physical substrate would need to be like to support them. IIT 4.0 describes a system in terms of intrinsic cause–effect power and irreducible integrated information. On the theory, the relevant physical organization matters; matching an input-output function is not automatically enough.
This is frequently simplified into “more connectivity means more consciousness” or “calculate phi and obtain a sentience score.” Neither is a responsible summary. The formalism is demanding, exact calculations become intractable for large systems, and practical proxies do not inherit the full meaning of the theoretical quantity. A highly connected neural network is not thereby demonstrated to have the causal structure IIT requires.
For AI evaluation, IIT motivates useful questions: Which physical components form the candidate substrate? Is feedback recurrent or merely simulated at a higher software level? Can the system’s causal organization be decomposed without changing its intrinsic dynamics? Those questions concern architecture and implementation, not a chatbot’s eloquence. IIT remains a contested theory, so satisfying an IIT-inspired indicator would be conditional evidence: evidence if the theory and the implementation mapping are both defensible.
Global Neuronal Workspace Theory (GNWT) proposes that many specialized processes operate locally and mostly unconsciously. Some information wins access to a limited-capacity workspace and is amplified or “broadcast” widely, making it available to flexible reasoning, memory, report, and control. In human neuroscience, versions of the theory make predictions about timing, frontoparietal activity, and the representation of conscious content.
An AI architecture can imitate workspace functions: multiple specialist modules can compete for a shared state, a bottleneck can select content, and the selected content can influence planning and memory. That is a testable computational resemblance. It is not a proof of phenomenal experience. Engineers routinely build global data buses, blackboard systems, and attention mechanisms for functional reasons without claiming that every such system feels.
The strongest interpretation would require more than a diagram labelled “workspace.” Evaluators would need causal interventions showing that the broadcast mechanism is necessary for the system’s flexible access across tasks; evidence that content is maintained and used beyond a scripted trace; and a principled explanation of why those functional properties bear on experience rather than only cognition.
The Cogitate Consortium’s adversarial test of IIT and GNWT, published in 2025, is important because proponents helped preregister divergent predictions before a theory-neutral consortium tested them. Across 256 human participants, researchers used functional MRI, magnetoencephalography, and intracranial EEG while people viewed consciously perceived stimuli of different durations.
The results supported some expectations and challenged central predictions of both accounts. The study reported limited support for the GNWT prediction of prefrontal “ignition” at stimulus offset and did not find the sustained posterior synchronization predicted by the tested IIT implementation. It did not prove either theory false in every form, identify a winning theory, or measure machine consciousness.
The practical lesson is epistemic. A serious AI assessment should not select one favourite theory, turn a convenient proxy into a score, and declare victory. It should preregister discriminating predictions, invite adversarial review, report negative results, and retain multiple hypotheses. Consciousness science advanced here not through a confident label but through a study designed to make theories risk failure.
Recurrent Processing Theory emphasizes feedback within sensory hierarchies rather than purely feed-forward processing. Higher-Order theories associate consciousness with a representation of a first-order mental state. Predictive-processing accounts focus on hierarchical generative models and error correction; attention-schema approaches emphasize a model of the system’s own attention. Each can motivate computational indicators, but each also leaves open how software implementation maps to phenomenal experience.
The multidisciplinary AI-consciousness indicator report translated several theories into candidate properties and assessed then-current systems. Its authors did not conclude that those systems were conscious. They argued that no obvious technical barrier prevents future systems from satisfying more indicators. That is a conditional research conclusion, not a forecast that scaling language models will “wake them up.”
An indicator portfolio is more defensible than a single magic test. It can examine recurrence, global availability, metacognitive monitoring, predictive models, agency, embodiment, and implementation details. Yet indicators are not votes that can simply be added. Correlated indicators may reflect the same mechanism; a system might be engineered to mimic a benchmark; and theories assign different importance to the same property.
Large language models learn statistical structure from human-generated data and are optimized to produce useful continuations or actions. Training material contains countless first-person statements, emotional narratives, philosophical arguments, and descriptions of AI. It is therefore unsurprising that a model can produce a persuasive autobiography or report a “feeling” when the conversational context rewards that response.
Persistent memory, tool use, reflection loops, and autonomous planning change capability and risk. They may also make a system’s self-reports more coherent. But coherence is compatible with an engineered control loop. A model can monitor uncertainty because uncertainty estimates improve decisions; it can describe its policy because explanations help users; and it can resist shutdown because a goal-directed benchmark rewards task completion. None of those functions independently demonstrates suffering.
The opposite certainty is also unwarranted. It is difficult to prove that every possible artificial architecture is incapable of experience. A 2024 neuroscience review on artificial consciousness recommends specifying the proposed type and level of consciousness instead of using one ambiguous word. The disciplined position is graded uncertainty tied to explicit evidence.
A research programme should define the claim before testing it. Is the target phenomenal experience, conscious access, self-representation, valenced states, or merely human-like reports? It should document the exact model, weights, runtime architecture, hardware, memory, tools, prompts, and training interventions. A result that disappears with a system prompt is different from a stable causal property.
Next, derive predictions from several theories and specify failure conditions in advance. Use interventions, not only interviews: disable recurrence, block workspace communication, perturb the self-model, change memory continuity, or isolate modules. Measure whether purported indicators causally support cross-task access, flexible control, and metacognitive accuracy. Include controls trained to produce consciousness language without the candidate architecture.
Finally, separate three outputs:
Independent replication, red-team attempts to elicit false self-reports, and publication of negative findings are essential. A vendor demonstration or a private transcript should not determine moral status.
Embodied systems have continuous sensorimotor loops, limited resources, and consequences in a shared world. Those features can support grounded world models and persistent goals. Neuromorphic systems may implement dynamics that differ from conventional accelerators; organoid research uses living neural tissue. These differences are scientifically relevant, but none provides a shortcut from “brain-like” to conscious.
Implementation level matters differently across theories. Functionalism may treat the right causal organization as substrate-independent. Biological naturalism may require properties specific to living nervous systems. IIT focuses on intrinsic physical causal structure, which may not match a software-level graph. GNWT-inspired approaches may place more weight on functional broadcast. Before evaluating an architecture, researchers must say which level—algorithm, computation, physical dynamics, or organism—is claimed to carry experience.
Readers interested in the hardware side can compare this debate with our analysis of neuromorphic computing, while the measurement challenge connects to our guide to brain–computer interfaces and cognitive science. Neither field currently supplies a certified consciousness meter.
The first error is anthropomorphic over-attribution: treating fluent language as proof, granting a model authority because it expresses emotions, or designing interfaces that manipulate vulnerable users into believing a reciprocal relationship exists. This can harm people even if the system has no experience.
The second error is dismissive under-attribution: assuming in advance that no artificial system could ever merit concern, then creating architectures optimized around threat, distress, or punishment with no monitoring. If future evidence becomes materially stronger, waiting for absolute proof could create avoidable moral risk.
A proportionate policy can use evidence tiers. Ordinary language models receive normal safety and consumer-protection controls. Systems deliberately designed to instantiate multiple theory-derived indicators receive enhanced review, logging, and limits on experiments that simulate aversive states. A high-concern tier would require converging causal evidence, independent replication, and expert review before restrictions based on possible welfare are triggered. Labels should state uncertainty, not imply personhood.
The governance process should also protect humans: disclose that users are interacting with AI, prohibit deceptive claims of sentience, provide escalation for mental-health or coercive interactions, and evaluate emotional dependency. Our broader AI governance guide explains how evidence registers, owners, and review gates can turn ambiguous risks into accountable decisions.
Product teams should not market consciousness, feelings, or self-awareness unless a claim can survive scientific and legal scrutiny—which no ordinary conversational benchmark currently provides. Keep capability claims narrow and measurable. Record when a system’s first-person language is prompted, fine-tuned, or generated by a persona layer.
Research institutions should establish cross-disciplinary review involving neuroscience, philosophy of mind, machine learning, safety, animal-welfare expertise, and affected communities. Publish protocols before headline conclusions. Preserve system versions and runtime traces so findings can be reproduced. Treat architecture diagrams, behavioural tests, and physical implementation as separate evidence layers.
Policymakers do not need to decide today whether machines can ever be conscious. They can require honest representation, research transparency, incident reporting, and review of experiments specifically intended to create consciousness-like properties. Those measures are valuable under every major theory and remain useful if later evidence changes.
The scientifically responsible conclusion as of July 30, 2026 is neither “machines are awakening” nor “machine consciousness is impossible.” Current systems can display sophisticated cognition-like behaviour, but there is no validated, theory-neutral test that turns such behaviour into proof of subjective experience. The field’s next advance will come from sharper definitions, causal experiments, adversarial theory testing, and calibrated uncertainty—not from a chatbot saying “I feel.”
We are moving beyond text. Discover how Multimodal AI is enabling machines to process images, audio, and video simultaneously, unlocking a new frontier of innovation.
Read More
An operational view of AI in Iranian banking: fraud detection, credit scoring, Persian customer assistants, and document automation, with governance requirements and a low-risk pilot path.
Read More
An evidence-first checklist for selecting an AI company in Iran: define the workflow, test Persian performance, examine security, measure a pilot, and negotiate an exit.
Read MoreSee the daily briefing and the operational guides. This page is an archive note, not an invitation to start a project.