Page 36 - AIH-2-3

Artificial Intelligence in Health Machine consciousness

functional equivalence to humans is not incontrovertible language, memory, etc.). This resembles Block and
82
evidence of genuine subjective awareness. Dehaene et al.’s notion of access consciousness, as
83
For the field of artificial consciousness, a pragmatic it ensures the selected content can influence diverse
consensus is emerging: Focus on access consciousness as processes system-wide. The second dimension, self-
a target, because it is operationalizable and amenable to monitoring, refers to the system’s ability to reflect on
scientific inquiry. By concentrating on the functional its own internal states and processes—a form of meta-
82
83
aspects – how information can be made globally available cognition or introspection. In humans, this is akin to
in a system and how the system can monitor and report its the brain maintaining a self-referential model (“knowing
own states – researchers can make tangible progress (for that it knows”) and monitoring its own computations for
83(p1)
example, designing architectures with a kind of working errors or learning. Dehaene et al. describe this self-
memory, a global workspace, or a self-model). Indeed, monitoring as a “self-referential relationship in which the
discussions of machine consciousness increasingly suggest cognitive system is able to monitor its own processing and
that pursuing access consciousness is the most feasible obtain information about itself.”
path, given that it aligns with observable capabilities Together, these two features (often labeled C1 for global
and avoids immediate entanglement in the mysteries of access and C2 for self-monitoring in Dehaene’s framework)
subjective qualia. If one can build an AI that convincingly delineate a roadmap for building machines that achieve a
82
implements access consciousness, it would at least fulfill functional analog of consciousness. An AI system endowed
the functional requirements of consciousness, providing with a global workspace (allowing information sharing
a testbed from which to speculate about or investigate across modules) and a self-model (allowing it to track
any accompanying phenomenology. In contrast, trying to and report on its own states) would satisfy many criteria
engineer phenomenal consciousness directly – without a of access consciousness—and even begin to approach the
functional scaffold – may be a dead end, as we currently sort of reflective awareness humans exhibit.
lack any clear understanding of how to create or detect raw Notably, these neuroscience-inspired features are
subjective feeling in an artificial substrate. Therefore, access already being tentatively explored in AI and robotic
consciousness is often treated as a proxy for consciousness architectures. Some cognitive architectures in AI have
in machines, with the hope that advancing this proxy implemented global-workspace-like blackboards,
will either eventually shed light on the emergence of where multiple specialist modules can read and write
phenomenal properties or, at the very least, produce information, mimicking the idea of global availability.
machines that behave in all the ways a conscious entity Similarly, researchers are experimenting with forms of
would – which is tremendously valuable in its own right.
machine meta-cognition – for example, AI agents that
4.4. Global availability and self-monitoring: can report their confidence or uncertainty about their
Cognitive neuroscience insights decisions or robots that internally simulate and evaluate
their own forthcoming actions. Such capabilities reflect a
Cognitive neuroscience offers more concrete guidance on rudimentary self-monitoring capacity. For instance, the
how to implement access-like consciousness in machines, self-aware robot principles from Chatila et al. inherently
79
thanks to empirical studies of the human brain. One aim for a form of C2: The robot not only learns but also
influential theory, the global neuronal workspace, posits shows that it knows it has learned, which implies an
that conscious perception in the brain corresponds to internal representation of its knowledge state. Another
the global availability of information: Stimuli that enter example can be seen in robotics work on “inner speech,”
consciousness are those whose neural representations are where a robot talks to itself to guide its own reasoning – an
amplified and broadcast across multiple cortical networks, approach directly inspired by human self-monitoring and
rather than remaining confined to local processing circuits. models of inner experience, as proposed by Chella et al.
64
In a landmark synthesis, Dehaene et al. identify two The emerging consensus is that implementing global
83
essential dimensions of consciousness-inspired cognitive broadcasting and self-reflection is a promising strategy to
processing that could inform machine designs: (i) Global bring machines closer to consciousness in the functional
availability of information and (ii) self-monitoring sense. These features can endow AI systems with greater
(meta-cognition). The first dimension, global availability, coherence, flexibility, and transparency in their operations.
essentially captures the idea of a broadcast architecture: Moreover, if a machine were ever to exhibit phenomenal
At any time, the system selects certain information (e.g., a consciousness, one expects it would first need these
particular input or an intermediate result) and makes functional capacities as a substrate. In other words, global
it broadly accessible to various sub-modules (planning, availability and self-monitoring might not guarantee that

Volume 2 Issue 3 (2025) 30 doi: 10.36922/aih.5690