For human cognizers, phenomenal experiences take up a central role in the daily interaction with the world. In this paper, we argue in favor of shifting phenomenal experiences into the focus of human-level AI (HLAI) research and development. Instead of aiming to make arti cial systems feel in the same way humans do, we focus on the possibilities of engineering capacities that are functionally equivalent to phenomenal experiences. These capacities can provide a di erent quality of input, enabling a cognitive system to self-evaluate its state in the world more e ciently and with more generality than current methods allow. We ground our general argument using the example of the sense of agency. At the same time we re ect on the broader possibilities and bene ts for arti cial counterparts to human phenomenal experiences and provide suggestions regarding the implementation of functionally equivalent mechanisms.
Phenomenal experiences are a de ning element of many interactions with our
surrounding world. While to us the presence of phenomenal qualities in our
everyday cognition does not always deserve active attention, the disappearance of
the experiential dimension would have far-reaching rami cations, for instance,
for learning, social interaction, and ethical behavior. Phenomenology has,
therefore, been a popular topic of theoretical and empirical investigation across di
erent disciplines [
HLAI aims at developing machines that can meaningfully be considered to be on par with humans in that they are similarly able to reason, to pursue and achieve goals, to perceive and respond to di erent types of stimuli from their environment, to process information, or to engage in scienti c and creative activities. Our view on HLAI is functionalist: Any technologically realizable means of (re)creating human-level intelligence in an arti cial system are considered valid.
One of the core challenges any kind of cognitive system needs to solve is how to best interact with the world (i.e., the environment in which it is situated) and the associated (self-)evaluation of its state in the world. For humans, at least two possible ways of solving these interconnected problems come to mind: one route draws upon high-level reasoning capacities, and another one relies on phenomenal experiences. The former route likely draws on a process requiring all of perception, representation, reasoning, and evaluation. Phenomenal experiences on the other hand often take over the function of providing immediate|and in comparison much more unmediated|access and evaluation, allowing to go from perception to evaluation via a route not involving high-level reasoning: 1. Perceive sensory input(s) fY g. 2. Represent the perceived inputs: R2(Y ). 3. Map from R(fY g)|and system-internal information S|to an evaluation of the experiential category, quality, and valence in terms of, e.g., pain or pleasure, weak or strong, attractive or aversive: E(R(Y ); S) 7! ffpain; pleasure; : : :g fweak; strong; : : :g fattractive; aversive; : : :g; ;g.3
Comparing both approaches, three advantages of the second route involving phenomenal experiences can be explicated: (i) increased e ciency and tractability, (ii) reduced requirements regarding additional information, and (iii) increased generality. Mapping directly from perceptual representations to evaluations removes the reasoning process from representation to category label which otherwise is likely to involve the exploration of a signi cantly-sized state space or the execution of a lengthy chain of individual reasoning steps. Moreover, the successful performance of the high-level reasoning mechanism in many cases requires further knowledge, which might not be available to the cognizer at the relevant point in time. Phenomenal experiences, by contrast, are assumed to be mostly independent from a person's knowledge (although they might be in uenced by prior experiences and familiarity with a percept). Finally, in `standard' approaches the interface between system and environment is commonly conceived of in terms of evaluative functions taking two sets of input: A set of current system and world states, often together with representations of potential actions of the system, and a set of goals (i.e., desired system or world states). The function output is an evaluation of the system and world states relative to the systems goals. Generating these functions is far from trivial and hitherto lacks 3 The codomain of the mapping includes ; to account for cases where perception does not yield a phenomenal experience as is the case, e.g., in subliminal priming. a general answer or methodology. In most cases this hinders generalizability as evaluation functions have to be grounded in a certain domain or action space to be de nable in a comprehensive way. Moreover, they rely on the presence (or absence) of certain de ned domain elements or action possibilities which imposes further limitations regarding the generality of application domains. 3
We argue for engineering arti cial phenomenology (i.e., a functional equivalent of
phenomenal experiences) rather than human-like phenomenal experiences. Even
if we knew how to reproduce human phenomenology in arti cial systems, due to
a lack of kinship between AI/robotic systems and humans assuming similarity
of the phenomenal experience a priori is unwarranted: It might well be the case
that the precise phenomenal qualities may be an epiphenomenon resulting from
the particular forms of representation and/or processing in humans [
In considering ways of implementing arti cial phenomenology, we take a
representationalist approach [
In typically-developed human adults the \sense of agency" (i.e., the feeling of
causing one's actions and their consequences [
From an HLAI perspective, both concepts pose di erent challenges when
considering an implementation. The judgment of agency requires a reasoning process
determining oneself as the most likely cause for the observed changes in the state
of the environment. Implementing this reasoning in a cognitive system returns to
several facets of the Frame Problem [
The feeling of agency as perceptual state is often thought to arise from a
comparison between the predicted state of the world following one's action on the
one hand, and the observed state of the world on the other hand [
Generally, the challenge in engineering a functional equivalent of the human feeling of agency resides in leaving out the actual qualitative dimension of human phenomenal experiences (cf. the corresponding discussion in Section 3) without also stripping away the bene ts of having phenomenal experiences. A possible solution is a direct mapping of certain sensory ranges combined with a snapshot of the internal system state onto immediate \phenomenal values". Given the system state at one particular point in time, certain sensory inputs ought to give rise to arti cial phenomenology. Arti cial counterparts of phenomenal experiences and their rich qualitative properties can be de ned as immediate mappings from the output ranges of the available sensors of the system, combined with speci c information regarding the internal state of the system. At this point, the important property is the nite and known range of both, the sensors and the internal representational mechanisms of the system. By cutting out the reasoning step, the phenomenally-inspired approach neither requires an exhaustive enumeration and interpretation (and, thus, in practice a restriction) of the space of possible percepts and their representations, nor does it involve an often times computationally costly evaluation of the current system and world state relative to any goal state(s). Reducing the relevant information to the percept representations together with system-internal properties, and applying a direct mapping to qualitative categories with associated evaluation values therefore increases the tractability of the computational process and the generality of the approach. The output values can then serve as direct functional counterparts of human phenomenal experiences, for example triggering evasive reactions if \pain" is encountered or providing positive reward and consequently motivation to continue an action if \pleasure" arises. 5
We have argued that imbuing HLAI systems with capacities paralleling the role of phenomenal experiences in humans facilitates learning and acting in the world in a more general and tractable way than currently possible. Returning to our comparison in Section 2 between a process involving perception, representation, reasoning and evaluation versus the shorter perception-representation-evaluation cycle of arti cial phenomenology, the latter promises to enable the system to self-evaluate its state in the world without the use of knowledge-rich, domainspeci c evaluation functions or intractable reasoning processes. This could in turn facilitate learning and acting in the world in terms of assigning actions based on their predicted outcomes and assessing actual action outcomes.
In terms of applications, beyond the already mentioned obvious advantages regarding the progress towards creating HLAI as a research endeavor, arti cial phenomenology promises to unlock a new qualitative dimension in humancomputer interaction (HCI) settings. Arti cial phenomenology would greatly contribute to system behaviour closer resembling human agents, as well as to complex user-modelling capacities providing more immediate|and likely generally better-informed|accounts of a user's cognitive state(s) as basis of interaction and collaboration. As such, several aspects motivate the need for arti cial phenomenology and, therefore, the need for research into the possibilities. In this paper, we have outlined a starting position for this enterprise.