NOTE-991: Apps & Tsakiris — The free-energy self

Contribution

Before this paper, the self-recognition literature (mirror and self-face tasks, self-voice, body ownership) had findings and no theory tied to a general account of cortical function, and the bodily illusions had a separate explanation each. The paper puts all of it under the free-energy principle and predictive coding. Self-recognition becomes ordinary hierarchical inference about the causes of sensory input, the self is the cause that best explains correlated input across the senses, and the rubber hand, enfacement and rubber-voice illusions become one effect. It also makes motor signals one kind of prediction among others, so that self-recognition does not need action.

Key insight

“Me” is a hypothesis, not a label fixed in advance. One’s own body is the object that, alone among objects, is reliably felt when it is seen touched, moves when one predicts movement, and so on. A brain that minimises prediction error learns that it is the single best explanation for all these correlations. Because it is a probabilistic belief, it can be revised: give another hand or face the same correlations, and some of the probability moves to it, and away from one’s own.

Assumptions

• The free-energy principle and predictive coding are taken as given. “The aim of this paper is not to discuss the validity or evidence supporting the free-energy principle and predictive coding”; the theory “is still in its infancy and therefore is neither largely supported or refuted as yet” (§3). Its mathematics is not given.
• Two classes of neurons. Representational units carry predictions and error units carry prediction errors; unexplained error goes up the hierarchy and predictions come down to “explain away” error (§3).
• Empirical Bayes in a hierarchy. Beliefs at intermediate levels are “empirical priors”, optimised by both top-down and bottom-up influences; high-level ones are “abstract multimodal beliefs about states of the world that change slowly over time”, learned from associations between congruent sensory events (§4).
• Motor signals are predictions. Under active inference, “cortical signals that drive movement are descending predictions about the proprioceptive consequences of movement”, better called corollary discharge than efference copy; corollary discharge “is just like any other descending prediction” except that it elicits movement (§4).
• Self-recognition tasks stand in for self-awareness. Recognising one’s own face or voice is taken, after Gallup, as a “behavioural marker” of self-awareness (§1). The paper is about recognising the physical self, not about the self as subject.

Key results

There are no new data, models or simulations. The paper’s claims, by section:

• The self as a probabilistic prior (§4). “One’s own body is the one which has the highest probability of being ‘me’ as other objects are probabilistically less likely to evoke the same sensory inputs.” It is recognised as “self” only when streams are integrated: a self-face is “a face, that I have seen before, that I am familiar with, and that is associated with congruent corollary discharge, vestibular, somatosensory and interoceptive information when seen on a reflective surface”. Touch seen on the skin and felt becomes associated, while touch seen between two other objects never evokes a felt touch. “In short, the notion that there is a ‘self’ is the most parsimonious and accurate explanation for sensory inputs.” The authors add that this “parsimonious accuracy” is “exactly the quantity that is optimised when minimising free energy”. To issue parallel predictions to several senses, “there must be a high level representation (of self)”.
• The mirror (§4). An infant matches felt with seen sensorimotor behaviour and so learns what she looks like. A reflection is surprising in space (allocentric), in time (it moves with one’s corollary discharge) and in form, and multisensory areas explain the surprise away. One’s face becomes “me” because “it is typically the face that is processed when looking in a mirror, and that face rarely violates expectations instantiated by the agent’s actions”.
• What sets the account apart (§4). Earlier accounts tied self-processing to one “self” network (Northoff et al. 2006) or to the match of motor efference with reafference alone (Legrand and Ruby 2009). Here “information from any system can be used to explain away information in any other system”, so “information from each and every sensory system” can modulate self-recognition.
• The illusions (§5.1). Reported from the cited studies:
  • Rubber hand illusion: synchronous touch on congruent locations of the hidden real hand and a seen rubber hand produces ownership of the rubber hand and a drift in felt hand position toward it (Botvinick and Cohen 1998; Tsakiris and Haggard 2005).
  • Full-body versions: Ehrsson (2007) and “Leggenhanger et al (2007)” [sic], with synchronous stroking of a seen virtual back making participants feel “as if the virtual body was their body”. These show multisensory input determining a minimal first-person perspective, self-location and self-identification, “three conditions that are critical for the experience of selfhood (Blanke and Metzinger, 2009)”.
  • Enfacement: touch felt on one’s face while watching another’s face touched in synchrony makes participants respond on self-other tasks “as if the others face was more like their own face” (Tsakiris 2008; Sforza et al. 2010), and report it as more like “me” (Tajadura-Jimenez et al. 2012).
  • Rubber voice: synchronous feedback of a stranger’s voice while speaking is reported as a distorted version of one’s own voice (Zheng et al. 2011, the only report).
• One explanation of the three (§§5.1–5.2, Fig. 1). In each illusion there is “considerable bottom up sensory surprise evoked in one system” (touch felt where it was not predicted, a felt touch coinciding with a seen one). It “will be explained away by top-down effects from multisensory areas”, and perceptual learning updates the representation of one’s body, face or voice. After the RHI, participants rated their hand and the rubber hand as more similar (Longo et al. 2009). Illusions are “Bayes optimal” inference under the prior “that most of our sensations are caused by ourselves”. The Figure 1 caption: before stimulation the rubber hand has “a low probability of being one’s own hand”; during the illusion “touch on the rubber hand is no longer surprising”.
• Ownership as competition (§5.2). “The driving effect for each illusion is an increase in the probability that the other object … will be represented as part of the body and a decrease in the probability that one’s actual body will be represented as ‘self’.” The decrease is read in reports that the real hand cools (Moseley et al. 2008; Hohwy and Paton 2010) and its histamine level rises (Barnsley et al. 2011), “consistent with the notion that the real limb is being partially rejected”. Which model wins is compared with binocular rivalry (Hohwy et al. 2008): competing generative models, with perception “a function of the relative fit”.
• Context as prior (§5.3). Masked and unmasked self-primes speed self-recognition in another domain. Self-face bias differs by culture: Chinese participants reversed it in a supervisor’s presence (Liew et al. 2011), and religious participants showed a weaker bias than atheists (Ma and Han 2012). The RHI is reduced or absent for a non-corporeal object, a hand of another skin tone, or a hand in an incongruent place. These are read as priors in force before the stimulus.
• Three anatomical predictions (§6).
  1. “No unimodal areas will exclusively be engaged by self-stimuli” (§6.1). Face, voice and body areas respond more to self-stimuli, but process all faces, voices and bodies; the extra response “may therefore reflect the surprise evoked by self-stimuli”.
  2. Multimodal areas will be engaged, without being self-specific (§6.2). TPJ, IPS, anterior insula and IFG are active in self- recognition and in the RHI and full-body illusions, are mutually connected, and “may comprise a core circuit which is engaged when recognising one’s self and creates a sense of ownership over one’s body”. Their connections show they also process other things. The anterior insula is called “primary interoceptive cortex”.
  3. Predicted self-stimuli will suppress activity (§6.3): one cannot tickle oneself, with reduced SI activity (Blakemore et al.); seeing one’s own face while touched reduces SI and IFG activity compared with another’s face (Cardini et al. 2011).
• Relation to Kilner and Friston (§6). Their free-energy account of the mirror-neuron system predicts others’ actions from simulated corollary discharge. This paper extends it to how a stimulus comes to be labelled “self”, including “in situations where actions are not a driving factor”.
• Caveats (§7). “There is, as yet, little direct evidence that predictions and prediction errors related to self-stimuli are processed anywhere in the brain”, though “there is also none that refute it”. The theory is “hard to falsify using standard neuroscientific methods”, because 3T BOLD mixes prediction and error populations, so that “a highly surprising event and a highly predicted event may … both evoke large responses”. Whether the self is represented probabilistically, not deterministically, has had no direct test. Proposed remedies: designs that make error and probability orthogonal, multivoxel analysis, laminar high-field fMRI, model-based behavioural tasks and dynamic causal modelling.
• Summary (§8). “Empirical data neither largely supports nor refutes our account of self-recognition.”

Claims

id	claim	strength	support
C1	One’s own body, face and voice are represented as a high-level multimodal prior, the most probable cause of correlated multisensory input	weak: theoretical; stated as following from the framework	§4; no direct test, as §7 grants
C2	The rubber hand, enfacement and rubber-voice illusions are one process: unimodal surprise from synchronous stimulation explained away by updating which body part, face or voice is “me”	weak to moderate: fits the cited findings; no rival explanation is shown to fail	§§5.1–5.2, Fig. 1
C3	During the illusion the probability that one’s own hand is “me” falls as the rubber hand’s rises	weak: an interpretation of cited temperature and histamine findings	§5.2
C4	Context (object type, posture, skin tone, priming, culture) acts as a prior on self-recognition and ownership	moderate for the cited effects; the prior reading is interpretive	§5.3, cited studies
C5	No unimodal area is self-specific; heightened responses to self-stimuli are surprise passed up the hierarchy	moderate for the first half, from cited imaging; weak for the second	§6.1
C6	TPJ, IPS, anterior insula and IFG form a non-self-specific multimodal circuit for self-recognition and body ownership	moderate: cited imaging and anatomical connectivity	§6.2
C7	Predicted self-generated or self-related stimuli suppress sensory and multimodal activity	moderate: tickle and self-face studies, cited	§6.3
C8	Any sense, not only motor efference against reafference, can drive self-recognition, which can occur without action	weak: follows from the framework; argued against Legrand and Ruby and Northoff	§4, §6
C9	The account is hard to falsify with standard 3T fMRI	moderate: argued from laminar mixing of prediction and error signals	§7

Concepts

• Free energy — “the upper bound on the level of surprise”; minimised by acting to sample predictable input and by updating beliefs about causes (§2).
• Surprise / prediction error — the discrepancy between predicted and actual input; phasic and local at each node, unexplained remainder passed up (§§2–3).
• Explaining away — top-down predictions from a higher level cancel error at the level below (§3).
• Empirical prior — a belief at an intermediate level of the hierarchy, constrained from above and below; high-level ones are abstract, multimodal and slow (§4).
• Corollary discharge — a prediction of the sensory consequences of movement; set against efference copy, a copy of the motor command (§4).
• Self-recognition — recognising a stimulus (face, voice, body part) as one’s own; distinct here from self-processing in general.
• Enfacement illusion — synchronous felt and seen facial touch making another’s face more like one’s own in recognition and report.
• Rubber voice illusion — synchronous feedback of another’s voice heard as a distortion of one’s own.

Connections

The paper builds on Friston’s free-energy and predictive-coding papers, among them Friston, Daunizeau, Kilner and Kiebel 2010 (LIT-1181) and Friston, Mattout and Kilner 2011 on active inference, and on Clark’s “Whatever next?” (then in press). It sets itself against two accounts: a self-specific cortical network (Northoff et al. 2006) and self-processing as motor efference matched with reafference (Legrand and Ruby 2009). It cites Seth, Suzuki and Critchley’s interoceptive predictive coding model of conscious presence (2011) as the interoceptive version of the latter. Its account of the illusions draws on Tsakiris’s own multisensory work and on Hohwy’s predictive-coding treatment of the RHI and binocular rivalry. It extends Kilner and colleagues’ predictive-coding account of the mirror-neuron system. It cites Botvinick and Cohen 1998, Lenggenhager et al. 2007 and Blanke and Metzinger 2009, which are being filed alongside it.

Bearing on the record

• Metzinger’s self-model (LIT-450, NOTE-372; THEORY-064). NOTE-372’s C6, that body ownership can be manipulated by multisensory conflict, rested on findings the record did not hold. This paper gives those findings a mechanism: ownership is a posterior that synchronous stimulation shifts, and the real hand loses probability as the rubber hand gains it. Its “the notion that there is a ‘self’ is the most parsimonious and accurate explanation for sensory inputs” is close to THEORY-064’s claim that what is experienced as a self is the content of a self-model. The paper, though, says nothing about transparency, and does not deny a self. It explains how the body is modelled as one’s own, not why the model is experienced as a self rather than as a model. That is where Metzinger’s claim does its work, and it is not tested here. Putting the two together is mine; the paper cites Blanke and Metzinger 2009 only for the three conditions of selfhood.
• Active inference (LIT-1181, NOTE-960). The paper adopts NOTE-960’s C2 (action fulfils proprioceptive predictions, with no separate motor command) and draws a consequence the Friston paper does not: if motor signals are only proprioceptive predictions, motor efference has no special standing for the self, and self-recognition can run on passive, purely sensory correlations. The RHI, with passive stroking, is the case.
• Agency and ownership (LIT-1009, NOTE-827; THEORY-322; LIT-453). Haggard (NOTE-827 C1) and Gallagher separate ownership from agency. This paper is about ownership and recognition and leaves agency aside. Its §6.3 suppression result is the same sensory attenuation that Haggard reports the comparator model predicts, and Haggard’s objection to the comparator (it explains non-agency but gives no positive signal, NOTE-827 C8) has a counterpart here. The account gives the self a positive representation, the high-level prior, but its own §7 says that representation has not been found. The comparison is mine.
• Interoception (LIT-542, NOTE-434; LIT-135; THEORY-045). The paper lists interoception among the inputs to the self-prior, calls the anterior insula “primary interoceptive cortex” and cites Seth et al. 2011. But its worked cases are all exteroceptive, visuo-tactile or auditory, and interoception is named, not modelled. Its anatomy differs from Barrett’s: NOTE-434 records Barrett placing primary interoceptive cortex in the mid/posterior insula and visceromotor predictions in the ventral anterior insula, whereas this paper calls the anterior insula itself “primary interoceptive cortex”. It gives THEORY-045 no new support, since it says nothing about affect. The interoceptive side of the bridge is better taken from Seth 2013, filed alongside.
• The scope of the free-energy principle (LIT-598, LIT-603). The paper states that it does not argue for the free-energy principle (§3), and it uses no Markov blankets. The objections in those critiques do not reach it. As far as this reading can tell, its arguments use only hierarchical predictive coding, not the free-energy bound itself; the paper notes the debate over whether predictive coding needs that bound (Clark; Friston 2012b) and does not take a side. That reading is mine.
• QUESTION-048. The illusions are measured by questionnaire, by proprioceptive drift, by self-other recognition tasks and by physiology (temperature, histamine). The paper treats them as converging and does not ask which of them measures the experience.
• No THEORY holds the predictive account of body ownership. A grep of theory.d, claims.d and questions.d for “rubber hand”, “body ownership”, “self-recognition” and “enfacement” found only THEORY-060, which mentions self-recognition in development and does not bear on this account. THEORY-769, filed on 2026-10-20, now holds it. No instruction for machine-learning practice; nothing for the anthology.

Limitations

• A theory with no test of its own. The authors say there is “little direct evidence” for its core claims and that the data “neither largely supports nor refutes” it. The evidence they give was gathered under other accounts, and they have “largely looked for and reported evidence that supports” the theory (§7).
• Hard to falsify, by the authors’ account. The anatomical predictions are weak as stated: “will be engaged” and “not self-specific” are met by almost any multimodal finding. The paper names no result that would count against it.
• The physiological evidence for “rejection” is interpretive. Cooling and histamine in the real hand are read as falling probability of ownership; no mechanism links the two.
• A review by an author whose group ran many of the reviewed studies (Tsakiris on the RHI and enfacement). No effect sizes are given.
• Manuscript slips. “Leggenhanger” for Lenggenhager, “Clark et al In Press” for a single-author paper, and typos (“damming”, “somtosensory”). They do not affect the argument.

Open questions

• Is ownership graded and probabilistic, as the account says, or all-or-none? The authors propose model-based behavioural tasks to tell, and report none done.
• What would the high-level self-prior look like in the brain, and how would it be told apart from multisensory integration in general?
• If the self is the best explanation of sensory input, why does the RHI fail for a non-corporeal object given the same correlations? The account answers with a prior, which explains the result but would explain the opposite result with a different prior.
• Does the account extend to the sense of agency, where Haggard finds a prospective component (THEORY-322), or only to ownership?