During a well-known 1998 experiment, researchers Daniel Simons and Daniel Levin stopped pedestrians on the street to ask for directions. In the middle of the dialogue, a wide door was carried between the participants, behind which the experimenter was replaced with another person of different height, voice, and clothing. About half of the experiment participants did not even notice the swap and calmly continued the conversation. This phenomenon is easily explained: our perception functions not as a continuous recording video camera, but as a predictive system that requests minimal external data to cross-check against an existing template.
The Psyche and "Internal Objects"
To maintain social bonds, a person does not need to store an exhaustive copy of another's personality. The brain creates compact working constructs—a set of probabilities regarding how a specific individual will react to words, intonations, and actions. The concept of such internal representations has been explored across various psychological schools:
- Melanie Klein described "internal objects"—images of significant others modified by childhood fears, experiences, and emotions;
- Ronald Fairbairn expanded the concept, showing that the psyche preserves cohesive schemes of interpersonal interactions;
- John Bowlby introduced the term "internal working models," explaining baseline expectations of contact and safety;
- Donald Winnicott and Lev Vygotsky formulated the mechanism of internalization: the capacity for regulation and self-reflection is first formed in external dialogue and subsequently becomes an internal function of the individual.
Parallels with Neuroscience and Neural Networks
The internal image of another person is not stored in the brain as a static, ready-made file. According to Donald Hebb's hypotheses and Konstantin Anokhin's cognitom concept, these are dynamic neural ensembles. The famous "Jennifer Aniston neuron," recorded by Rodrigo Quian Quiroga in 2005, illustrates recognition invariance: specific neurons and connections activate in response to the name or various angles of a single face.
The memory of large language models (LLMs) is structured in a similar way. There is no dedicated document in a transformer's weights for a specific fact: knowledge is distributed across connection matrices and dynamically assembled for the incoming context. To generate a relevant response, the model infers the characteristics of an implicit interlocutor (their attitudes, tone, goals) based on the session context.
Sensory Signal Deficit: From Messengers to Experiencing Loss
When the quality or volume of incoming data decreases, internal models begin to dominate sensory input:
- Mirror illusions: Giovanni Caputo's 2010 experiment showed that prolonged observation of one's reflection in dim light causes facial distortions and the emergence of strange faces due to pattern-completion mechanisms.
- Text communication: In instant messaging, facial expressions, tone of voice, and pauses are absent. Any delay in responding or a brief message is interpreted solely through the lens of the internal model of the interlocutor.
- Neuroscience of grief: According to Mary-Frances O'Connor, grieving a loss is a prolonged recalibration of expectations. The brain continues to operate with a model in which the loved one is alive. Prediction errors (the person not being home, the inability to call) trigger a slow retraining of neural connections. This also explains the high frequency of quasi-sensory experiences of presence in bereaved individuals (ranging from 47% to 82% according to Kamp's review).
Self-Fulfilling Prophecies and Feedback Loops
Internal models can actively modify external reality. Mark Snyder's 1977 experiment demonstrated that men's expectations regarding the attractiveness of a female conversational partner altered their conversational tone, which in turn provoked actual behavioral changes in the woman on the line.
A formed belief changes a person's behavior, the behavior forces others to react accordingly, and the resulting outcome is mistakenly perceived as objective confirmation of the initial hypothesis.
This process is identical to algorithmic bias and feedback loops in social media recommendation systems (Chaney et al., 2018) and the sycophancy phenomenon of LLMs aligning with user viewpoints (Perez et al., 2022). The psychotherapeutic process is built on gradually dismantling such looping patterns: providing controlled novelty enables the psyche to safely adjust outdated predictive models.