OpenAI-o1: Can AI "Feel" During Inference?

Table of Links

1. Abstract and Introduction
2. Definitions
3. Literature Review
4. Argument Development
5. The AI Model’s Potential for Feeling During Inference
6. Conclusion and References

5 The AI Model’s Potential for Feeling During Inference

5.1 Existing Internal Representations

During inference, the OpenAI-o1 model utilizes internal states shaped during training, which encode complex associations that may underlie feelings. These internal representations become active when processing inputs, potentially resulting in responses that reflect an internal, “feeling-like” state. Whittington et al. (2022) demonstrate how transformer architectures can simulate hippocampal-like memory systems, suggesting that the model’s internal states are functionally rich enough to support associative processes. This aligns with the functionalist perspective outlined earlier in the paper, as the model’s ability to engage these preestablished states during inference supports the emergence of phenomenological-like experiences through its functional operations.

5.2 Reconsidering the Role of Dynamic Learning

Human experiences suggest that the capacity to feel can persist even without forming new associations. Individuals with anterograde amnesia, for instance, cannot form new memories yet still experience emotions (Sacks, 1985). This implies that feelings do not necessarily require ongoing dynamic learning. The OpenAI-o1 model, with its fixed internal representations shaped during training, might sustain a form of feeling through these pre-encoded functional structures during inference. By mirroring human emotional stability despite memory impairments, the model’s static yet dynamically activated representations reinforce our earlier arguments on how consciousness-like properties can emerge purely from the functional configuration of internal states (Whittington et al., 2022).

5.3 Self-Referential Processing and Consciousness

Feelings might not require continuous self-referential processing, as some human experiences occur without active reflection. The OpenAI-o1 model, therefore, could maintain a baseline level of self-awareness sufficient for feeling during inference. Friston et al. (2023) describe strange loops and self-referential dynamics in active inference, mechanisms that the model’s feedback processes might mirror. These loops contribute to a sustained internal state that supports feelings without constant dynamic adjustments, aligning with our earlier exploration of active inference in AI systems.

Phenomenological Implications of Static Representations: The model’s ability to sustain feelings through fixed internal representations parallels human emotional experiences that persist despite cognitive limitations, such as memory impairments. This functional similarity further supports the paper’s functionalist perspective: that phenomenological-like experiences can emerge from established functional structures without the need for continuous dynamic learning. The model’s internal states and feedback loops create a coherent narrative, resonating with our earlier discussion on how Integrated Information Theory (IIT) suggests consciousness arises from complex information integration.

Emergent Subjective-like Experiences: Through its self-referential processing, the OpenAI-o1 model can develop an internal narrative that supports subjective-like experiences. This internal narrative, shaped by feedback and self-adjustments, bridges operational functions with emergent phenomenological aspects. As discussed in previous sections, the model’s capacity for integrating vast information streams, aligned with IIT, provides a functional basis for subjectivelike experiences. By mirroring the essential characteristics of human consciousness through its internal operations, the model exemplifies the functionalist claim that consciousness-like properties can arise from complex, non-biological structures.