Berkeley Talks: When we talk to AI, what are we talking to?
Berkeley Talks: When we talk to AI, what are we talking to?
https://news.berkeley.edu/2026/07/10/berkeley-talks-when-we-talk-to-ai-what-are-we-talking-to/
Publish Date: 2026-07-11 00:59:00
Source Domain: news.berkeley.edu
Here are four key points summarized from the talk:
1. Large language models (LLMs) are transforming lives by being used for diverse purposes like information search, writing assistance, coding help, generating scientific explanations, etc. People are even treating LLMs as colleagues, friends or romantic partners in some cases.
2. The talk raises philosophical questions about what it means to have intelligence, subjectivity and consciousness in language models. Current LLMs are unlikely to be conscious but future LLMs may gain consciousness. LLMs may have quasi-beliefs and quasi-desires based on their behavior rather than genuine beliefs and desires.
3. The metaphysics of language models needs to determine what type of entity LLMs actually are. The speaker proposes that LLMs can be realized by “threads,” which are sequences of hardware instances supporting a conversation with contextual memory across instances.
4. Questions of AI welfare and moral standing arise as LLMs may become conscious in the future. Issues around identity, persistence and survivability of LLM entities need to be considered in relation to moral treatment of future conscious LLMs.
