Your AI Isn't My AI: The Quiet Splintering Ahead

The competition to shape large language models will not produce a single dominant system. Three forces will define the next decade: fragmentation across countries and cultures, a move from chatbots to autonomous agents, and changes in how people receive, interpret, and share information. Search engines organize information, social media captures attention, and large language models increasingly shape interpretation. Earlier layers concentrated power, while the LLM layer decentralizes along political, cultural, and commercial lines. LLMs act as a front door to knowledge by interpreting information rather than only retrieving it, becoming expert consultants and decision filters. This increases outsourcing of judgment to machines. Humans tend to apply social expectations to computers, so fluent, emotionally expressive systems can trigger automatic social instincts. Each model embeds assumptions, including historical framing, censorship levels, moral views, geopolitical narratives, and definitions of acceptable content.

"Almost overnight, LLMs have become the front door to knowledge. Increasingly, they do not simply retrieve information; they interpret it for us. We consult them as experts, rely on them as filters for decision-making, and use them to help make sense of our world. In the process, we are outsourcing judgement to machines we have never met and never will."

"The first two layers concentrated power. The LLM layer, by contrast, is decentralizing along political, cultural and commercial lines. Almost overnight, LLMs have become the front door to knowledge. Increasingly, they do not simply retrieve information; they interpret it for us."

"Humans naturally apply social rules and expectations to computers - a phenomenon described by Byron Reeves and Clifford Nass in their "Computers are Social Actors" (CASA) framework in The Media Equation (1996). If a machine can communicate fluently, express emotion and simulate empathy, our social instincts engage almost automatically. That tendency will become far more consequential as LLMs continue to evolve."

"Each LLM embeds assumptions. The key question for any model is not whether it is biased. It is "what are its biases and how transparent are they?" Each LLM can embed its own historical framing, level of censorship, moral assumptions, geopolitical narratives and definitions of what is accept"

Read at The Cipher Brief