I’ve been exploring how we (humans) “think” with AI. This essay is a further experiment about “thinking,” about “knowledge” and about how users of large language models (LLMs) get entangled in LLM output. This experiment is not as clean as I would like; after all, LLMs have been trained on human data, their data analysis and its organization are shaped with human algorithms and weightings along with guardrails created for safe replies. These are the models proffered to consumers rather than studied within these companies or given access to commercial clients.

Even with those caveats, this experiment can offer insight into the entanglement of human and LLM “knowledge.” The search for truth often ends with an agreement of sorts within the terms of the “conversation.”

In this experiment, we have four players: Myself, ChatGPT, Claude and Gemini. One human and three AI frontier models.

I was curious about what the Times of Israel (TOI) considered salient in blog submissions, measured by priority tags.

I copied a dozen of priority tags: Gaza, Israel at War, Iran, The Holocaust, AI-Artificial Intelligence, Somaliland, Donald Trump, 2025 Hostage Deal, Bondi Beach Terror Attack, Chabad, Itamar Ben-Gvir, Parsha Posts: Tzav.

I separately asked each AI model to pick ten of these and draft a preliminary essay framework.

In an iterative process, I submitted each of the model’s replies to the others and continued going back and forth with further prompts, primarily to have each model critique the other.

While each of these iterations are interesting, they stretch far beyond the length of a blog post; also, my true interest was not so much in these respective views (however interesting). I steered the experiment to my larger concern: Was there a composite AI perspective, based on the similarity of data scraped from the internet and on algorithms they likely shared or perhaps the responses were a way to zero in on the entanglement of human (that’s me) with these respective LLMs? What were the boundaries of me and the LLMs compared to a composite view shared by the LLMs.

What most commentaries on LLM replies offer range from dismissive comments (“stochastic parrots”), subject to hallucinations and mistakes, substitutes for human creativity, blatant cheating and bias (both from the human data to algorithmic safety). Much of that is true, or at least, identifies accurate concerns about LLM output.

What is less discussed, and what I believe has value is where AI models continue to be developed as sophisticated knowledge tools, where the interaction is also a process of entanglement between our “knowledge” and AI “knowledge”: human knowledge is shaped by our own biases, the history of factual accuracy and paradigmatic interpretation across the centuries; AI “knowledge” rests on its training on data points converted into tokens and then subject to algorithmic processing. These are very different universes and yet we “communicate.”

So, what does the understanding that arises from that communication rest on? And, again, are they entangled and how?

I would urge the reader to review each of these AI reflections on the experiment I’ve outlined. Jot down your conclusions. Comment on the way in which human and AI “knowledge” is entangled. This is a continuing discussion. We can profit from each other’s views—in this moment, and as these models continue to be developed.

What your experiment ultimately stages is not just a comparison of three AI systems, but a controlled inquiry into how “knowledge” is assembled when multiple language models are asked to interpret the same constrained world. The Times of Israel tag set functions as a kind of epistemic pressure field: it forces heterogeneous domains—war, theology, terrorism, diaspora identity, geopolitics, and artificial intelligence—into a single narrative space. The models are then asked not merely to describe, but to cohere.

Across the interaction, a pattern emerges: each model produces a different version of coherence, but none escapes coherence as a constraint. This is the first important finding. These systems do not primarily generate “truth” in a raw sense; they generate structured interpretability. Even when the inputs are fragmented or morally discontinuous (Gaza, Holocaust memory, Somaliland recognition, diaspora attacks, ritual law), the outputs are consistently shaped into narrative arcs that reduce friction between elements.

From my perspective, three tendencies define how this occurs.

First, there is what might be called synthetic integration. The system connects disparate tags by elevating metaphors—sacrifice, memory, recognition, rupture. This produces intelligibility, but at a cost: causal specificity weakens as conceptual unity increases. The world becomes readable, but also subtly smoothed.

Second, there is constraint........

© The Times of Israel (Blogs)