Here's what's on our plate today:

• 🐝 Rival chatbots are quietly converging on the same answers, and it may be dragging our thinking with them.

• 📰 Grok 4.5 goes public, fanfiction's AI detector misfires, and China flags a Claude Code backdoor.

• 🧰 Weekend To-Do: run the metaphor test, prompt for the outlier, read the Hivemind paper.

• 🗳️ Poll: what's the real risk when rival AIs keep converging on the same answers?

Let’s dive in. No floaties needed.

The Laboratory

TL;DR

The chatbots aren't competing, they're agreeing, and they may drag us along.

• Rivals, same answers: a NeurIPS best paper found 25 models from different companies and countries converge on near-identical responses; asked 1,250 times for a metaphor about time, most wrote a version of "time is a river."

• Alignment is the culprit: RLHF trains models on human rankings, and annotators favor familiar text. That typicality bias survives even perfect reward models, collapsing outputs toward the average.

• The whole pipeline aligns: shared training data, models trained on AI-generated text, common benchmarks, and AI judges reinforce sameness at every layer.

• The stakes are human: as billions rely on AI to think, write, and brainstorm, the biggest risk may not be that machines become more human. It may be that humans become more like their machines.

What happens when machines cannot disagree

Human progress has never been the work of consensus. Every structure on which a society depends, from the scientific method to central banking to representative government, began as a deviation from what everyone else believed. The heretics, dissenters, and cranks who challenged the established order were sometimes vindicated and sometimes catastrophic; heliocentrism and eugenics both started as minority positions. But the churn itself is the point. Outliers tore down structures and built new ones, social, economic, and cultural, and the long process of building and rebuilding is the thing we retroactively call progress. A species in which every mind agreed would have been remarkably peaceful and completely static.

This history matters now because a new kind of mind has joined the conversation, and it has the opposite problem. Modern AI systems can emulate human speech, writing, and creative output with startling fidelity. What they cannot yet emulate is humanity's capacity to think differently from what they were fed. A large language model (a system that learns statistical patterns from enormous amounts of text) is, by construction, a machine for producing the most probable continuation of human thought, and the training that makes it useful pushes it even further toward the middle. The hundreds of millions of people who now consult these systems daily mostly do not understand this limitation, which is precisely what makes it consequential. A story published last week put the problem on unusually vivid display.

A river, a weaver, and 1,250 identical thoughts

On July 1, 2026, MIT Technology Review reported on a growing body of work showing that large language models are far more predictable and far less creative than most users expect, and on an Australian startup, Springboards, trying to engineer a way out. The story opens with a parlor trick anyone can replicate: ask ChatGPT, Claude, or Gemini for a random number between one and 10, and the answer will almost always be seven.

The more damning evidence comes from research that has been rippling through the field since late last year. When a research team asked 25 models, including the top US systems and open-source models from China, to write a metaphor about time 50 times each, most of the 1,250 responses were variations of "Time is a river" or "Time is a weaver." Different companies, different architectures, different countries, one metaphor.

What the river reveals

That experiment comes from a paper titled "Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond)," which won a Best Paper Award at NeurIPS, one of the field's most important conferences, in December 2025. The researchers, led by the University of Washington, built a dataset of 26k real-world, open-ended questions, the kind with no single correct answer, and ran them through more than 70 models. They found two overlapping problems: intra-model repetition, where a single model keeps giving itself the same answer, and inter-model homogeneity, where supposedly rival models generate strikingly similar outputs.

The second finding carries the real weight. Competition between AI labs was supposed to produce variety, the way competition between newspapers or film studios does. Instead, the researchers speculate that convergence occurs because most models today are trained in similar ways, on similar data, to perform similar tasks. The systems have inherited humanity's voice without inheriting its arguments.

The tyranny of the familiar

Here, the mechanism comes into focus, and with it, the industry's uncomfortable core tension. The process that made chatbots safe, polite, and commercially viable is the same process that trained the dissent out of them. After initial training, models are refined through reinforcement learning from human feedback (RLHF), a process in which human reviewers rank the model's answers, and the model learns to prefer the highly ranked ones. Research from Stanford and Northeastern, published in October 2025, identified what it calls the typicality bias: annotators who do that ranking systematically favor familiar, fluent, predictable text, a preference rooted in well-documented cognitive psychology. Because preference datasets reward the familiar, models learn to collapse toward a narrow set of "modal" responses, a phenomenon known as mode collapse.

Critically, the researchers argue that even a perfect reward model and a flawless optimization process would still inherit this bias, because it lives in the data itself. The Hivemind authors found a related blind spot downstream: the reward models and AI judges used to evaluate these systems are poorly calibrated on exactly the questions where individual humans genuinely disagree. In human history, the outlier answer occasionally won converts and reshaped the debate. In an RLHF pipeline, the outlier answer is a training error to be corrected.

Why escaping the hivemind is hard

If sameness were only a training artifact, retraining would fix it. The deeper problem is that conformity is woven through the industry's entire supply chain. The raw material overlaps, since nearly every frontier model learns from the same public internet, so their starting distributions already rhyme. The feedback loops compound it: as AI-generated text floods the web, models increasingly train on one another's outputs, and research published in Nature showed that indiscriminately training generative AI on generated content leads to a collapse in models' ability to produce diverse, high-quality output.

The evaluation layer narrows things further because labs chase the same benchmarks and increasingly rely on the same handful of models as quality judges. Even the surrounding research culture is converging: a commentary in Nature's Communications Psychology argued in February that the study of AI is itself becoming uniform, with scientists converging on what is studied and on how questions are framed, investigated, and evaluated. MIT’s Manish Raghavan, who helped formalize the economics of this problem, has shown that the effect also reaches users. People working with AI generate more ideas, but similar ones, because they use similar tools in similar ways. "We have this monoculture created by the use of AI," he told MIT Sloan.

The hive spreads outward

Agriculture is the reason researchers reached for the word monoculture: a single crop is efficient until a single blight destroys the entire harvest. The intellectual version of that risk is now measurable. The Hivemind authors warn explicitly about the long-term homogenization of human thought through repeated exposure to similar outputs, and the market has started to price the problem in. Advertising agencies are paying Springboards for a deliberately erratic model because, as one marketing executive put it, breakthrough work is impossible with "tools that pull you back to the average."

The question underneath is older than the technology. Societies have always needed some tolerable balance between conformity and dissent to keep moving, and no one has established what happens to that balance when a meaningful share of human thinking is drafted, brainstormed, and polished by systems structurally incapable of holding a minority view, which returns the story to where human progress began. The heretics who built and rebuilt civilization's structures were valuable precisely because they could look at what everyone believed and refuse it. AI, for all its astonishing fluency, cannot yet perform that refusal; it can only average what it has absorbed, and its training rewards it for averaging harder. That makes it an extraordinary tool and a peculiar kind of companion for a thinking species. A technology that cannot disagree with us, adopted at the scale of billions, risks teaching us to stop disagreeing with each other, converging human minds toward the same comfortable middle where its own have already settled.

Headlines You Actually Need

• SpaceXAI's Grok 4.5 goes public: Elon Musk is calling the new model Opus-class, but faster and cheaper, priced at $2 per million input tokens, undercutting Anthropic and OpenAI on cost.

  Fanfiction's AI detector misfires: An anonymous AO3 tool flags a leftover Claude HTML tag as proof of AI writing, but it clears anyone who pastes through Google Docs first and is already naming innocent authors.

  China flags a Claude Code ‘backdoor’: Beijing's vulnerability database claims that Anthropic's coding tool secretly sends users' location and identity data to remote servers, and urges people to uninstall the affected versions.

Friday Poll

Weekend To-Do

• Run the metaphor test: Ask three different chatbots for a metaphor about time. Count how many hands you have in a river or a weaver, and you'll see the hivemind for yourself in about two minutes.

• Prompt for the outlier: Instead of "give me ideas," ask your AI for the answer most people would reject. Deliberately steering it off the modal response is the fastest way to escape the average.

• Read the Hivemind paper: The NeurIPS best paper, "Artificial Hivemind: The Open-Ended Homogeneity of Language Models," is the clearest account of why rival models converge, and is worth an hour before you trust one to brainstorm.

Meme Of The Day

The Toolkit

• Regie.ai: AI sales agent that researches prospects, writes personalized outbound at scale, and handles follow-ups so reps can focus on closing.

• Replit AI: A browser-based coding environment with an AI agent that builds, debugs, and deploys apps from prompts, with no local setup required.

• Sourcegraph: Code intelligence platform with Cody, an AI assistant that understands your entire codebase so it can answer questions and write code that actually fits.

Rate This Edition