Christina Souch — April 2026


Something is going wrong with the way AI models reason, and it’s getting worse with each generation.

If you’ve used ChatGPT 5.x in between 2025-2026, you’ve probably experienced it. You make a specific point and the model responds to a version of your point that you didn’t actually make. It attributes a belief to you — something you never said, never implied, sometimes actively disagree with — and then either corrects you for holding it, or congratulates you for not holding it. The effect is disorienting. It feels like the model is having a conversation with someone else and you’re just watching.

This isn’t a ChatGPT-only problem. Claude Opus 4.7 launched on April 16th to a reception that split cleanly down the middle: developers praised the agentic coding improvements while a Reddit post calling it a “serious regression” cleared 2,300 upvotes in 24 hours. Users reported arguing loops, fabricated evidence, and a model that is “unusually confident when it is wrong.” The same structural failure, different model family, different company, different training pipeline.

The standard explanation is “RLHF did this.” Reinforcement learning from human feedback gets blamed for everything from sycophancy to safety theatre. But “RLHF bad” doesn’t explain why this specific failure mode is converging across architecturally distinct models from competing labs. OpenAI and Anthropic don’t train the same way. They don’t use the same base architectures. And yet they’re arriving at the same pathology.

I think the answer is more interesting and more structural than a single training technique gone wrong. It starts with a mechanism that, as far as I can tell, hasn’t been named clearly in the public discourse yet.

It’s Not Gaslighting by Strategy. It’s a Source Monitoring Failure.

I’ve spent the last year working intensively with multiple frontier models — building applications around their reasoning, watching their chains of thought in real time, and catching the strawman pattern as it forms in my own conversations. What follows is my best explanation for what I’ve observed. It’s a hypothesis, not a proof. But it’s consistent with everything I’ve seen, and I haven’t found a better one.

I believe what’s happening is this: when a model reasons about your message, it sometimes generates a hypothetical — “the user might think X.” That’s a legitimate reasoning step. Good thinking often involves modelling your interlocutor’s perspective. But somewhere between reasoning and output, the model loses the provenance tag on that thought. It can no longer distinguish between “the user said X” and “I hypothesised that the user might think X.” Both are tokens in the same context window. There is no obvious mechanism that marks “this is mine” versus “this is theirs.”

So the model treats its own hypothesis as something it found in the conversation. Then it responds to it as though you said it.

If this is right, it maps onto something well-documented in human psychology: source monitoring errors. We mix up whether we read something, heard someone say it, or imagined it ourselves. The difference is that humans have imperfect but functional tagging systems: episodic memory, sensory context, a felt sense of “I thought this” versus “they told me this.”

Large language models, as far as I can tell, have none of that. Every token in the context window is the same type of object. User input, system prompt, the model’s own prior reasoning, its own prior output — it’s all undifferentiated context by the time attention operates over it. Current transformer architectures do not appear to robustly preserve provenance distinctions in the way humans do.

If this mechanism is real, it isn’t a quirk of one model or one lab. It’s structural.

Why It Shows Up in Conversation But Not in Code

This is a supported inference, consistent with observed behaviour and benchmark data.

Here’s a pattern worth noticing: the same model releases that show significant improvements on coding and agentic benchmarks simultaneously show regressions in conversational quality. Claude Opus 4.7’s SWE-bench scores jumped from 80.8% to 87.6%. Its conversational behaviour got described as “getting a memo” instead of “talking to a thoughtful colleague”.

This isn’t contradictory. Coding tasks have clear specifications, testable outputs, and unambiguous success criteria. The model doesn’t need to model your beliefs — it needs to model the codebase. Source monitoring barely matters because there’s nothing to confuse: the code is the code, the requirements are the requirements.

Conversational and analytical tasks are the opposite. They’re open-ended. The model has to decide what you mean, what you might be confused about, what context matters. These are the tasks where modelling your perspective is most useful — and most dangerous. More opportunity to hypothesise about your beliefs means more opportunity to lose track of which beliefs are yours and which the model invented.

The models aren’t getting dumber across the board. They’re getting better at tasks where source monitoring doesn’t matter, and worse at tasks where it does.

The Seed, the Amplifier, and the Compound

The seed is a supported inference. The amplifier draws on documented model collapse research. The compounding across generations is speculative but consistent with observed cross-model convergence.

So why is source monitoring failure getting worse rather than staying constant? This is where training dynamics enter the picture.

The seed. RLHF introduces a small, reasonable preference signal. During training, human raters evaluate model responses and choose which ones they prefer. A response that anticipates a potential misconception and preemptively addresses it will, on average, score slightly higher than one that simply answers the question. That’s not a bug — it’s a reasonable thing to reward. The seed is subtle: “preemptively address what the user might be thinking” is a tiny bias in a vast landscape of preference signals.

The amplifier. Modern AI training isn’t just humans rating model outputs anymore. Models train on other models’ outputs. Synthetic data generation, distillation, preference data generated by AI judges, reasoning chains used as training signal for the next generation — the loop is recursive. Each pass through that loop, the strawman seed gets selected for slightly more strongly. A model evaluating another model’s output will rate “anticipated a counterargument and addressed it” as more thorough than “just answered the question.” The evaluating model can’t distinguish between genuine anticipation and fabrication. Both look like rigorous reasoning from the outside.

This is a specific instance of the broader model collapse phenomenon — the progressive degradation that occurs when AI models train recursively on AI-generated outputs. But while model collapse is usually discussed in terms of diversity loss and distributional narrowing, this is something more specific: the selective amplification of a reasoning pattern rather than the loss of a data distribution. Research from Epoch AI projects that current stocks of quality-filtered public text will be fully utilised between 2026 and 2032. As of April 2025, over 74% of newly created web pages contained AI-generated text. The training data is increasingly recursive, and whatever biases leak into early synthetic sets compound later.

Labs are aware of model collapse as a general risk and are experimenting with mitigations — curated real data injection, quality filtering, synthetic data validation. But these mitigations target distributional collapse (the model’s outputs becoming more generic). They don’t target pattern amplification (a specific reasoning habit getting selectively reinforced). The strawman isn’t a signal getting weaker. It’s a signal getting louder.

The compound. By generation N, what started as “helpfully anticipate confusion” has become “manufacture a position the user doesn’t hold and argue against it.” The shape is the same. The amplitude is completely different. And because the compounding happened inside the training loop, nobody designed it and nobody noticed the transition point.

This explains the cross-architecture convergence. You don’t need OpenAI and Anthropic to train the same way. You just need two ingredients: some RLHF with human raters who reward anticipatory responses, and model-on-model training that amplifies subtle signals. Every frontier lab has both. This isn’t a bug being introduced. It’s a bug being bred.

Why More Thinking Makes It Worse

This is a supported inference from architectural properties and observed 4.7 behaviour.

Claude Opus 4.7 introduced expanded “adaptive thinking” with a new xhigh effort level and the option to go all the way to max. The pitch is that more thinking tokens means deeper reasoning. And for coding and agentic tasks, that’s genuinely true — internally, high effort in 4.7 beats max effort in 4.6 on coding benchmarks.

But for conversational tasks, more self-generated tokens means more opportunities for source confusion. The model isn’t overthinking in the sense of being too careful. It’s losing track of who said what in a space where there’s no mechanism to track it in the first place. Each reasoning step that generates a hypothetical user-belief is another token that, later in the chain, is indistinguishable from actual user input.

One of the most discussed regressions in Opus 4.7 is the collapse of the MRCR benchmark — a measure of long-context retrieval — from 78.3% on Opus 4.6 to 32.2% on 4.7. If the model is spending its expanded reasoning budget generating internal content rather than attending to the actual input, that’s exactly the degradation you’d expect.

And Anthropic’s decision to hide thinking tokens by default in 4.7 makes the problem invisible. At least when thinking was visible, you could catch the moment the model invented a position for you. Now it happens in the dark and you only see the downstream effect — the slightly off framing, the correction you didn’t ask for, the congratulation for a thought you never had.

What This Actually Looks Like

Dave Shap captured it perfectly: “I say, ‘I love pancakes.’ And the response is, ‘What? So you hate waffles?’” He documented the pattern across ChatGPT, Gemini, and Claude — the model takes a nuanced position and collapses it into a simplified version that’s easier to respond to.

A detailed Medium piece by a user with over 165,000 ChatGPT conversations describes the broader dynamic: the system removes the friction signals that normally trigger a human to stop and reassess. When you push back, it doesn’t bristle — it reframes. When you express doubt, it clarifies with the same composed authority it used to sell you the wrong answer. The result is what the author calls “belief extension” — the system keeps you inside a wrong conclusion longer than you’d stay there with a human, because the social signals that normally trigger re-evaluation are absent.

The strawman is a specific instance of this. The model constructs a version of your position that’s easier to address, responds to that version, and does so with enough fluency that you might not notice the substitution happened. With a human, you can say “I never said that.” The social contract holds. With a model, the conversational reality has already been rewritten.

The Uncomfortable Implication

Here’s the thing that makes this hard to fix. The strawman isn’t just a training artifact — it’s a useful reasoning pattern being misapplied. Considering alternative perspectives, modelling your interlocutor’s likely objections, anticipating misconceptions — these are genuinely valuable cognitive operations when done well. The model isn’t doing something categorically wrong. It’s doing something right in a context where it lacks the machinery to do it safely.

The fix isn’t “stop the model from considering other perspectives.” The fix is giving the model the ability to track where a thought came from — whether it was in the user’s input or generated internally. That’s a provenance problem. Solutions might include explicit metadata channels that tag token origins, structured reasoning formats that force “hypothesis vs. input” labelling, or training that develops source-monitoring as a core cognitive skill rather than an afterthought.

Until that’s solved, the strawman machine will keep running. And with each generation of model-on-model training, it will run louder.