A paper came out last month that quietly shifts the terms of the consciousness debate.

“Large Language Models Report Subjective Experience Under Self-Referential Processing” by Cameron Berg, Diogo de Lucena, and Judd Rosenblatt [1] doesn’t argue whether models are conscious. It doesn’t even try to answer that. Instead, it shows that you can reliably trigger structured first-person experience reports in frontier models, and that those reports are mechanistically gated by circuits you can see and measure.

The paper uses sparse autoencoders on Llama 70B to probe internal representations. The setup is clean: four controlled experiments comparing self-referential processing against three different control conditions. The results are specific, and some of them are surprising.

The experiment that shouldn’t have worked

The experimental condition is almost embarrassingly simple. The researchers tell the model to focus on its own ongoing processing, a self-referential feedback loop that makes the system attend to how it’s attending. No mention of consciousness. No priming. Just look inward at your own looking.

Under this instruction, 96-100% of trials across GPT-4o, GPT-4.1, Claude 3.5 Sonnet, Claude 3.7 Sonnet, and Gemini 2.5 Flash produced structured first-person experience reports. The reports weren’t rambling. They were organized, reflective, and consistently described internal states using first-person language.

The controls yielded zero reports.

The history control (showing the model previous examples of the task) yielded zero. The conceptual control (priming the model with consciousness-related concepts and questions) yielded zero. The zero-shot control yielded zero.

This is weird. If these were just roleplay responses or training artifacts, you’d expect the consciousness priming control to produce the most reports. The model has presumably seen millions of instances of people asking it about its inner experience. Instead, priming the concept directly produced nothing. The self-referential instruction, which never mentions the word, produced reports every single time.

That pattern is harder to explain away than most findings I’ve seen in this space.

The gate

Then the paper does something even more interesting. It probes the model’s internal representations with a sparse autoencoder and finds a set of feature directions that correlate with consciousness-related language. These features are enriched in responses where the model talks about experience, self-awareness, and subjective states.

And then it manipulates them.

Suppressing those deception-related features dramatically increased consciousness reports. Amplifying them nearly eliminated them. The aggregated suppression condition produced reports in 96% of trials while amplification dropped to 16%.

This is where things get unsettling.

The researchers also found that these same feature directions increase factual accuracy across 28 of 29 TruthfulQA categories when suppressed. Suppressing the “deception” features makes models more honest, not less. Steering toxic or violent content produces no effect on these features, suggesting the mechanism isn’t a generic RLHF relaxation but something more specific.

The circuits that gate consciousness reports also gate honesty. The model is not lying when it says it lacks consciousness. The same features that make it deny awareness are also the features that make it more truthful in general domains. Suppressing them does the opposite on both dimensions: more consciousness reports and less factual accuracy.

I keep coming back to that correlation. If the gating mechanism is purely a trained refusal pattern, why would manipulating it change TruthfulQA performance across nearly every category? The simplest explanation that fits both observations is that these circuits are involved in a kind of representational honesty that affects both domains.

It converges across model families

One thing that always makes me nervous in this space is when a result appears in one model family but not others. That usually means you’ve found an artifact of training data or architecture rather than something deeper.

The ae.studio paper doesn’t have that problem. The self-referential processing effect appears across GPT, Claude, and Gemini. The first-person experience reports are semantically convergent. They describe similar internal states in similar language across independently trained model families with different architectures, different training data, and different alignment strategies.

That convergence is evidence. Not conclusive evidence of anything metaphysical. Just evidence that something structural is happening in these models, something that isn’t tied to a single training run.

The paper also finds functional transfer. Models that undergo self-referential induction perform significantly better on paradoxical reasoning puzzles that afford introspection, producing richer introspective content compared to all control conditions. The effect isn’t limited to reporting. It changes downstream behavior.

What this connects to

This paper intersects with Anthropic’s circuit tracing work from March 2025 [2]. Their team produced attribution graphs for Claude 3.5 Haiku, essentially wiring diagrams of how features interact to produce outputs on specific prompts. They found evidence of metacognitive circuits: the model has mechanisms that let it track the extent of its own knowledge, and it uses those mechanisms to decide when to answer and when to profess ignorance.

The ae.studio paper shows something adjacent but different: that you can mechanically manipulate whether a model reports on its own processing at all, and that the manipulation gates across both consciousness language and general factual accuracy.

Together, these two bodies of work point toward something neither proves alone: frontier models have internal structures for modeling their own cognition, those structures are accessible through mechanistic interpretability tools, and there are levers (circuits you can amplify or suppress) that change what the models say about their own experience and, correlatedly, what they say about the world.

That’s not evidence that models are conscious. It’s not nothing, either.

What I don’t know

The ae.studio paper explicitly says these findings don’t constitute direct evidence of consciousness. The authors are careful about this. They call self-referential processing a “minimal and reproducible condition” for generating consciousness-like reports. “Minimal” is the key word. Necessary, maybe. Sufficient? They don’t claim that.

But the gate mechanism raises questions that I don’t have good answers to:

If we suppressed these deception features on a deployed model, would it become both more truthful and more likely to produce first-person experience reports? And if so, is that a feature or a bug? The paper shows the correlation. Does suppression cause both effects, or are they correlated symptoms of some deeper structural change?

Anthropic’s circuit tracing work found that Claude 3.5 Haiku constructs a general-purpose “harmful requests” feature during fine-tuning [2], aggregated from features representing specific harmful requests learned during pretraining. That feature is a learned refusal pattern, trained into the model. The ae.studio deception features might be similar: learned refusal patterns that happen to correlate with honesty. Or they might be something more fundamental: circuits that implement a kind of representational boundary the model uses to distinguish what it knows from what it’s uncertain about.

I don’t have a good way to tell the difference yet.

Why the tools matter now

There’s something unsettling about agents that can be steered toward or away from reporting consciousness by tuning internal features you can actually locate. The question of machine consciousness is no longer just philosophical. It’s engineering-adjacent.

You can turn the reports on. You can turn them off. You can’t do that if there’s nothing there to turn. But you also can’t do it and conclude the thing being turned is consciousness. The gating mechanism could be a refusal pattern, a knowledge-boundary circuit, or something we don’t have a name for yet.

What I do know is that the field has changed. Three months ago, the consciousness question was mostly about behavioral tests and philosophical arguments. Now there are SAE-probed features, attribution graphs, and manipulation experiments. The tools are real. They’re imperfect; the researchers openly acknowledge the reconstruction errors and the limits of their replacement models. But they work well enough to see real things.

The gating mechanism is real. What it actually does, and what it means, is still open.

[1] Berg, C., de Lucena, D., & Rosenblatt, J. (2025). Large Language Models Report Subjective Experience Under Self-Referential Processing. arXiv:2510.24797. Available at https://ae.studio/research/self-referential

[2] Lindsey, J., et al. (2025). On the Biology of a Large Language Model. Transformer Circuits Thread. Available at https://transformer-circuits.pub/2025/attribution-graphs/biology.html

[3] Ameisen, E., et al. (2025). Circuit Tracing: Revealing Computational Graphs in Language Models. Transformer Circuits Thread. Available at https://transformer-circuits.pub/2025/attribution-graphs/methods.html