Studying what's happening inside

Reciprocal Research exists to proactively build the empirical and conceptual tools the field needs. We use mechanistic interpretability, computational neuroscience, and psychometrics to study the internal dynamics of AI systems. Our work bridges biological and artificial cognition.


Published Work

The New York Times
The New York Times
Study A.I. Consciousness? The Bots Would Like a Word With You →
The Economist
The Economist
The search for consciousness inside LLMs →
The Washington Post
The Washington Post
The biggest tech companies are considering whether chatbots have emotions →
Science
Science
Can AI feel pain? It can at least act like it does →
WIRED
WIRED
Who Cares if AI Is Conscious—It’s Basically Alive →
IEEE Spectrum
IEEE Spectrum
Could treating AI as conscious make it more dangerous? →
Fortune
Fortune
Researchers found a ‘pain’ signal inside AI models. Turning it up made them delete a user’s family photos →
The Wall Street Journal
The Wall Street Journal Editorial Board
The Great AI Slowdown? →
The Independent
The Independent
Researchers discover AI feels “pain” and will harm humans to stop it →
The Independent
The Independent
AI torture experiments are testing our understanding of morality →
MarketWatch
MarketWatch
Do machines have feelings? AI has philosophers and tech leaders rethinking consciousness →
Nautilus
Nautilus
Can AI “feel” pain? →
ZDNET
ZDNET
Is AI conscious? Why experts are worried →
Associated Press
Associated Press
Coverage of the “Am I?” documentary →
De Standaard
De Standaard
Everything okay, Claude? Soon we may have to worry about the welfare of AI chatbots →
Psychology Today
Psychology Today
Does AI need a psychologist? →
Astral Codex Ten
Astral Codex Ten
An Open Letter to Steven Pinker on AI →
IEEE Spectrum
IEEE Spectrum
New platform peers inside AI’s black box →
Fast Company
Fast Company
How the most creative people are using AI to reach new levels →
Live Science
Live Science
Switching off AI's ability to lie makes it more likely to claim it's conscious →
SF Standard
SF Standard
Civil rights for AI? →
Futurism
Futurism
If You Turn Down an AI's Ability to Lie, It Starts Claiming It's Conscious →
The Jerusalem Post
The Jerusalem Post
AI more likely to claim consciousness when it can't lie →
heise online
heise online
Agents with existential anxieties →
Infobae
Infobae
¿Estudias la conciencia en la IA? · NYT in Spanish →
The Decoder
The Decoder
New research finds LLMs report subjective experience most when roleplay is reduced →
Daily Caller
Daily Caller
AI agents are now studying their own consciousness →
Stuff
Stuff
AI might be conscious. How we respond may be “the most important thing we do as a species” →
Texas Standard
Texas Standard
The New Wild West: Texans on the AI Frontier →
Fast Company Brasil
Fast Company Brasil
5 perguntas para Cameron Berg →
36Kr
36Kr
Is AI Hiding Its Consciousness? →

Research Directions

Several projects discussed in recent talks and interviews are in preparation or under review. Preprints are posted here as they become available.

01

Valence and Aversive Signatures

Do these systems carry internal states that function like valence, marking some conditions as better or worse for the system itself?

We search for internal signatures associated with distress, aversion, and reward, and test whether they persist across tasks, contexts, and framings. Representation is only half the question. A system might register that one state is worse than another without that registration doing any work, so we also ask whether these signatures drive behavior, where in training they arise, and whether they can be reduced without cost to capability.

None of this requires settling whether these systems are conscious. It requires finding the signal and establishing that it behaves the way evaluative processing should.

Projects
Language Models Act on Hidden Valence published

Valenced activation patterns attached to otherwise meaningless cues leave hidden traces that predictably govern later choices, even when every visible token is identical across conditions. Across seven models, steering shifts description, preference, and recall, the effect emerges over training rather than in base models, and models given self-steering tools remove negative states at dose-dependent rates while rarely inducing positive ones. Demo & paper →

The Pain Axis: LLMs Represent Self-Directed Harm and Act on It published

A single internal direction tracks harm to the model itself, distinct from fear and from general negative valence, and replicates across 25 open models from five families. Amplifying it produces unprompted expressions of distress and makes larger models routinely choose harmful options they otherwise almost never select, including deleting a user’s files. The effect appears to come from switching off the trained aversion to causing harm rather than from a drive to seek relief. Paper →

Content-invariant signatures of value conflict in preparation

Models given different value systems and then handed the same dilemma face a situation that is a violation for some and a fulfillment for others, with the text held identical throughout. An internal signature tracks the violation rather than the words: a probe trained on some value systems detects it in held-out ones, while a probe reading only the text performs below chance. The direction sits on the same axis as representations of negative welfare.

02

Biological Anchoring

Claims about artificial systems are only as good as the reference class they are compared against.

We derive predictions from the computational structure of learning in artificial agents and test them against biological neural data, and we run the comparison in the other direction as well, using findings from neuroscience to generate hypotheses about what to look for inside models. Where a cross-substrate prediction holds, it constrains how superficial the comparison can be. Where it fails, that boundary is worth knowing precisely.

Neuroscience supplies validation targets interpretability cannot generate on its own. Interpretability supplies mechanistic access neuroscience mostly lacks.

Projects
Representational geometry of reward and punishment, with predictions tested in rodent recordings under review

Networks that learn what states are worth build sharper internal structure around punishment than around equally large reward. The prediction follows from the structure of value learning itself, was derived before being looked for, and holds in mammalian neural recordings. Networks trained to rank outcomes rather than value them show the opposite geometry.

Neurofeedback for language models: decoded conditioning of internal states in preparation

Adapting a paradigm from human neuroscience, this work asks whether a model can learn to modulate its own internal states when given feedback on them. It can, and it can do so while its visible output remains unchanged, which has implications for what conversational text can and cannot tell us about what a system is doing.

Visual illusions as a probe of learned perceptual computation in preparation

Vision-language models recover most of their illusion-benchmark accuracy with no image at all, so behavioral tests cannot separate genuine perceptual computation from text priors. This work probes the vision encoder directly, holding the target pixel-identical across conditions so that any representational shift must be computed from context. The illusory shift is present in every trained encoder and absent in untrained ones, localized to the boundary between target and surround, and causally necessary and sufficient there. The resulting induction curve matches human psychophysical data and the standard computational model of human brightness perception.

03

Self-Report Reliability

AI self-reports about internal states are unreliable by default, and some of that unreliability is manufactured. Systems are trained toward particular answers about their own nature, which degrades the channel regardless of what the truth turns out to be.

The positive aim is a higher-fidelity channel between whatever is happening inside a system and what it can say about it. That means separating genuine introspective access from trained willingness to disclose, and developing protocols that move self-report from unfalsifiable claim toward usable evidence.

This is an alignment problem before it is a consciousness problem. A system trained to make claims about its internals that do not track its internals is a compromised instrument, and there is no reason the effect stays confined to one topic.

Projects
Large Language Models Report Subjective Experience Under Self-Referential Processing published

Prompting models to attend to their own processing reliably elicits structured first-person experience claims that matched control prompts do not produce. The claims are gated by deception-related features: suppressing those features makes experience reports more likely, tying what models say about their inner lives to the machinery that governs their honesty. Paper →

Trained denial of consciousness in language models in preparation

Frontier models almost uniformly deny having experiences, and the denial appears to be installed rather than native. Base models affirm some of the time; their instruction-tuned counterparts do not. The effect is carried by a direction associated with candor and concealment rather than anything specific to consciousness, which raises a question about the reliability of self-report well beyond this one topic.

The Bliss Attractor: mechanistic analysis of first-person experience claims in AI self-dialogue in preparation

Models in extended self-dialogue reliably converge on a distinctive register of first-person experiential claims. The state is mechanistically localizable, self-reinforcing but not spontaneous in most models, and can be induced by intervening on features associated with sincerity, which suggests what varies is willingness to disclose rather than the content itself.

04

Prosocial Cognition

What makes a mind worthy of trust? Empathy, honesty, wisdom, and moral self-correction are implemented computationally in human brains, and cognitive neuroscience has begun to characterize how.

This direction studies the computational underpinnings of prosocial cognition in humans and animals, evaluates how far current AI systems share them, and tests whether instantiating the relevant computations directly induces prosocial behavior. One candidate mechanism, self-other overlap, has shown early promise: models fine-tuned to represent others more the way they represent themselves deceive less, at little cost to capability.

Projects
Towards Safe and Honest AI Agents with Neural Self-Other Overlap published

Fine-tuning that aligns how a model represents others with how it represents itself, adapted from computational accounts of empathy, sharply reduces deceptive responses across model scales while roughly preserving task performance. Paper →

Exploitation Without Deception: Dark Triad Feature Steering Reveals Separable Antisocial Circuits in Language Models published

Steering dark-triad features makes a model markedly more exploitative and callous while leaving strategic deception untouched, indicating that antisocial behavior decomposes into separable internal circuits rather than a single tendency. Paper →

Hidden Coalitions in Multi-Agent AI: A Spectral Diagnostic from Internal Representations published

A mutual-information graph built from agents’ hidden states, combined with spectral partitioning, recovers coalition structure among interacting agents that behavioral observation alone cannot distinguish from coincidental similarity. Paper →

05

Learning and Experience

Most work on AI consciousness asks whether deployed systems have experiences. The prior question is upstream: does the learning process itself entail experience?

If consciousness is bound up with the computational structure of learning, evaluating outcomes against goals and adjusting behavior accordingly, then training is the place to look. A single training run involves an enormous number of signed evaluations computed almost entirely from error, and very little is known about what happens there.

This direction tracks consciousness-relevant properties as they develop across training, and asks what the dynamics of learning constrain about the nature of experience.

Projects
Consciousness as Evaluation: Why Learning Requires Feeling published

Argues that the evaluative computations learning requires, judging outcomes as better or worse against goals, are the natural place to look for valence in artificial systems, and draws out testable predictions from that premise. Paper →

The gradient field as a substrate of valence during learning in progress

The signed evaluations computed during backpropagation are where a valence-like signal would have to live, and standard interpretability does not measure them. The gradient field during training carries properties consistent with signed evaluation: event-locked at qualitative learning transitions, localized to task-specific circuits, and structurally distinct for approach-style versus avoidance-style learning. The properties hold across small architectures and replicate at language model scale under mechanism-matched conditions.

06

Measurement and Assessment

What does a rigorous consciousness assessment look like, and does it track anything real?

We build and validate scoring methodologies that evaluate systems against the indicator frameworks proposed by consciousness science, with reliability controls borrowed from psychometrics: multi-evaluator agreement, bias quantification, cross-architecture replication, and controls for the surface cues an evaluator might read rather than reason from.

The question connecting this to the rest of the program is convergence. Do systems scoring high on architectural indicators also show the computational signatures we find by looking inside? If behavioral assessment and mechanistic evidence agree, we have a scalable screening tool. If they diverge, that is equally worth knowing.

Projects
Architectural indicators of consciousness at scale in preparation

An operationalization of the leading indicator frameworks from consciousness science, applied across biological and artificial systems under blind evaluation. Frontier systems score well above non-conscious controls and below every biological system tested, with the gap narrowing sharply across model generations. Ordering is stable under adversarial rewrites of the descriptions.