Stefano Palminteri

Do Large Language Models Vindicate Skinner’s Approach to Language? What large language models teach us about language, learning, and the long-running debate between Skinner and Chomsky

including historical querelles 🧐🙂https://medium.com/@stefano.palminteri/do-large-language-models-vindicate-skinners-approach-to-language-b8a323682b46

🚨 Updated preprint 🚨

Excited to share this updated preprint, in which @ruimata.bsky.social and I discuss five ways LLMs can help address longstanding challenges in cognitive science and psychology, examining both opportunities and pitfalls.

🔗 arxiv.org/abs/2511.00206

Beyond computational equivalence: the behavioral inference principle for machine consciousness Abstract. Large Language Models (LLMs) have rapidly become a central topic in AI and cognitive science, due to their unprecedented performance in a vast ar

🚨 New article in #NCONSC

Beyond computational equivalence: the behavioral inference principle for machine consciousness
academic.oup.com/nc/article/2...

@stepalminteri.bsky.social

#consciousness
🧠🧪

Charley's amazing laboratory is hiring!

A week ago, we had an open conversation amongst senior academics in leadership roles and their use of LLMs (in both their leadership capacities and individually). Small group, not representative, but it already taught me a great deal about the diversity of perspectives (& their consequences). 1/

Beyond computational equivalence: the behavioral inference principle for machine consciousness Abstract. Large Language Models (LLMs) have rapidly become a central topic in AI and cognitive science, due to their unprecedented performance in a vast ar

🧠🤖 New paper on AI & consciousness

We challenge the idea that computational equivalence is the right test for machine consciousness. In cognitive science, consciousness is never “read off” from code, it’s inferred inductively from behavior

with @thecharleywu.bsky.social

doi.org/10.1093/nc/n...
👇

We solve the p-zombie problem in the box of the paper. You can give it a look if you want and time, but in short p-zombies are a false problem

Publicly wondering if your product might be conscious is good hype @anthropic.com, but poor science.
Consciousness attribution should integrate prior expectations with public/shareable facts, not by computational equivalence alone or anecdotes. If that makes us "vulgar behaviorists", we are proud.

When evaluating AI risk or AI welfare, consciousness may not be the most relevant starting point; agency and goal-directedness are more immediately pertinent. The article argues that “the questions of consciousness and danger are often confounded,” even though an AI system may be highly intelligent and goal-directed without being conscious. From an ethical and safety perspective, “the control problem and other existential AI safety issues are perhaps better addressed not by discussing consciousness, but rather their capacity for agency,” particularly as reinforcement-learning systems increasingly shape behavior.

When evaluating AI risk or AI welfare, consciousness may not be the most relevant starting point; agency and goal-directedness are more immediately pertinent.

Applying consciousness criteria to LLMs will likely require stringent, theory-informed behavioral tests, explicitly accounting for priors, mimicry, and architectural opacity. Because LLMs are trained to display behaviors typically associated with consciousness, attribution is complicated by “gaming or mimicry,” reflecting Goodhart’s law that “a measure ceases to be a good measure once it becomes a target.” The Behavioral Inference Principle therefore integrates priors about training and architecture and relies on “behavioral tests that are specifically designed to be diagnostic of the computational process of interest.”

Applying consciousness criteria to LLMs will likely require stringent, theory-informed behavioral tests, explicitly accounting for priors, mimicry, and architectural opacity.

The Behavioral Inference Principle frames consciousness attribution as the inference of a latent computational process from behavior, grounded in the inductive methodology of cognitive science. Under the Behavioral Inference Principle, “consciousness is attributed if it is useful to explain (and predict) a given set of behavioral observations,” treating consciousness as a latent construct rather than a directly observable property. This approach explicitly follows inference to the best explanation, echoing Harman’s formulation that when we infer mental states, “we are inferring that the latter fact explains better than some other explanation what he does” (Harman, 1965)

The Behavioral Inference Principle frames consciousness attribution as the inference of a latent computational process from behavior, grounded in the inductive methodology of cognitive science.

Computational equivalence is not a viable requirement for machine consciousness: the relevant computations are neither agreed upon nor readable in black-box AI systems. The paper argues that computational equivalence presupposes both “an explicit hypothesis regarding the structure of the target computation” and “a computational architecture of the candidate system [that] is transparent enough to allow one to verify the presence (or absence) of the target computation.” In practice, neither condition is met: there is no consensus on the computations underlying consciousness, and LLMs are “extremely complex black-box systems… whose computational processes cannot be directly inspected but must be inferred from their behavior.”

Computational equivalence is not a viable requirement for machine consciousness: the relevant computations are neither agreed upon nor readable in black-box AI systems.

Within computational functionalism, the dominant view holds that “an artificial system can be said to be conscious if it processes information by implementing the same computational processes that characterize consciousness in other systems already known to possess this capacity.” This position is explicitly defended by many, including Schneider and colleagues, who argue that “the system processes information in a way analogous to how a conscious human or non-human animal would respond when in a conscious state »

Computational equivalence is frequently assumed to be the appropriate route for attributing consciousness to artificial systems.

Beyond computational equivalence: the behavioral inference principle for machine consciousness Abstract. Large Language Models (LLMs) have rapidly become a central topic in AI and cognitive science, due to their unprecedented performance in a vast ar

🧠🤖 New paper on AI & consciousness

We challenge the idea that computational equivalence is the right test for machine consciousness. In cognitive science, consciousness is never “read off” from code, it’s inferred inductively from behavior

with @thecharleywu.bsky.social

doi.org/10.1093/nc/n...
👇

Feedback-induced attitudinal changes in risk preferences - Nature Communications Normative theory predicts that feedback should not affect decisions under risk, but past findings disagree. Here, the authors show that feedback shifts risk-taking by changing attitudes rather than th...

Here it is share.google/M9fHctRwDOqU...

More broadly, our findings challenge the idea that feedback is a panacea for improving rationality, suggesting instead that feedback often reshapes preferences and emotions rather than correcting beliefs: an insight with important implications for decision-making research, behavioral interventions, and the long-standing debate on human rationality. Our findings also raise a cautionary tale for data-driven approaches to modeling, as focusing on quantitative prediction rather than fine-grained, design-based analyses may lead to diagnostic behavioral signatures being overlooked.

More broadly, our findings challenge the idea that feedback is a panacea for improving rationality, suggesting instead that feedback often reshapes preferences and emotions rather than correcting beliefs: an insight with important implications for decision-making research, behavioral interventions.

Key behavioral signatures of our results (namely, the first trial effect and the negativz recency effect) are not captured by the influential BEAST model, highlighting a critical limitation of learning-based accounts that attribute feedback effects to experience-driven value updating.

Key behavioral signatures of our results (namely, the first trial effect and the negativz recency effect) are not captured by the influential BEAST model, highlighting a critical limitation of learning-based accounts that attribute feedback effects to experience-driven value updating.

Trial-by-trial analyses further rule out reinforcement learning, as participants exhibit negative recency (a gambler’s fallacy–like pattern), becoming less likely to repeat a risky choice after positive outcomes rather than more.

Trial-by-trial analyses further rule out reinforcement learning, as participants exhibit negative recency (a gambler’s fallacy–like pattern), becoming less likely to repeat a risky choice after positive outcomes rather than more.

Critically, the effect emerges from the very first trial (before any outcome can be experienced) indicating that feedback operates through attitudinal mechanisms. This attitude changes is driven by epistemic curiosity under partial feedback and anticipated regret under complete feedback.

Critically, the effect emerges from the very first trial (before any outcome can be experienced) indicating that feedback operates through attitudinal mechanisms.

This attitude changes is driven by epistemic curiosity under partial feedback and anticipated regret under complete feedback.

We find that feedback robustly increases risk-taking, while leaving expected-value maximization unchanged (if ever slightly impaired!). These counterintuitive results go against the received wisdom that feedback (through learning) corrects biases and improves decision-making.

We find that feedback robustly increases risk-taking, while leaving expected-value maximization unchanged (if ever slightly impaired!).

These counterintuitive results go against the received wisdom that feedback (through learning) corrects biases and improves decision-making.

Across seven incentivized experiments and a large reanalysis, we systematically manipulated the presence and type of post-choice feedback in repeated risky decisions to test whether feedback shapes behavior through learning mechanisms or through anticipatory changes in preferences.

Across seven incentivized experiments and a large reanalysis, we systematically manipulated the presence and type of post-choice feedback in repeated risky decisions to test whether feedback shapes behavior through learning mechanisms or through anticipatory changes in preferences.

Very happy that @PNASNews agreed to publish our (w/ @romanececchi.bsky.social) response to Prakhar's thought-provoking study! You can find the final version at the link below. See the following tweet for Prakhar's response to our response. Happy to hear your thoughts!
www.pnas.org/doi/10.1073/...

Reply to Cecchi and Palminteri: On the need to model temporal variation in learning rates | PNAS Reply to Cecchi and Palminteri: On the need to model temporal variation in learning rates

Here it is! www.pnas.org/doi/10.1073/...

Please consider following early-career researchers @romanececchi.bsky.social and @prakhargodara.bsky.social to stay updated on their future amazing research!

Very happy that @PNASNews agreed to publish our (w/ @romanececchi.bsky.social) response to Prakhar's thought-provoking study! You can find the final version at the link below. See the following tweet for Prakhar's response to our response. Happy to hear your thoughts!
www.pnas.org/doi/10.1073/...

I had a great time in Cambridge, great people, great science and lovely place. Thanks a lot to @orbenamy.bsky.social @georgiaturner.bsky.social @sjblakemore.bsky.social

lovely walk and chat with a great mentor and friend

Thank you for your comments and engagement!

Good point. I guess that in Fabien's paper the uncertainly is internal (do I know it? or not?) while in lotteries it is external. In one case curiosity could be overshadowed by the fear of being wrong, in the second less so (after all the outcomes are determined "mechanically").

... reward integration. In our case anticipated emotional states (the pleasure of resolving uncertainly, anticipated regrat) seem to influence behavior BEFORE rewards are experience (and therefore before learning can occurr). So I guess again they are not much related?

Hi Nicole, at the micro level Bennett and co. model is about how mood shape learning: in our case we demonstrate that there is no learning at all. So in that case I think it is fair to say that the works are not much related. At the "macro" level, they show that emotion/mood influence...