Gary Marcus: rules, reasoning and the search for trustworthy AI

How a learner reaches beyond examples

Marcus’s route into AI began with questions about human language. His 1993 MIT doctorate, supervised by Steven Pinker, examined rules and exceptions in inflectional morphology: how words change their form. The title already contained a tension that would recur throughout his work. Learning involves remembering particular cases, but also finding relationships that remain useful when the cases change.

At New York University, where he is now Professor Emeritus, that concern developed into experimental work on early cognition. Studying infants offered a demanding way to investigate learning before fluent language could conceal how much a learner had understood.

What infants can tell us about rules

In a 1999 Science paper, Marcus, S. Vijayan, S. Bandi Rao and P. M. Vishton reported three experiments with seven-month-old infants. After hearing syllable sequences arranged in repeating patterns, infants listened longer to sequences that violated the familiar structure, even when the test syllables were new. The authors interpreted this as evidence that infants could extract and extend abstract relationships, rather than merely remember familiar strings.

The distinction mattered for computational models. Yet the paper explicitly limited its criticism to particular neural architectures; it did not rule out neural implementations of symbolic operations. Nor did looking-time differences reveal the brain’s machinery directly. The experiment supplied evidence about behaviour, from which competing mechanisms still had to be evaluated.

Subsequent research made the picture richer. In 2007, with Keith Fernandes and Scott Johnson, Marcus reported that prior exposure to patterns in speech helped infants learn corresponding patterns in nonspeech sounds. A later meta-analysis and experiment by Hugh Rabagliati, Brock Ferguson and Casey Lew-Williams emphasised the relevance of meaningful stimuli across domains. The broader lesson is that generalisation depends on both the learner’s capacities and the conditions under which those capacities are tested.

Building a bridge between neurons and symbols

The Algebraic Mind (2001) developed a constructive account of how neural systems might support operations over symbols. A variable is useful because it stands for something that can change: a rule can apply to a new object without having to be relearned for that object alone. Marcus’s proposal was that the organisation of a neural system matters, including its ability to represent such reusable relationships.

His books also explored the limits of human cognition. The Birth of the Mind examined how genetic processes contribute to mental development, while Kluge treated the mind as an imperfect product of evolution. Together they complicate any simple appeal to human intelligence as a flawless engineering blueprint. In Guitar Zero (2012), learning became personal: his efforts to learn guitar provided a setting for considering practice, musical skill and adult learning.

Taking ideas into companies

Marcus helped found Geometric Intelligence in October 2014 with Zoubin Ghahramani, Kenneth Stanley and Douglas Bemis. Uber’s announcement of 5 December 2016 records its acquisition and the creation of Uber AI Labs, with Marcus as the initial director. The venture brought cognitive science into a commercial research setting. The acquisition itself establishes an institutional milestone, rather than a scientific demonstration that one learning approach had surpassed another.

He was also a founder and early chief executive of Robust.AI, a robotics company established in 2019. Those are historical roles. The company announced Marin Tchakarov’s appointment as CEO in August 2026; its later products and leadership should not be read as a description of Marcus’s current responsibilities.

A programme for dependable intelligence

His 2018 critical appraisal of deep learning acknowledged advances in perception and games while raising concerns about data requirements, generalisation and reasoning. The argument was that further components would be needed for general intelligence. Its status is that of a research critique, rather than a theorem setting permanent limits on every future neural model.

With Ernest Davis, Marcus brought the case for common sense and understanding to a wider audience in Rebooting AI (2019). His 2020 research proposal made the constructive agenda more explicit: combine learning and symbolic operations, acquire substantial knowledge, reason with it, and maintain cognitive models of situations. These models would represent entities and their relationships so that a system could interpret events and anticipate consequences. They are related to the broader ambition of modelling the world, but are not synonymous with learned video predictors or the latent dynamics used in reinforcement learning.

In a 2023 paper with Doug Lenat, he explored what language models might gain from Cyc’s tradition of explicit knowledge and inference. The proposal envisaged integration: generative abilities could be combined with structured reasoning and traceable premises. It also recognised the computational difficulty of expressive logic. A valid inference remains dependent on its premises and on how accurately an everyday problem has been represented.

The argument continues in 2026

With Vaishak Belle, Marcus co-authored a perspective published in the AAAI proceedings in March 2026. It surveys several ways of joining neural learning with symbolic reasoning, including logical constraints, knowledge representations and external solvers. This is a family of approaches, with different assumptions and engineering demands. The authors also acknowledge that hybridisation alone cannot guarantee trustworthy general intelligence.

Their position invites empirical comparison. Brenden Lake and Marco Baroni’s 2023 Nature study demonstrated human-like systematic generalisation in a controlled task through a neural network trained by meta-learning. That result challenges blanket claims that neural systems cannot generalise compositionally, while leaving open how broadly the ability transfers. Progress must be judged by specified tasks and evidence, rather than by treating either “neural” or “symbolic” as a guarantee of success.

Marcus also joined the large author group behind A Definition of AGI (2025), which proposes evaluating AI across multiple cognitive domains. It connects his longstanding interest in human cognition with the problem of measuring machine competence. The proposed framework is one approach to evaluation, not an agreed universal definition of intelligence.

From reliability to public responsibility

On 16 May 2023, Marcus testified at the United States Senate’s hearing on AI oversight, alongside Sam Altman and Christina Montgomery. In Taming Silicon Valley (2024), he extended his concerns into proposals for data rights, oversight and public accountability. The book is an intervention in policy and public debate; its prescriptions express a position about how AI should be governed.

Read across these different activities, his career turns on a persistent demand: an intelligent system should be able to carry what it learns into unfamiliar circumstances, and its reliability should be examinable. The enduring research question is how to build that capacity. His work helps formulate the question, while the success of any proposed answer remains something to demonstrate.

A career in seven moments

1. 1993 · Rules and exceptionsCompletes his doctorate at MIT under Steven Pinker.
2. 1999 · Learning in infancyCo-authors experiments on generalising patterns to new syllables.
3. 2001 · The Algebraic MindDevelops a case for neural systems capable of symbolic operations.
4. 2014–2016 · Geometric IntelligenceCo-founds the company, later acquired by Uber to form the core of its AI laboratory.
5. 2019–2020 · A constructive AI agendaWorks in robotics, publishes with Ernest Davis and sets out a programme for robust intelligence.
6. 2023–2024 · Knowledge and accountabilityWrites with Doug Lenat, testifies before the Senate and publishes Taming Silicon Valley.
7. 2025–2026 · Evaluation and hybrid systemsContributes to a proposed AGI evaluation framework and writes with Vaishak Belle on neuro-symbolic AI.