AgentsAutonomy & tool use
Michael Tomasello argues on Machine Learning Street Talk that LLMs have human-like language but lack the shared agency it grew from
On Machine Learning Street Talk, Duke professor Michael Tomasello argues that large language models gained human-like language without the shared-agency foundations it evolved from in humans: joint goals, joint attention and perspective-taking. On this view, fluent conversation is weak evidence of real agency.
From the Machine Learning Street Talk podcast episode of November 19, 2026 and cited sources: Tomasello says LLMs 'started at the endpoint.' They have the language humans use but skipped the shared agency underneath it, and in his view they do not create conventions. [1]
SOURCES[1] MLST episode page (Spotify) podcasters.spotify.com
AI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- Tomasello says LLMs 'started at the endpoint.' They have the language humans use but skipped the shared agency underneath it, and in his view they do not create conventions. [1]
- Original source
- Machine Learning Street Talk podcast episode, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 19, 2026 podcast episode published
- Covered
- Nov 20 edition
- Evidence
- 9 sources across 8 sites; no transcript
- 1MLST episode page (Spotify)podcasters.spotify.com
- 2The Agency Gap, CogSci 2026 (eScholarship)escholarship.org
- 3The Evolution of Agency (Penguin Random House)penguinrandomhouse.com
- 4How to make artificial agents more like natural agents…scholars.duke.edu
- 5The Evolution of Agency (MIT Press Bookstore)mitpressbookstore.mit.edu
- 6Michael Tomasello (Wikipedia)wikipedia.org
Reach: not measured yet
Tomasello says LLMs 'started at the endpoint.' They have the language humans use but skipped the shared agency underneath it, and in his view they do not create conventions. He also argues a Turing-style test falls far short as a test of agency. This is a theoretical argument that no reviewed source tests directly. [1]
His framework sets out agency as evolved layers. Goal-directed agency came first in early vertebrates, intentional agency in mammals, rational agency in apes and shared agency in humans. Each layer adds a capacity: control loops with executive oversight, simulating outcomes before acting, seeking information before committing, then goals shared with others. He proposes this as a design model for more natural artificial agents. [1] [3] [4] [5]
A 2026 CogSci study used a non-linguistic coordination game to compare people working with people and people working with AI partners. Adults and 7-year-olds paired with AI partners showed less commitment and coordinated less efficiently, even when they still succeeded at the task. Tomasello co-wrote the study, so it supports his position but does not confirm it independently. [2]
one study found weaker commitment and less efficient coordination in human–AI pairs.
Read the full assessment
Implication: teams building collaborative agents may need measures beyond task success, and fluent conversation says little about whether a system can share goals with people.
Executive brief
Michael Tomasello argues that large language models "started at the endpoint." In his account, they have human-like language without the shared-agency infrastructure that language grew out of in humans: joint goals, joint attention and perspective-taking. He makes the case on a Machine Learning Street Talk episode published 19 November 2026, according to the episode description. Some measured data backs the idea. A 2026 CogSci paper he co-wrote found that people working with AI partners showed less commitment and coordinated less efficiently than people working together, even when the outcomes were good (eScholarship). The episode is commentary. No independent coverage of it was found.
What changed and event timeline
Book sets out four kinds of agency
The Evolution of Agency (MIT Press) describes a sequence: goal-directed agency in early vertebrates, intentional agency in mammals, rational agency in apes and shared agency in humans ().
The argument is aimed at AI
In Trends in Cognitive Sciences 29(9), Tomasello argues that efforts to make AI more human-like could learn from how biological agent architectures evolved ().
Testing humans with AI partners
Cheng, Gao, Tomasello and Kushnir used a non-linguistic coordination game to compare human–human and human–AI collaboration ().
The MLST interview
Tim Scarfe and Tomasello test the theory on a traffic light, a thermostat and a robot vacuum. They then discuss shared agency, conventions and why a Turing test falls far short as a test of agency ().
Capabilities and access
- This is not a model release. No model or version is evaluated in the episode materials.
- The episode is free on Spotify and other podcast platforms, with a chapter list (episode page).
- The Agency Gap paper is open access (eScholarship). The TICS article is DOI 10.1016/j.tics.2025.07.004 (Duke Scholars).
Technical analysis for researchers and developers
- When the environment is predictable, evolution hard-wires behaviour. When it is not, evolution builds a decision-maker that runs control loops with executive oversight (episode description).
- Each stage adds a capacity on top of the last. Humans add shared goals with others (Penguin Random House).
- The Agency Gap game separates whether a pair succeeds from how they coordinate, measuring commitment and efficiency.
Read the full section
- Design principle. When the environment is predictable, evolution hard-wires behaviour. When it is not, evolution builds a decision-maker that runs control loops with executive oversight (episode description).
- Layered architecture. Each stage adds a capacity on top of the last. Mammals simulate outcomes before acting, like a squirrel picturing a fall. Apes go looking for more information before they commit. Humans add shared goals with others (Penguin Random House).
- Evaluation method. The Agency Gap game separates whether a pair succeeds from how they coordinate, measuring commitment and efficiency. It tested adults and 7-year-olds with several AI partners (eScholarship). The design is worth reusing for testing agents that work with people.
Claims and evidence
- Claim: LLMs skipped the infrastructure of shared agency and do not create conventions. This is Tomasello's theoretical argument as summarised in the episode description.
- Claim: humans commit less and coordinate less efficiently with AI partners. This was measured.
- Claim: a four-stage evolutionary sequence. This is a theoretical synthesis (MIT Press Bookstore).
Read the full section
- Claim: LLMs skipped the infrastructure of shared agency and do not create conventions. This is Tomasello's theoretical argument as summarised in the episode description. No reviewed source tests it directly.
- Claim: humans commit less and coordinate less efficiently with AI partners. This was measured. Shared agency is "largely absent" in standard reinforcement learning agents, and the paper calls its status in LLMs "unclear" (eScholarship). Tomasello co-authored the study, so it supports his own position and is not independent confirmation of it.
- Claim: a four-stage evolutionary sequence. This is a theoretical synthesis (MIT Press Bookstore).
Context and prior work
- Tomasello is a professor at Duke and Emeritus Director at the Max Planck Institute for Evolutionary Anthropology in Leipzig (Penguin Random House; Wikipedia).
- His earlier research on joint attention, norms and roles is the basis for his shared-agency argument (Duke Scholars profile).
- He has also discussed apes and children with other interviewers (Many Minds; Mindscape).
Read the full section
- Tomasello is a professor at Duke and Emeritus Director at the Max Planck Institute for Evolutionary Anthropology in Leipzig (Penguin Random House; Wikipedia).
- His earlier research on joint attention, norms and roles is the basis for his shared-agency argument (Duke Scholars profile).
- He has also discussed apes and children with other interviewers (Many Minds; Mindscape).
Limitations, safety and contested findings
- The Agency Gap results come from one minimal coordination game. The abstract does not establish whether they carry over to tasks that use language.
- Adults and children often adapted to AI partners and still succeeded.
- No independent replication or published rebuttal was found among the reviewed sources.
Read the full section
- The Agency Gap results come from one minimal coordination game. The abstract does not establish whether they carry over to tasks that use language.
- Adults and children often adapted to AI partners and still succeeded. The gap the paper reports is in how people coordinated, not whether they succeeded (eScholarship).
- No independent replication or published rebuttal was found among the reviewed sources.
Business and practitioner implications
- If you are building agents that collaborate with people, measure commitment and coordination efficiency, not just whether the task gets done (eScholarship).
- Treat fluent language as weak evidence of agency. Tomasello argues a Turing-style test is far from sufficient (episode description).
- Research teams looking for new agent designs could draw on evolutionary architectures: control loops, simulating outcomes before acting, and seeking information before committing (Duke Scholars).
Read the full section
- If you are building agents that collaborate with people, measure commitment and coordination efficiency, not just whether the task gets done (eScholarship).
- Treat fluent language as weak evidence of agency. Tomasello argues a Turing-style test is far from sufficient (episode description).
- Research teams looking for new agent designs could draw on evolutionary architectures: control loops, simulating outcomes before acting, and seeking information before committing (Duke Scholars).