“Making them work reliably in the real world — and not just as demos - will require common sense – about objects, people, and entities in the world.”
Our reading:
AI agents will require common sense about objects, people, and entities to work reliably in the real world.
Assessment: Unresolved
What happened
Did the forecast come true?
Outcome: Unresolved
We checked, and the evidence so far doesn’t settle it.
Judged on a plain reading. They named no measure, so we used the reading a typical informed reader would take. It is spelled out under “How we judged it”.
Evidence up to . Checked .
Why
Deployment evidence collected from April through November 2025 documents constrained autonomy and human verification, without establishing sustained reliability attributable to the agents themselves [e1]. The January 13, 2026 evaluation reports contextual-inference failures in simulated workplace tasks [e2]. The deployed interviewing agent improved from March to August 2026, but the reported improvement establishes neither an absolute reliability standard nor its commonsense competence [e3]. These studies do not pair established reliable operation with separate commonsense tests of the same agents, so they settle neither the necessity claim nor its proposed counterexample [e1–e3].
Caveats
- The evidence does not establish reliable real-world operation under a stated standard while distinguishing agent performance from human corrections [e1, e3].
- Simulated failures do not establish that commonsense competence is necessary for every route to reliable operation [e2].
- The packet contains no qualifying counterexample; it does not establish that none exists [e1–e3].
Evidence (3 items)
Context only: Pan et al., arXiv; accepted at ICML 2026,
strict environmental and autonomy constraints
Why it matters: The study collected deployment evidence from April through November 2025. It documents practical approaches to reliability, but does not independently establish sustained agent performance and separately measure the relevant commonsense competence.
- Publisher
- Pan et al., arXiv; accepted at ICML 2026
- Kind of source
- Peer-reviewed paper
- About events on
- Published
Context only: Ritchie et al., arXiv,
contextual inference
Why it matters: The evaluation identifies failures relevant to Marcus's account of commonsense competence. Its simulated e-commerce tasks neither establish reliable real-world deployment nor show that such competence is universally necessary.
- Publisher
- Ritchie et al., arXiv
- Kind of source
- Preprint, not peer reviewed
- About events on
- Published
Context only: Zhang and Esposito, arXiv,
reliability increased
Why it matters: This is evidence of improvement in actual deployment. A relative improvement does not establish an absolute reliability standard, and the reported result does not test whether the agent lacks relevant commonsense competence.
- Publisher
- Zhang and Esposito, arXiv
- Kind of source
- Preprint, not peer reviewed
- About events on
- Published
How we judged it
- Kind of claim
- Conditional forecast
- Deadline
- None given. We don’t add one.
Conditions
- working reliably in the real world, not just as demos (their words)
Where we had to interpret
- The claim states a necessary condition: agents that work reliably in real-world use must possess relevant commonsense competence.
- I read reliability as consistently completing the tasks entrusted to the agent under ordinary operational variation. It does not mean flawless performance or competence at every possible task.
- I read common sense as practical understanding of relevant entities, their properties and identities, and the implications of ordinary interactions. Competence must be assessed separately from overall task success to avoid circular reasoning.
- The claim requires possessing competence, not a recognized research advance, a particular architecture, or evidence about how that competence developed.
- Human oversight does not automatically disqualify a deployment, but human corrections must be distinguished from the agent's own reliable performance.
- The antecedent is established reliable real-world operation. Deployment alone, relative improvement, and simulated workplace evaluations do not establish it.
Right if: After reliable real-world operation is established, evaluations of those same agents demonstrate relevant commonsense competence across the situations encountered in deployment. Consistent findings across deployments would support the claim within the tested scope, although coexistence alone cannot conclusively establish universal necessity.
Wrong if: An agent consistently performs real-world tasks involving relevant objects, people, or entities, while separate, valid evaluations establish that it lacks the commonsense competence those situations call for. Its reliability must not be attributable to humans supplying the missing judgments.
We wrote this test after the outcome was known.
What would settle it
Dated operational records measuring sustained task completion, errors, ordinary variation, and human intervention under a stated reliability standard, paired with separate commonsense tests of the same agents. Reliable operation despite demonstrated absence of relevant competence, without humans supplying the missing judgments, would refute the claim. Relevant competence across reliably operating deployments would support it within the tested scope, without proving universal necessity.
We found 17 claims in this post. See them all.
Full record
Source and dates
Marcus on AI, AI still lacks “common” sense, 70 years later, paragraph 40.
- Said
- We read the source
- live page, full text, public
- Evidence up to
- Checked
We read this wording on 2026-10-28. The post may have been edited after 2025-01-05.
Where it was said (1)
Said once in the posts we read.
: Marcus on AI, AI still lacks “common” sense, 70 years later
Making them work reliably in the real world — and not just as demos - will require common sense – about objects, people, and entities in the world.
- Where
- paragraph 40
- Published
- 2025-01-05
- We read it
- , live page
- Access
- full text, public
History
Versions
- Current: Unresolved. Evidence up to 2026-10-28.
What happened to this record
PublishedBy us
Published after both AI reviews and the final check passed.
Held backBy us
A reviewer said our test wasn’t fair to what was actually said.