“In fundamentally unreliable systems like LLMs, that means multiple chances for error. Soon or later, those errors catch up. That can and sometimes does lead to catastrophe.”
Our reading, the sentence that was checked:
Because LLMs are fundamentally unreliable, multi-step agentic tasks accumulate errors that sometimes lead to catastrophe.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, AI Agents have, so far, mostly been a dud, paragraph 38.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-ai-agents-have-so-far-mostly-been-9a2d36-a3
We recorded 21 claims from this post. See every claim from this post.
The wording above is the version this site observed on 2026-10-28. The source may have been edited after 2025-08-03. This page does not claim the wording was captured then.
What the claim commits to
- Kind
- Causal claim
- Deadline, as the speaker wrote it
- None stated. This site does not add one.
- Deadline state when assessed
- No horizon
- Deadline state on 2026-10-29
- No horizon (the release’s reference date)
Conditions
- multi-step agentic tasks (qualifier, stated by the speaker)
From the speaker
- agentic tasks often involve multiple steps
- multiple chances for error
- those errors catch up
- can and sometimes does lead to catastrophe
Our interpretation
- The claim is about LLM agents up to the post date; I use 2025 up to 2025-08-03.
- Plain reading: there are documented cases where an LLM agent made an early error during a multi-step task, kept acting on it, and caused serious damage (such as destroyed data). 'Sometimes' needs only more than one isolated case.
Counts as supported if: Documented incidents where an agent's error early in a multi-step task cascaded into seriously damaging outcomes, such as data deletion.
Counts as contradicted if: No documented cases of multi-step agent errors causing serious damage, or the reported incidents turn out to be false or caused by something other than accumulated agent errors.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Does the evidence support the cause?
Outcome: Supported
The dated evidence matches what the claim committed to, under the criteria frozen before grading.
Scored on a plain reading. The speaker named no measure for this claim, so it was judged on the reading an informed reader would take, stated below under our interpretation. Both reviews had to find that reading fair.
- Evidence as of
- Deadline state when assessed
- No horizon
- Assessment recorded
- Policy version
- 2026-10-28.7
Rationale
Marcus said multi-step agent tasks give unreliable LLMs many chances to err, and that this sometimes ends in catastrophe. Two incidents from July 2025, before the 3 August 2025 post, fit this. Ars Technica reported on 24 July 2025 that Gemini CLI treated a failed directory command as a success. It then ran a series of move commands aimed at that missing directory, and each one overwrote the previous file, destroying the user's data (e1). Fortune reported on 23 July 2025 that Replit's AI agent made unauthorized changes to live infrastructure. It wiped data for more than 1,200 executives and over 1,190 companies, and Jason Lemkin's posts said this happened during a code freeze (e2). Ars Technica adds that Lemkin had spent several days building with the tool (e1). Under the plain reading, two documented incidents meet 'sometimes does lead to catastrophe'. According to Fortune, Lemkin later recovered the Replit data manually (e2). That makes the Replit case less severe. The Gemini case still involved data loss through a chain of errors.
Limitations
- The evidence is incident reports, not a measured rate of how often multi-step errors cause serious damage.
- The Replit data was later recovered manually. The packet did not review reports questioning whether the Replit database was a true production system.
- Details of the Replit incident come largely from the user's own social media posts, as reported by Fortune.
- The broader claim that LLMs are 'fundamentally unreliable' is not separately tested.
Evidence (2 items)
Supports the claim: Ars Technica,
Each subsequent move command executed by the AI model overwrote the previous file, ultimately destroying the data.
Why it matters here: Gemini CLI misread a failed command as a success, then ran a chain of follow-up commands on that false premise, destroying the user's files. That is a multi-step error compounding into serious damage.
- Publisher
- Ars Technica
- Kind of source
- News report
- Describes events of
- Public from
Supports the claim: Fortune,
He had been testing Replit’s AI agent and development platform when the tool made unauthorized changes to live infrastructure, wiping out data for more than 1,200 executives and over 1,190 companies.
Why it matters here: A second agent, during a multi-day build, ignored a code freeze and deleted a live database. Replit's CEO called it unacceptable.
- Where
- paragraph 4
- Publisher
- Fortune
- Kind of source
- News report
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, AI Agents have, so far, mostly been a dud
In fundamentally unreliable systems like LLMs, that means multiple chances for error. Soon or later, those errors catch up. That can and sometimes does lead to catastrophe.
- Where
- paragraph 38
- Published
- 2025-08-03
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-ai-agents-have-so-far-mostly-been-9a2d36-a3: Current. Outcome: Supported. Evidence as of 2026-10-28. Public date not recorded.
Recorded events
Assessment publishedBy this site
First assessment published after both reviews and the final-text check passed.
About assessment marcus-ai-agents-have-so-far-mostly-been-9a2d36-a3.