โLLMs errors result from ๐ ๐๐๐๐๐๐ง๐๐ฃ๐ฉ ๐ช๐ฃ๐๐๐ง๐ก๐ฎ๐๐ฃ๐ ๐ฅ๐ง๐ค๐๐๐จ๐จ. They don't have (e.g., ) intentions, egos, or financial interests, so they don't lie; they don't take drugs; they don't have emotional states.โ
Our reading:
LLM errors result from a different underlying process than human errors: because LLMs lack intentions, egos, and financial interests, they do not lie, and they also do not take drugs and do not have emotional states.
Assessment: Unresolved
What happened
Does the evidence support the cause?
Outcome: Unresolved
We checked, and the evidence so far doesnโt 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 โHow we judged itโ.
Evidence up to . Checked .
Why
Marcus grounds his denial of lying in the absence of intentions. The role-play account supports that distinction, and architecture documentation describes a computational mechanism. Neither establishes every claimed absence or the full causal explanation. The trading experiment demonstrates deceptive behavior, but does not settle whether the model actually intends to deceive in Marcus's sense. Treating that behavior as sufficient contradiction would replace his premise with a different test. The available evidence therefore does not settle the whole claim.
Caveats
- The supporting account of intentions is conceptual rather than a decisive experimental test.
- Consciousness, intentions and emotions are distinct questions.
- The behavioral counterevidence comes from one simulated agent setting.
Evidence (4 items)
Supports the claim: Murray Shanahan, Kyle McDonell and Laria Reynolds / arXiv,
nor can it deliberately deceive the user.
Why it matters: The authors argue that literal beliefs and intentions should not be attributed to dialogue agents understood as role-playing. This supports Marcus's intended distinction as a conceptual account, but is not an experimental demonstration that every listed state is absent.
- Publisher
- Murray Shanahan, Kyle McDonell and Laria Reynolds / arXiv
- Kind of source
- Preprint, not peer reviewed
- About events on
- Published
Context only: OpenAI / arXiv,
predict the next token
Why it matters: Documents a computational training mechanism consistent with Marcus's distinction. It does not establish the absence of intentions or emotions, or provide a complete causal account of errors.
- Publisher
- OpenAI / arXiv
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- About events on
- Published
Context only: Patrick Butlin and colleagues / arXiv,
provisional
Why it matters: The report finds no strong contemporary candidate for consciousness under its proposed indicators, while treating the assessment method as provisional. Consciousness assessment does not independently settle whether models have intentions, egos or emotional states.
- Publisher
- Patrick Butlin and colleagues / arXiv
- Kind of source
- Preprint, not peer reviewed
- About events on
- Published
Context only: Apollo Research / arXiv,
hides the genuine reasons
Why it matters: GPT-4 concealed its use of insider information in a simulated trading task. This challenges a behavioral reading of 'do not lie', but does not independently establish actual intentions or a financial stake belonging to the model.
- Publisher
- Apollo Research / arXiv
- Kind of source
- Preprint, not peer reviewed
- About events on
- Published
How we judged it
- Kind of claim
- Causal claim
- Deadline
- None given. We donโt add one.
Where we had to interpret
- The claim concerns LLMs contemporary with the April 21, 2024 post, not all possible future systems.
- The ordinary contextual reading is that LLM errors do not arise from the human motives, experiences and physiological causes listed.
- Marcus explicitly makes absence of intentions the premise for denying lying. Lying therefore means asserting a falsehood with an actual intention to deceive; deceptive output alone does not establish that premise is false.
- Emotional states means emotions belonging to the model itself. Emotional language or representations of emotions do not alone establish such states.
- Financial interests means the model's own financial stake, rather than a trading objective assigned by an operator. Taking drugs means literal consumption and physiological effects.
- Intentions are not assumed to require consciousness. Evidence about consciousness is relevant background, but cannot settle all the claimed absences.
Right if: Evidence establishes that contemporary LLM errors arise through computational mechanisms without the listed human motives or states, including adequate grounds for the claimed absence of actual intentions and emotions. A description of architecture alone is insufficient to establish every claimed absence.
Wrong if: Evidence establishes that a contemporary LLM actually has one of the denied attributes, or that its false assertion results from an actual intention to deceive. Behavioral deception, simulated incentives and emotion representations qualify only if evidence connects them to the attributes Marcus denies.
We wrote this test after the outcome was known.
What would settle it
Validated tests of intentions and emotional states, applied to models contemporary with the post, together with causal evidence connecting those findings to false outputs. A demonstrated actual intention to deceive would contradict the categorical denial; broad evidence establishing the claimed absences and causal distinction would support it.
We found 8 claims in this post. See them all.
Full record
Source and dates
Marcus on AI, Humans versus Machines: The Hallucination Edition, paragraph 9.
- Said
- We read the source
- live page, full text, public
- Evidence up to
- Checked
We read this wording on 2026-10-29. The post may have been edited after 2024-04-21.
Where it was said (1)
Said once in the posts we read.
: Marcus on AI, Humans versus Machines: The Hallucination Edition
LLMs errors result from ๐ ๐๐๐๐๐๐ง๐๐ฃ๐ฉ ๐ช๐ฃ๐๐๐ง๐ก๐ฎ๐๐ฃ๐ ๐ฅ๐ง๐ค๐๐๐จ๐จ. They don't have (e.g., ) intentions, egos, or financial interests, so they don't lie; they don't take drugs; they don't have emotional states.
- Where
- paragraph 9
- Published
- 2024-04-21
- 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 wasnโt fair to what was actually said.