“As most experts realize, the reality is that current AI doesn’t “decide to teach itself”, or even have consistent beliefs.”
Our reading, the sentence that was checked:
Current AI does not decide to teach itself and does not have consistent beliefs.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Stop Treating AI Models Like People, paragraph 3.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-stop-treating-ai-models-like-people-db97c6-a1
We recorded 9 claims from this post. See every claim from this post.
The wording above is the version this site observed on 2026-10-29. The source may have been edited after 2023-04-17. This page does not claim the wording was captured then.
What the claim commits to
- Kind
- Observation
- 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)
From the speaker
- current AI
- doesn’t “decide to teach itself”
- or even have consistent beliefs
Our interpretation
- The context concerns mainstream conversational language models, particularly ChatGPT, GPT-4 and Bard, rather than every experimental AI architecture.
- The ordinary behavioral reading is that these systems do not independently initiate persistent learning and do not reliably maintain consistent answers across equivalent questions. This does not establish whether they possess subjective beliefs.
- For the ongoing state, I use January 1 through April 17, 2023: the recent months discussed in the post and the latest publicly documented systems before publication.
- Learning information within a supplied conversation is distinct from independently initiating changes to the trained model.
Counts as supported if: Contemporary technical documentation describes externally controlled training without autonomous learning from use, and reproducible tests show inconsistent answers to equivalent inputs.
Counts as contradicted if: Contemporary documentation or reproducible tests demonstrate independently initiated persistent learning in the named chatbot class, or establish the consistent positions that the claim denies.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Was it true for the period described?
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
The behavioral reading is supported. GPT-4's technical report describes externally controlled training and says it does not learn from experience. A contemporary study found that ChatGPT changed predictions across equivalent inputs. These records support the practical warning against treating chatbot answers as stable personal beliefs. They do not settle a philosophical claim about subjective mental states or a universal claim about every AI architecture.
Limitations
- The result depends on reading beliefs behaviorally and current AI in the chatbot context.
- Consistency can hold on particular tasks even when it fails elsewhere.
- The training evidence is a developer disclosure, not an independent audit.
Evidence (2 items)
Supports the claim: OpenAI,
fine-tuned using Reinforcement Learning
Why it matters here: The report describes externally organized training and explicitly states in its limitations that GPT-4 does not learn from its experience. This supports the operational reading of the self-teaching clause.
- Publisher
- OpenAI
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Supports the claim: arXiv; Myeongjun Jang and Thomas Lukasiewicz,
not completely trustworthy regarding semantic consistency
Why it matters here: Tests of the January 30 ChatGPT version found inconsistent predictions for paraphrased inputs. The same paper reports improved negation consistency, so it does not show that every answer or position was inconsistent.
- Publisher
- arXiv; Myeongjun Jang and Thomas Lukasiewicz
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Stop Treating AI Models Like People
As most experts realize, the reality is that current AI doesn’t “decide to teach itself”, or even have consistent beliefs.
- Where
- paragraph 3
- Published
- 2023-04-17
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-stop-treating-ai-models-like-people-db97c6-a1: 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-stop-treating-ai-models-like-people-db97c6-a1.