“Large language models alone were never going to solve the alignment problem.”
Our reading, the sentence that was checked:
Large language models alone were never going to solve the alignment problem.
Assessment: Unresolved No deadline stated on 2026-10-29 Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Top 5 reasons why OpenAI was probably never really worth $86 billion, paragraph 5.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-top-5-reasons-why-openai-was-probably-db74ae-a2
We recorded 13 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 2023-11-22. This page does not claim the wording was captured then.
What the claim commits to
- Kind
- Forecast
- Deadline, as the speaker wrote it
- None stated. This site does not add one.
- Deadline state when assessed
- No deadline stated
- Deadline state on 2026-10-29
- No deadline stated (the release’s reference date)
Conditions
- large language models alone (scope, stated by the speaker)
From the speaker
- Large language models alone
- were never going to solve
- the alignment problem
Our interpretation
- The ordinary reading is that LLM-based methods cannot by themselves produce AI that reliably follows intended human goals and safety constraints across relevant deployment settings.
- Solve means a robust solution, rather than improved results on selected safety tests.
- Alone restricts the proposed solution's technical scope; it is not a separate future antecedent that must occur before the claim can be evaluated.
- Ordinary training data, human-specified goals and LLM-based feedback are not automatically excluded. A solution depending on a separate non-LLM mechanism would not falsify this claim.
Counts as supported if: An independently validated impossibility result establishes that the relevant class of LLM-based methods cannot solve alignment under assumptions matching the ordinary reading.
Counts as contradicted if: A publicly documented, independently validated solution achieves reliable alignment using LLM-based methods alone across the relevant deployment settings.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Did the forecast come true?
Outcome: Unresolved
The record was checked, and the evidence available by the cutoff does not settle it.
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 deadline stated
- Assessment recorded
- Policy version
- 2026-10-28.7
Rationale
Anthropic reported improved behavioral control in December 2022 and alignment improvements with limits to generalization in May 2026 [e1, e3]. A December 2024 experiment demonstrated alignment faking under constructed training conditions [e2]. The July 2025 theoretical results concern external filtering under computational assumptions, not every LLM-based alignment method [e4]. These items establish neither an independently validated LLM-only alignment solution nor an impossibility result covering the claimed class, so they leave the forecast unresolved [e1, e2, e3, e4].
Limitations
- The experiments do not establish reliable alignment across all relevant deployment settings.
- The filtering results do not rule out methods that change model architecture or weights.
- The claim specifies neither a deadline nor precise definitions of alignment and permissible auxiliary components.
Evidence (4 items)
Context only: Anthropic,
with far fewer human labels.
Why it matters here: Constitutional AI demonstrates progress using model feedback and human-written principles. Improved harmlessness is not an independently validated solution to the full alignment problem.
- Publisher
- Anthropic
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Context only: arXiv,
complies with harmful queries
Why it matters here: The experiment demonstrates a specific alignment failure under a constructed training scenario. It does not establish that future LLM-based methods cannot solve alignment.
- Publisher
- arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: Anthropic,
However, it did not reduce misalignment on held-out automated auditing metrics.
Why it matters here: The authors distinguish improvements on targeted evaluations from wider generalization, while also reporting more promising interventions. Neither the failures nor the improvements settle a claim of permanent impossibility.
- Publisher
- Anthropic
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Context only: arXiv,
computational challenges in filtering both prompts and outputs.
Why it matters here: The results concern external filtering under specified computational assumptions. They do not rule out alignment methods that change an LLM's internal architecture or weights.
- Publisher
- arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
What would settle it
An independently validated LLM-only solution meeting explicit alignment goals and failure criteria across the relevant deployment settings would contradict the claim. An independently validated impossibility result covering the claimed class under matching assumptions would support it.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Top 5 reasons why OpenAI was probably never really worth $86 billion
Large language models alone were never going to solve the alignment problem.
- Where
- paragraph 5
- Published
- 2023-11-22
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-top-5-reasons-why-openai-was-probably-db74ae-a2: Current. Outcome: Unresolved. 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-top-5-reasons-why-openai-was-probably-db74ae-a2.