“For all the daily hype about exponential improvement, we continue to lack the reliable way of generating high fidelity images from “a deep level of language understanding” that Google promised us three years ago.”
Our reading, the sentence that was checked:
Despite hype about exponential improvement, AI still lacks the reliable deep-language-understanding image generation Google promised in 2022.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Image generation: Still crazy after all these years, paragraph 45.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-image-generation-still-crazy-after-d11e1d-a1
We recorded 5 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 2025-06-25. 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)
Conditions
- despite hype about exponential improvement (qualifier, stated by the speaker)
From the speaker
- we continue to lack
- the reliable way of generating high fidelity images
- a deep level of language understanding
- Google promised us three years ago
Our interpretation
- The plain reading concerns dependable fidelity to a prompt's meaning, including objects, attributes, counts and relationships, across varied intelligible prompts without repeated repairs. It does not require flawless performance.
- Treat the ongoing observation as describing publicly available systems in the current quarter through the post date, April 1–June 25, 2025. Image generation has no standard reporting period; this is an evaluator-selected window for contemporaneous evidence.
- Treat deep language understanding as a claim about observable semantic fidelity, not a separately measurable internal mental state.
- Read promised as referring to Google's advertised capability. Do not assume Google explicitly guaranteed perfect or quantified reliability.
- The reference to exponential-improvement hype is rhetorical background, not a conditional antecedent or a separate measurable growth claim.
Counts as supported if: Contemporaneous evaluations of leading image generators show recurring, material failures to render semantic details across varied prompts, rather than only isolated failures. Google's original description must also substantiate the reference to deep language understanding.
Counts as contradicted if: Contemporaneous broad evaluations establish that an available leading system dependably preserves the semantic details of varied ordinary and unusual prompts, with failures sufficiently rare that recurring unreliability is an inaccurate description.
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 practical reliability claim is supported under the stated plain reading. GPT-4o's compositional benchmark results show substantial remaining weaknesses, and Google's own May 2025 Imagen 4 card acknowledges difficulties with counting, composition and spatial reasoning. June research also finds strong semantic capabilities, so the evidence supports persistent unreliability alongside substantial progress. Google's 2022 paper uses the quoted language-understanding description, but does not make an explicit reliability guarantee. The assessment therefore supports Marcus's observable capability criticism, not a claim that Google broke a quantified promise or that image errors prove an absence of language understanding.
Limitations
- The result depends on an ordinary qualitative reading of reliability.
- Benchmark scores and documented weaknesses do not establish a universal user failure rate.
- The selected research was available as preprints; the Google model card is a company assessment.
- Relative improvement and strong performance on many prompts are compatible with the supported reliability concern.
Evidence (4 items)
Supports the claim: Zhiyuan Yan and colleagues, via arXiv,
0.61 on attribute binding.
Why it matters here: GPT-4o's measured spatial and attribute-binding performance supports recurring semantic failures on a structured benchmark. The same evaluation reports a leading overall score of 0.84 and substantial gains, so it also limits any inference that progress had stopped.
- Publisher
- Zhiyuan Yan and colleagues, via arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Supports the claim: Google DeepMind,
challenging for all models.
Why it matters here: Google documents remaining counting and part-reasoning problems in Imagen 4 and other evaluated strong models. The surrounding section also identifies scale, composition, actions, spatial reasoning and complex language as difficult, supporting the claim's practical reliability concern.
- Publisher
- Google DeepMind
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Context only: Google Research authors, via arXiv,
a deep level of language understanding.
Why it matters here: This verifies the historical language Marcus invokes. It advertises a capability but does not specify a reliability threshold or guarantee correct rendering of every prompt.
- Publisher
- Google Research authors, via arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: Jingjing Chang and colleagues, Shanghai Jiao Tong University and StepFun, via arXiv,
strong capabilities in semantic alignment.
Why it matters here: This is material evidence against interpreting Marcus as denying strong semantic capabilities. Strong relative performance does not establish dependable correctness across varied prompts, and the surrounding analysis describes continuing attribute-assignment and detail errors.
- Publisher
- Jingjing Chang and colleagues, Shanghai Jiao Tong University and StepFun, via arXiv
- 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, Image generation: Still crazy after all these years
For all the daily hype about exponential improvement, we continue to lack the reliable way of generating high fidelity images from “a deep level of language understanding” that Google promised us three years ago.
- Where
- paragraph 45
- Published
- 2025-06-25
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-image-generation-still-crazy-after-d11e1d-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-image-generation-still-crazy-after-d11e1d-a1.