Pundit ClaimsAI claims by Gary Marcus and Ed Zitron, with the evidence

Gary Marcus Marcus on AI Published

Spinal Tap meets John Malkovich

We recorded 4 claims from this post. 2 were checked against evidence. 2 can’t be checked. 3 more are held. Read the post at its source.

Claims from this post (4)

In the order they appear in the post. Each quotation is the speaker’s words in this post. A label such as “Supported” or “Can’t be checked” is the claim’s current assessment; its page has the evidence and the reasons.

  1. Assessment: Can’t be checkedSaid

    “These systems can draw things that kinda look like the world, but they obviously don’t know what’s what or how things work outside of their training data.”

    What we tested: Generative image systems can produce world-like images but do not understand what things are or how they work outside their training data.

  2. Assessment: SupportedSaid

    “No model could have done this in 2022. The graphics are vastly better than they used to be.”

    What we tested: Image generation graphics quality has improved vastly since 2022, to a level no 2022 model could reach.

    What happened: In August 2022, CyclingTips tested Stable Diffusion, Midjourney, DreamStudio, Craiyon and Enstil on cycling prompts. It judged most of the realistic cycling images unusable, which fits the claim that 2022 tools could not reliably draw riders on correct cycles. (Velo / CyclingTips, )

  3. Assessment: Can’t be checkedSaid

    “Because they live in the land of correlations between images and text, not in the land of ideas.”

    What we tested: The image-generating AI systems under discussion are unpredictably right or wrong because they operate in the land of correlations between images and text, not in the land of ideas.

  4. Assessment: UnresolvedSaid

    “Watch the tuners come and go! So much for object permanence.”

    What we tested: Grok's generated video failed to keep objects consistent, with the bass tuners appearing and disappearing.

Held from this post (3)

These records did not pass every check, so nothing about them is published. Held is not a finding.

  • The two reviews reached different assessments. Our reading, not assessed: When the image-generating AI systems under discussion are run many times, they are sometimes right and sometimes wrong, and you never know which result you will get.
  • The criteria are ambiguous. Our reading, not assessed: The model that produced the rebuttal image (ChatGPT-5 Thinking) was clearly better than the model whose unicycle image Marcus had ridiculed.
  • The two reviews reached different assessments. Our reading, not assessed: Better graphics in the image-generating AI systems under discussion do not necessarily mean a better understanding of the world.