“And like other generative AI systems, there is no way to encode (and guarantee) constraints like “be truthful” or “obey the laws of physics”or “don’t just invent (or eliminate) objects”.”
Our reading:
Sora, like other generative AI systems, has no way to encode and guarantee constraints such as truthfulness or obeying physics.
Assessment: Unresolved
What happened
Was it true for the period described?
Outcome: Unresolved
We checked, and the evidence so far doesn’t 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 “How we judged it”.
Evidence up to . Checked .
Why
Sora's documented failures show that its original outputs lacked consistent physical correctness. They do not establish that there was no way to encode and guarantee the named properties. The theoretical papers address narrower formal questions and do not settle Sora's constraint capabilities. The turbulence GAN enforces a specialized conservation constraint, so it does not contradict the contextual claim about general real-world generation. Neither side meets the stated criteria.
Caveats
- An unsuccessful search for a guarantee does not prove that none was possible.
- Published qualitative failures do not establish an architectural impossibility.
- The assessment concerns methods available in February 2024, not later model improvements.
Evidence (4 items)
Context only: OpenAI,
physically implausible motion
Why it matters: The original model's outputs did not consistently obey physical expectations. A failure of that implementation does not establish the absence of any applicable enforcement method.
- Publisher
- OpenAI
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- About events on
- Published
Context only: arXiv; Ziwei Xu, Sanjay Jain, and Mohan Kankanhalli,
a formal world
Why it matters: The paper argues that LLMs cannot learn every computable ground-truth function without errors. Applying that result to Sora's physical or object-continuity constraints requires an additional argument that this source does not supply.
- Where
- Abstract, version 1
- Publisher
- arXiv; Ziwei Xu, Sanjay Jain, and Mohan Kankanhalli
- Kind of source
- Preprint, not peer reviewed
- About events on
- Published
Context only: arXiv; Dima Tretiak, Arvind T. Mohan, and Daniel Livescu,
hard constraints
Why it matters: The method enforces mass conservation in a GAN for incompressible turbulence. That establishes a specialized constraint mechanism, not a guarantee of general real-world physics in a Sora-like generator.
- Where
- Abstract
- Publisher
- arXiv; Dima Tretiak, Arvind T. Mohan, and Daniel Livescu
- Kind of source
- Preprint, not peer reviewed
- About events on
- Published
Context only: arXiv; Adam Tauman Kalai and Santosh S. Vempala,
may mitigate
Why it matters: The impossibility result concerns particular facts under a calibration condition, while other errors may be mitigated. It neither supplies a general truthfulness guarantee nor establishes that Sora cannot enforce the named constraints.
- Publisher
- arXiv; Adam Tauman Kalai and Santosh S. Vempala
- Kind of source
- Preprint, not peer reviewed
- About events on
- Published
How we judged it
- Kind of claim
- Observation
- Deadline
- None given. We don’t add one.
Where we had to interpret
- The ordinary contextual reading concerns the original Sora and comparable general-purpose generative systems, across their intended real-world uses.
- The claim concerns enforceable truthfulness, physical consistency, or object continuity, rather than an improvement in average output quality.
- A guarantee must cover the relevant property throughout the system's intended domain. Enforcing one conservation equation in a specialized scientific generator does not establish that a Sora-like system obeys real-world physics.
- This is a claim about the technical position at the time of the post, not a prediction that no future method could ever succeed.
- The target period is February 15–16, 2024, the latest public Sora announcement and demonstration period before the post.
- The distinction between lacking an existing guarantee and having no way to implement one is material. Output failures alone establish only the former.
Right if: Architectural analysis or sufficiently comprehensive technical evidence establishes that Sora and comparable general-purpose systems then lacked a way to encode and enforce these broad real-world constraints throughout their intended outputs.
Wrong if: A method available by the post demonstrably encodes and guarantees at least one of the named broad properties in Sora or a comparable general-purpose system across its intended domain. A narrow formal constraint or improved benchmark score is insufficient.
We wrote this test after the outcome was known.
What would settle it
Technical evidence addressing the original Sora and comparable systems: either an applicable analysis establishing the stated enforcement limitation, or a method available by February 16, 2024 that demonstrably guaranteed a named broad property across the intended domain.
We found 15 claims in this post. See them all.
Full record
Source and dates
Marcus on AI, Sora’s Surreal Physics, paragraph 22.
- 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 2024-02-16.
Where it was said (1)
Said once in the posts we read.
: Marcus on AI, Sora’s Surreal Physics
And like other generative AI systems, there is no way to encode (and guarantee) constraints like “be truthful” or “obey the laws of physics”or “don’t just invent (or eliminate) objects”.
- Where
- paragraph 22
- Published
- 2024-02-16
- 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 wasn’t fair to what was actually said.