“Statistical approximation through generative AI is rarely reliable.”
Our reading:
Statistical approximation through generative AI is rarely reliable.
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
The research documents persistent failures, task-dependent performance, and useful successes. It does not establish that dependable performance was rare across generative AI's ordinary uses. Counting selected failures would overstate the evidence; treating productivity gains as general reliability would also overstate it.
Caveats
- The evidence covers selected language and image tasks rather than the full range of uses.
- Some evaluated models predate those available at publication.
- The conclusion is sensitive to undefined task weights and reliability standards.
Evidence (3 items)
Context only: Ghosh, Hajishirzi and Schmidt / arXiv,
In particular, two object performance has risen in the recent IF-XL and SD-XL models, while position and attribute binding tasks remain difficult overall.
Why it matters: The benchmark finds meaningful differences between simple and compositional image tasks. It supports the existence of reliability problems but does not measure their prevalence across generative AI's ordinary uses.
- Publisher
- Ghosh, Hajishirzi and Schmidt / arXiv
- Kind of source
- Preprint, not peer reviewed
- About events on
- Published
Context only: Liang and colleagues / Transactions on Machine Learning Research, arXiv version,
We observe that the relationship between accuracy and calibration (§4.4: metrics-calibration) depends on the scenario and adaptation procedure (Figure 24, Figure 25).
Why it matters: HELM shows that conclusions depend on the task and evaluation method. Its earlier model sample and language-only scope do not settle the prevalence claim about generative AI at publication.
- Publisher
- Liang and colleagues / Transactions on Machine Learning Research, arXiv version
- Kind of source
- Peer-reviewed paper
- About events on
- Published
Context only: Shakked Noy and Whitney Zhang / MIT,
Our results show that ChatGPT substantially raises average productivity: time taken decreases by 0.8 SDs and output quality rises by 0.4 SDs.
Why it matters: The writing experiment documents useful performance and improved average quality. It is material counterweight to a failure-only account, but productivity gains do not establish dependable performance across uses.
- Publisher
- Shakked Noy and Whitney Zhang / MIT
- 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 reading is that generative AI seldom performs dependably across its ordinary uses. Dependability means consistently meeting the task's requirements, not never making an error.
- Marcus supplies neither a numerical meaning of 'rarely' nor a distribution of tasks. I do not invent either.
- The target is the state of generative AI at publication. There is no regular reporting cycle or stated beginning, so the period has an unspecified start and ends on the post date. Earlier evaluations of relevant systems are admissible; experiments need not have occurred in January.
- Selected failures establish limitations, while selected successes establish capabilities. Neither alone establishes how common dependable performance is.
Right if: Representative evaluations of systems available at publication show that dependable performance is uncommon across ordinary uses, with results robust to reasonable choices of tasks and reliability standards.
Wrong if: Representative evaluations of those systems show that dependable performance is common across ordinary uses, with results robust to reasonable choices of tasks and reliability standards.
We wrote this test after the outcome was known.
What would settle it
Representative evaluations of systems available by January 25, 2024, showing whether dependable performance was uncommon or common under several reasonable task distributions and reliability standards.
We found 11 claims in this post. See them all.
Full record
Source and dates
Marcus on AI, Where’s Waldo? The Elephant in the Room, paragraph 10.
- 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 2024-01-25.
Where it was said (1)
Said once in the posts we read.
: Marcus on AI, Where’s Waldo? The Elephant in the Room
Statistical approximation through generative AI is rarely reliable.
- Where
- paragraph 10
- Published
- 2024-01-25
- 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.