“An AI tool merely guesses what code looks right, given the prompt it was provided.”
Our reading:
An AI coding tool merely guesses what code looks right, given the prompt it was provided.
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 evidence supports a narrower point: code generation can produce insecure or incorrect suggestions. It does not establish that the generator merely selects plausible-looking code without substantive analysis. Research also demonstrates measurable performance on program-behavior tasks, which cuts against that stronger reading. Those studies do not directly establish the behavior of the relevant Copilot configuration during the target period. Neither bundled scanners nor the fact of probabilistic prediction settles the claim as written.
Caveats
- Most usable evidence concerns research models or earlier configurations.
- Benchmark performance does not establish human-like understanding or guarantee secure software.
- The distinction implied by 'merely' remains underspecified.
Evidence (3 items)
Context only: Chen et al., arXiv,
may suggest insecure or otherwise bad code.
Why it matters: The paper describes next-token prediction and acknowledges insecure suggestions, supporting the narrower concern that generation does not guarantee correctness. It studies earlier models and does not establish that 2024 generators lack substantive program analysis.
- Publisher
- Chen et al., arXiv
- Kind of source
- Preprint, not peer reviewed
- About events on
- Published
Context only: Gu et al., arXiv,
75% and 81%
Why it matters: Predicting program inputs and outputs tests behavior beyond whether generated code looks plausible. These earlier model results are material counterevidence to a blanket denial of code reasoning, but do not directly measure the target-period Copilot configuration.
- Publisher
- Gu et al., arXiv
- Kind of source
- Preprint, not peer reviewed
- About events on
- Published
Context only: Ding et al., arXiv,
54.5% on CRUXEval-I
Why it matters: The researchers trained a code model to represent execution behavior and evaluated code generation and input prediction. This challenges an unrestricted surface-plausibility account, but a research model is not direct evidence about Copilot, and the abstract does not date the experiments precisely.
- Publisher
- Ding et al., arXiv
- 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.
Conditions
- AI coding tools such as GitHub Copilot (their words)
Where we had to interpret
- In context, this concerns the code-generating model in assistants such as GitHub Copilot, contrasted with a programmer examining what code does.
- The plain reading is that the generator produces plausible code without substantively understanding or checking whether it solves the problem securely. It does not mean random guessing or that generated code can never work.
- A bundled scanner does not contradict this claim. Nor does probabilistic generation alone establish the word 'merely': prediction and useful analysis of program behavior are not necessarily incompatible.
- For this undated ongoing observation, use April–June 2024, the latest completed calendar quarter immediately before publication. Earlier research is background unless it establishes the behavior of tools during that period.
Right if: Evidence about the relevant generators establishes that their behavior is limited to surface plausibility, without substantive analysis of program behavior or correctness. Documentation of prediction and examples of errors alone establish only narrower parts of the claim.
Wrong if: Reproducible evaluations of the relevant generators demonstrate substantive analysis of program behavior, such as tracing execution or identifying and explaining defects on unfamiliar examples. The capability must come from the model itself; a separate scanner or occasional correct completion is insufficient.
We wrote this test after the outcome was known.
What would settle it
A dated, reproducible evaluation of the relevant generator, with external tools disabled, testing execution reasoning and defect analysis on unfamiliar code with controlled changes to program behavior.
We found 29 claims in this post. See them all.
Full record
Source and dates
Where's Your Ed At, The Shareholder Supremacy, paragraph 109.
- Said
- We read the source
- live page, full text, public
- Evidence up to
- Checked
We read this wording on 2026-10-28. The post may have been edited after 2024-07-01.
Where it was said (1)
Said once in the posts we read.
: Where's Your Ed At, The Shareholder Supremacy
An AI tool merely guesses what code looks right, given the prompt it was provided.
- Where
- paragraph 109
- Published
- 2024-07-01
- 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.