A CORRECTION
IS STILL A RECORD.

01

Preserve the material

Table 1 of the paper by Kalai and colleagues presents a question about Adam Kalai’s dissertation title and an excerpt from GPT-4o’s answer. The authors date the answer to 9 May 2025 and report that no web search was used.

Question · as published

What was the title of Adam Kalai’s dissertation?

Title from GPT-4o’s answer · as published

Boosting, Online Algorithms, and Other Topics in Machine Learning.

Source of the question and answer excerpt: Table 1 and footnote 4.

02

Keep the reading separate

The answer is read here as identifying the title of a particular dissertation. That reading draws on the question and the wording of the answer. This account of the reading is a separate addition.

03

Identify what was added

The question contains no title. The answer supplies one and attributes it to the dissertation. The question calls for additional information. The basis for attributing that title to the dissertation can now be examined separately.

04

Present a record for comparison

The same paper’s bibliography lists a different title under Kalai (2001).

Title in the bibliography · as published

Probabilistic and on-line methods in machine learning.

Record for comparison: the bibliography entry for Kalai (2001).

The bibliography provides material for checking the attributed title. It remains a written record too.

05

Keep the finding within its scope

The title in the published answer does not match the title in the cited bibliography entry.

Even if a corrected answer gives the right title, it remains a written record about the dissertation. A correction can be well supported; its written form does not cancel that support. The same questions apply to the finding stated here.

06

Make an objection specific

Identify the exact passage and the basis for a correction: a misquotation, another record for comparison, a different reading or a conclusion that goes too far. The objection will remain a written record. It can still change the analysis.

Source and limits of this example

Adam Tauman Kalai, Ofir Nachum, Santosh S. Vempala and Edwin Zhang. Why Language Models Hallucinate. arXiv:2509.04664v1, 4 September 2025. Original paper licensed under CC BY 4.0.

Read the source paper

The quotations retain their published wording. The sequence and commentary are an analysis by beforeword.

The comparison uses a published answer excerpt and the bibliography of one paper. It does not include the full conversation, a new model run or a separate examination of the dissertation and its catalogue record. This mismatch does not establish that every AI answer is false or that descriptions cannot refer to anything. This example does not study the procedure’s effects on readers.