AI on the job: does correcting an answer correct its effects?
An assistant may produce an incorrect explanation that is subsequently corrected but has already entered a team’s reasoning. The useful question is specific: once an AI answer has circulated, how can managers check whether its correction also changes the conclusions that depend on it?
What the literature establishes
Hollyn M. Johnson and Colleen M. Seifert (1994, Journal of Experimental Psychology: Learning, Memory, and Cognition) investigated the continued influence of information after its retraction. In their experiments, initial information could still inform inferences even after participants had received a correction. Withdrawing a claim and removing its influence on reasoning are therefore distinct operations.
The explanatory gap left by a correction
Stephan Lewandowsky, Ullrich K. H. Ecker, Colleen M. Seifert, Norbert Schwarz and John Cook (2012, Psychological Science in the Public Interest) review research on misinformation and its correction. Their review highlights, among other mechanisms, the difficulty of abandoning an explanation when no alternative makes sense of the event. A correction may therefore be more useful when it provides a substantiated alternative explanation rather than simply stating that the earlier claim was false.
The premature conclusion
One might conclude that asking AI for a new explanation is enough to cancel out the first. That would confuse a cognitive mechanism with a guarantee of reliability: the replacement explanation must itself be checked. When the cause remains unknown, it is better to make that uncertainty explicit and reconsider the affected conclusions than to manufacture narrative continuity. (our executive and employee training programmes)
What these studies do not demonstrate
These publications do not test generative assistants in professional teams in Basel. They cannot establish either how frequently this persistence occurs at work or its magnitude in regulated activities. They do, however, support an operational distinction: confirming that a document has been corrected does not prove that reasoning based on its first version has been revised.
A practical check in Bâle
An SHR “IA on the job” workshop in Basel could use a fictional case linking pharma, life sciences, Rhine logistics and international R&D: an AI-assisted summary incorrectly attributes a sample’s delay to a transport restriction. After a correction supported by the case materials, ask participants individually to restate their explanation of the delay and their recommendation, without access to their first answer. Measure the proportion of answers that still use the invalidated restriction as a premise, using a rubric defined before the exercise and comparing them with the initial answers. This observation does not measure a productivity gain; it checks whether the correction has actually reached the reasoning rather than merely the document. To go further: explore the AI on the Job training in Basel, or browse our executive and employee training programmes in Switzerland.
In pictures: AI on the Job in Basel
- ai on the job
- Basel
- research
- training basel