ChatGPT Corrects 'strawberry' R-count but Errors Persist

9to5Google reports that OpenAI posted that "at long last," ChatGPT can correctly answer how many times the letter "R" appears in the word "strawberry." 9to5Google says the same coverage documents other examples of so-called "confident mistakes," including instances where ChatGPT reportedly replies that "cranberry" has one "R" and where the model previously recommended walking to a car wash 50 meters away after a logically malformed prompt.

Companies working with large language models commonly face persistent "confident mistake" behavior, also called hallucinations, where the model asserts incorrect facts with high confidence. Industry-pattern observations: teams often deploy targeted, hardcoded patches or prompt-layer rules to fix high-visibility failures; these fixes can close individual test cases but do not always generalize across similar inputs or underlying failure modes.

Industry context: for practitioners, the ChatGPT example underlines two predictable patterns, first, surface-level correctness improvements can be implemented quickly and publicized; second, a constellation of related edge cases typically remains. That combination matters for product teams and evaluators because it affects trust metrics, test coverage design, and escalation criteria for model outputs in user-facing flows.

Observers should track whether OpenAI publishes technical notes, a changelog, or unit test coverage for the fix, and whether community reports of related miscounts decline on developer and social channels. Industry-pattern observations: durable remediation is usually visible when a fix survives diverse paraphrases and adversarial rephrasings rather than only canonical prompts.