Persona Fails on Vertical IDs
The article discusses failures of AI persona systems when handling vertical identification documents (e.g., driving licenses, ID cards). It highlights that existing systems often break or produce poor results on vertical IDs due to training data biases toward horizontal layouts, leading to inaccurate data extraction and user experience issues.
Background
- The "persona rule" is an unwritten norm that prepended text (e.g., "I'm sorry, but I cannot...") should match the language of a user's prompt. There is no official rule from any AI company requiring this; it emerged informally in online AI communities and is inconsistently enforced.<br>- "Vertical IDs" are government-issued identification cards (driver's licenses, state IDs) that are physically oriented vertically rather than horizontally. In the US, vertical IDs have been the standard for under-21 IDs since the 2010s, while horizontal IDs are for those 21 and older.<br>- Many e-commerce, alcohol delivery, and age-verification systems have struggled to process vertical IDs because software was originally trained only on horizontal IDs, leading to false rejections. This is an ongoing practical problem in automated age verification.<br>- The blog post mketab.org draws a parallel between this real-world ID-scanning failure (where vertical IDs are incorrectly rejected) and a pattern in large language models where the "persona rule" (prepended text matching language) causes models to incorrectly refuse English prompts that include a non-English word or name.