Over reaching “safety layer” affecting legitimate workflow

Addressing the Impact of Safety Layer Restrictions on Critical Workflow: A Case Study from The Strangers Project

Introduction

In the realm of participatory art and community storytelling, authenticity and accessibility are paramount. The Strangers Project, an acclaimed initiative spanning over 16 years, exemplifies this ethos by collecting and sharing more than 100,000 anonymous handwritten stories from diverse individuals. These narratives encompass a broad spectrum—from lighthearted anecdotes to deeply personal accounts involving sensitive topics such as trauma, abuse, and mental health struggles.

The Role of Transcription in Preserving Artistic Authenticity

To ensure these stories reach a wider audience and are preserved digitally, transcription becomes an essential step. Initially, manual transcription was employed, but as the volume grew, reliance on Automated Optical Character Recognition (OCR) technologies seemed promising. Traditional OCR often struggles with varied handwriting styles, leading to inaccuracies.

Fortunately, recent advancements have integrated large language models like GPT, which can generate remarkably accurate transcriptions, significantly streamlining the workflow. This integration has been vital, enabling efficient archiving, curation, and dissemination of these personal narratives.

The Challenge: Safety Layer Interference

Recently, the project encountered a significant obstacle. GPT’s built-in safety filters, designed to prevent the generation or dissemination of harmful or sensitive content, began to interfere with transcription tasks. Specifically, when processing stories that reference topics such as suicide, abuse, or trauma—common in honest storytelling—the model increasingly refused to produce direct transcripts. Instead, it added disclaimers, altered outputs, or outright declined to transcribe.

While these safety mechanisms are understandable in casual or public contexts, they pose a serious issue for professional archival work that relies on faithful reproduction of original content. The core problem lies in the safety layer restricting the model’s ability to accurately transcribe sensitive but legitimate stories, which in this context are vital to preserving authenticity.

Implications for Workflow and Automation

Beyond manual transcription, the project aims to automate large-scale processing using GPT’s API. However, if these safety filters are active at the API level, they could impede automated workflows, making it difficult or impossible to efficiently transcribe thousands of stories without manual intervention or risk of content suppression.

A key concern is whether API calls invoking GPT encounter the same safety restrictions as the web-based interface. If so, this would significantly hinder the project’s scalability and efficiency.

Seeking Solutions and Workarounds

The creator of The Strangers Project is seeking advice on potential workarounds—such as ways to toggle off safety