Whisper

Risks and harms 🛑

Whisper has been associated with significant potential and actual harms, including: 

• Bias and discrimination. Whisper may perpetuate any manner of biases present in its training data.

• Cultural insensitivity. Whisper's performance varies across different languages and dialects, resulting in what can appear culturally insensitive and disrepctful outputs.

• Loss of active listening skills. Over-reliance on Whisper may reduce active listening skills.

• Factual inaccuracies. Whisper can produce transcription errors, especially with accented speech, background noise, or technical terminology, leading to misinformation and disinformation, and poor decision-making. Such errors can have a serious impact in high-risk domains, including law, healthcare and medicine, and politics and government.

• Dual/multiple use. Whisper can easily be misused for mass surveillance and other malicious purposes.

• Copyright abuse. OpenAI has not disclosed Whisper’s training sources, raising concerns that the system may violate third-party copyright.

• Privacy loss. Voice data collected by Whisper could potentially be to identify and profile individuals captured it its training data.

• Employment. The extensive use of Whisper and other automated transcription tools risks replacing human jobs.

Transparency and accountability 🙈

• Data sources. Critics have pointed out the lack of transparency regarding the datasets used to train Whisper, particularly concerning indigenous languages.

• User support. OpenAI does not provide ongoing support or integration assistance for users implementing Whisper, leaving them to manage issues independently - a lack of accountability that can lead to challenges in the ethical and effective use of the model.

• Environment. OpenAI have not revealed details of the energy or water consumed in Whisper's training or of the carbon emissions associated with its training and use.

Research, advocacy 🧮

• Koenecke A. et al. Careless Whisper: Speech-to-Text Hallucination Harms

• Keoni Mahelona, Gianna Leoni, Suzanne Duncan, Miles Thompson. OpenAI's Whisper is another case study in Colonisation

News, commentary, analysis 🗞️

• https://www.tomsguide.com/ai/meet-whisper-web-a-new-and-free-way-to-transcribe-audio

• https://www.techtimes.com/articles/305605/20240612/openai-transcription-tool-whisper-generate-harmful-violent-text-hallucinations.htm

• https://www.nytimes.com/2024/04/06/technology/tech-giants-harvest-data-artificial-intelligence.html

• https://www.healthcare-brew.com/stories/2024/11/18/openai-transcription-tool-whisper-hallucinationsQ

• https://interestingengineering.com/culture/why-did-whisper-take-a-million-hours-of-youtube-videos