[Air-L] OpenAI & Rev transcriptions: Ethics concern?

[kalev leetaru]

As a general note, it is important to note the massive difference between
newer non-deterministic ASR's like Whisper that prioritize "fluency over
fidelity" and classical and modern deterministic large model ASRs. In
short, with models like Whisper, the transcription will vary every time you
run the model and will often deviate from what the speaker actually said if
it is a statistically unlikely utterance, as they are designed to
essentially rewrite the utterances into a more readable and understandable
text, as opposed to capture what was actually said. For a live video
conference where understandability trumps accuracy, that might be just
fine, but for interview transcriptions to be used for content analysis and
even for providing the most accurate transcriptions for assistive needs,
keep those issues in mind. Also remember that models like Whisper can
hallucinate and truncate *badly* and recommended mitigations often don't
work on many content types:

https://blog.gdeltproject.org/a-deep-dive-exploration-applying-openais-whisper-asr-to-a-russian-television-news-broadcast/
https://blog.gdeltproject.org/a-deep-dive-exploration-applying-openais-whisper-asr-to-a-pbs-newshour-broadcast/
https://blog.gdeltproject.org/testing-the-new-openai-whisper-asr-large-v2-model-on-a-russian-tv-news-broadcast/
https://blog.gdeltproject.org/openais-whisper-asr-how-the-nato-threat-to-putin-becomes-putins-threats-to-nato-the-challenges-of-machine-translation/
https://blog.gdeltproject.org/experiments-with-whisper-asr-model-parameters-non-determinism-temperature_increment_on_fallback/
https://blog.gdeltproject.org/ais-pivot-to-fluency-existential-non-determinism-vs-fbis-bbcm-osint-the-centrality-of-fidelity-in-translation/

The same holds for translation - keep in mind that the newer generative
translation models like Meta's Seamless have a number of critical
departures from what we think of as traditional NMT:

https://blog.gdeltproject.org/experiments-with-speech-transcription-translation-metas-seamlessm4t-vs-openais-whisper-vs-gcps-sttgt-vs-gcps-usm-chirpgt/

And keep in mind that if you're using LLMs to summarize or otherwise
distill transcripts, hallucination will manifest in highly
unpredictable ways based on the alignment of the source topics and the
training data:

https://blog.gdeltproject.org/hallucination-in-summarization-when-chatgpt-hallucinated-new-stories-in-an-evening-news-broadcast/

Embedding models can be used for mitigation for monolingual tasks:

https://blog.gdeltproject.org/using-embedding-ranking-to-combat-llm-hallucination-in-generative-summarization-the-abc-news-chinese-spy-balloon-story/

But keep in mind that in multilingual tasks, even bitext models and the
most recent LLM embedders will still favor stilted NMT translations over
human translations, so they are counter-productive for multilingual
mitigation contexts:

https://blog.gdeltproject.org/authoritative-human-vs-nmt-llm-translation-embedding-based-quality-rankings-nmt-skew/

Kalev


On Wed, Sep 6, 2023 at 7:06 PM Hamlet Lopez via Air-L <
air-l at listserv.aoir.org> wrote:

> Hi Natalie,
>
> I won't talk about the ethics involved here. But you can get high
> quality automatic transcription of your interviews for free, from your
> personal computer, if you are not afraid of console programs. You can
> use open source software.
>
> Check https://github.com/ggerganov/whisper.cpp
>
> I use whisper.cpp for interviews in Spanish, with excelent results.
> The English language models are even better.
>
> Best wishes,
>
> Hamlet
>
> On 9/6/23, Natalie Rock via Air-L <air-l at listserv.aoir.org> wrote:
> > Ethics mavens, your insights would be appreciated: should I be concerned
> > about this note sent from Rev.com? I use Rev to create automated
> transcripts
> > of research interviews. Does this note imply my transcripts might be used
> > for AI training?
> >
> >
> > Begin forwarded message:
> >
> >> From: Your Team at Rev <yourteam at rev.com>
> >> Date: September 6, 2023 at 14:08:15 PDT
> >>
> >> In your Terms of Service and Data Processing Agreement with Rev, we
> commit
> >> to notifying you before any new third-party sub-processor starts
> >> processing applicable data. We’re sending this message because we’re
> >> planning to appoint a new third-party sub-processor for Rev to support
> an
> >> upcoming new feature.
> >>
> >> OpenAI - OpenAI is a company focused on AI research and deployment.
> Their
> >> API platform provides models, tools, and technologies for AI
> development.
> >> These services are utilized to enhance the efficiency, scalability, and
> >> flexibility of both human and Automatic Speech Recognition (ASR)
> services.
> >> This new third-party sub-processor has been verified to ensure it meets
> >> Rev security and privacy standards and will meet Rev’s data processing
> >> terms. No customer owned data will be provided to this third party until
> >> the sub-processor notification period has passed.‌
> >>
> >>
> > _______________________________________________
> > The Air-L at listserv.aoir.org mailing list
> > is provided by the Association of Internet Researchers http://aoir.org
> > Subscribe, change options or unsubscribe at:
> > http://listserv.aoir.org/listinfo.cgi/air-l-aoir.org
> >
> > Join the Association of Internet Researchers:
> > http://www.aoir.org/
> >
>
>
> --
> Dr.C. Hamlet López García
> Investigador
> Instituto Cubano de Investigación Cultural
> "Juan Marinello"
> Profesor
> Universidad de la Habana
> _______________________________________________
> The Air-L at listserv.aoir.org mailing list
> is provided by the Association of Internet Researchers http://aoir.org
> Subscribe, change options or unsubscribe at:
> http://listserv.aoir.org/listinfo.cgi/air-l-aoir.org
>
> Join the Association of Internet Researchers:
> http://www.aoir.org/
>