Artificial Intelligence (AI), specifically ‘machine learning’, is becoming increasingly embedded in our lives. Tools like ChatGPT, Mistral, Llama, Gemini, Copilot and Claude are changing the way people work by saving them time and improving accessibility. It is such a swiftly moving field spanning many different disciplines, sectors, legalities, applications and activities that it is very difficult to make many definitive assertions about AI as it applies specifically to what oral historians do. Many AI tools simply do some of the things we already do – such as transcription – but more quickly or in different ways.
In this initial OHS discussion paper we explore the risks and benefits of AI tools for oral history and invite debate and comment.
As historians we understand that technological change can also create inequalities, foster elite power structures benefiting only the few, and undermine democratic principles and institutions. So informed decision-making, safeguards, transparency and vigilance are vitally needed.
Many AI tools have trained their ‘large language models’ (LLMs) using copyright-protected materials already online without (in most cases) permission. Creative organisations are now pushing back against these infringements of copyright and intellectual property to protect the livelihoods of creative practitioners. For online oral history-based content which might have been ‘scraped’ to train AI tools for commercial gain, there are obvious risks that this has compromised the original consent of interviewees, infringed copyright and GDPR, and failed to properly acknowledge the authorship of data. This particularly applies to online texts like transcripts, though a tool like Intelion can search audio.
Some benefits of using AI for oral history:
- Transcription. AI voice-recognition tools such as MacWhisper, Otter, Trint, GoodTape and Descript can now generate remarkably accurate speech-to-text transcripts, especially where the audio quality is good, and the speaker is clear and unaccented. Some remote interviewing platforms like Riverside also offer built-in transcription functionality. Many are costed services.
- Documentation and analysis. AI tools can search audio files and large text-based data sources such as oral history transcripts to generate content summaries, indexes and other analytical outputs. They can also facilitate retrospective sensitivity-checking for GDPR and/or accessibility and publication purposes.
- Access and accessibility. AI can search across many data sources and create connections and links to enable aggregated data searches (such as Museum Data Service and Congruence Engine). There are huge research opportunities here. And AI, through such tools as captioning, can make audio more accessible to people with disabilities (for example the hearing-impaired). Some AI ‘copy generators’ can also help you produce project outputs such as publicity and marketing materials, and virtual tours and exhibitions.
Risks of using AI for oral history:
- Security and privacy. Most free AI tools – such as transcription software – will store and/or retain your data. Others will also claim rights over your data, which might infringe data protection (GDPR). Check what the AI tool does with your data, especially if it is in any way sensitive or confidential (much oral history is!). You should resist any rights transfer or sharing, and if possible, only use AI tools locally through applications installed on your own device, rather than through online applications, carefully checking the terms and conditions of any AI software you use. Sensitive or confidential oral history data will simply not be suitable for some AI applications. The risks of malicious creation of ‘fake oral history’ exist but have thus far not been well-documented.
- Accuracy: Data generated by AI tools can be inaccurate and misleading, so it is vital to check any AI-generated content against trusted sources. AI transcription tools might be useful in creating an initial transcript, but in all cases it needs to be carefully checked against the original audio. AI tools are very poor at understanding, for example, silences and emotional actions such as laughing or crying. They also have biases such as ‘hallucinations’, over-generalisations, and outdated language. For some poor-quality audio or accented/dialect speech a traditional human transcriber will be better. Similarly, AI-generated content summaries will rarely be as good as a human-generated version.
- Loss of ‘intelligent’ analysis. Compared to AI tools many oral historians have found that preparing their own summaries and transcripts requires deep listening, a process that is an important part of their research and analysis. The summary writing stage can also be part of the (GDPR) content sensitivity review process. Listening facilitates the identification of themes across interviews, and the selection of good audio extracts for project outputs such as podcasts, publications, web resources and exhibitions. Using speech-to-text/analysis tools side-steps these opportunities for getting to know your recordings.
- Transparency and openness. It is not always self-evident whether data has been AI-generated in all or part. It is vital that AI-originated data (such as an AI transcript) is tagged accordingly so users understand its provenance. We need to call-out AI-generated and manipulated data when its origins are not self-declared, and act quickly on any misuse.
- Bad for the planet. This is a developing area of research, but we know that AI tools use huge amounts of energy to power their data centres, creating significant environmental impact. ChatGPT for example uses 50 to 90 times more energy than standard search engines.
Practical recommended actions when using AI tools for oral history
- If you plan to make any specific use of AI tools, such as transcription software, in your oral history workflow then you need to carry out a careful risk assessment and be transparent about it throughout your project documentation and agreement forms with interviewees. So much commonly used software (such as Adobe Acrobat) uses AI nowadays that a blanket ‘no AI’ declaration on interview participation and agreement forms is not useful. This kind of declaration would have the effect of preventing all online access or computer processing.
- You should check the terms and conditions (and ongoing costs) of any AI tools or applications which you are using to ascertain whether or not the provider will retain a copy of your data and/or any rights in it and for how long. Retention of data or rights over it should be resisted in favour of using AI tools which don’t require uploading your data outside your own local systems.
- Assess the risks relating to the use of AI tools for confidential or sensitive oral history content where third parties might have insecure access to your interviews or might make use of your data in legal jurisdictions not covered by or aligned with UK/EU GDPR.
- Any interview data processed by AI tools needs to be carefully checked against the original audio file for accuracy, and project managers need to put in place rigorous quality control protocols. You should not be sharing or making public any AI-originated data which has not been checked.
Further comment
- We warmly welcome further comment, reflection and input into this evolving area of discussion. Please email [email protected] or contribute to [email protected] (OHS members only).
Some useful resources:
OHS blog on ‘Future oral history’ at https://ohs.org.uk/general-interest/future-oral-history/ includes three useful videos:
‘Doug Boyd on Artificial Intelligence and Oral History’: https://www.youtube.com/watch?v=DOg0iCefZJw
‘Using AI to Analyze and Organize Oral History’: https://deref-gmx.com/mail/client/eUroh7FdwWE/dereferrer/?redirectUrl=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DnAp5aOxQIss
‘Can AI Collect Oral Histories?’: https://deref-gmx.com/mail/client/vNkwbCzH8HE/dereferrer/?redirectUrl=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DNp1SRKUkjkM
British Library blogs about AI:
The challenges of AI for oral history: theoretical and practical issues – Digital scholarship blog
Andrew Flinn, Doug Boyd, Charlie Morgan and Julianne Nyhan. ‘Democratising access? The interface of new technologies and archived oral histories’, Oral History, vol 52 no 3, summer 2024, pp 99-119, online athttps://ohs.org.uk/wp-content/uploads/2024/08/OHJ-52-3-online.pdf
‘AI In OH: How new and evolving technologies will impact the profession’, online [US Oral History Association] Symposium, 15-19 July 2024 at https://oralhistory.org/ai/
Society of Authors AI statement: Artificial Intelligence – The Society of Authors
History Communication Institute AI statement: HCI A.I. Statement
TUC manifesto on ‘Artificial intelligence for creative workers’: https://www.tuc.org.uk/research-analysis/reports/artificial-intelligence-creative-workers
Special issue of AI and Society on the theme of ‘When data turns into archives: making digital records more accessible with AI’: https://link.springer.com/collections/eabccfhhbd
Featured image: https://www.vpnsrus.com/, CC BY 2.0 <https://creativecommons.org/licenses/by/2.0>, via Wikimedia Commons
