Skip to main content

Treasure Hunt with Generative AI: The Deep Dive Podcast

| By | Alcohol History,News,Resources

Have you ever heard of the Alcoholism Treatment Digest? An innovative publication for its time, the Alcoholism Treatment Digest (ATD) was a periodical founded by Mark Keller and published by the Center of Alcohol Studies (CAS) from 1950 to 1973 at Yale and Rutgers. Each issue featured short review articles on emerging topics in alcohol literature prepared by the staff of the editorial department of the Quarterly Journal of Studies on Alcohol (QJSA). Each review summarized multiple scholarly articles on a specific topic related to the latest advances in alcoholism treatment and rehabilitation. Primarily intended for physicians and other healthcare professionals, ATD was a subscription-based publication, typed and mimeographed. It was distributed five times a year to subscribers, who were encouraged to republish the reviews in their publications with proper attribution to QJSA.

MicrophoneFollowing the founders of modern-day alcohol studies to disseminate information innovatively, I have been experimenting with new advancements in another emerging field: Artificial Intelligence (AI). The result is a podcast, Deep Dive: A Treasure Hunt in Alcohol Studies History, presenting an fresh (albeit controversial) approach to starting a conversation about substance use by sharing knowledge about the early decades of an evolving field.

Selecting from our own previously published scholarly articles, blog posts, website contents, and notes, I trained Google’s NotebookLM on these texts to develop podcasts on topics of general interest. Each AI-generated episode explores historically significant topics from a fresh perspective, presenting them in an engaging style to appeal to fans of the podcast genre. A fitting demo for the experiment covers the Alcoholism Treatment Digest.

NotebookLM

NotebookLM is an experimental AI app developed by Google. To train AI, users can upload documents in an individual notebook as an information source and use the documents and the content AI generated as resources.

My workflow was as follows.

  • Starting at the “Sources” tab, I uploaded pre-selected sources (formats supported are PDF, txt, mp3, linked website or YouTube video, up to 50 by notebook), or pasted texts. For most topics, I chose about 2-5 source documents (higher numbers resulted in hallucinations, such as one more wife for Jellinek, although very specific: a German woman he married in Italy). AI immediately created a summary of the documents, which is editable for content.
  • Using the “Chat” function, AI generated notes related to each responding to questions suggested by NotebookLM and prompts added by me. After reviewing each note for accuracy and relevance, I selected the most valuable ones to convert into a new source, which was automatically added to “Sources.” If necessary, a new note can be added here at any time. I also took advantage of the option to create FAQs as a source. With minor editing for content and emphasis, it provided a decent overview of the topic.
  • Next, I used the “Audio Overview” function under “Chat” to create the podcasts. AI can automatically convert content from the sources into a podcast-style dialog. To generate a more engaging but accurate podcast, one can give prompts at this point, e.g., defining what the two speakers should focus on. Processing takes only a few minutes, and the end product varies from decent (i.e., shareable after minor edits) to hilariously useless (i.e., not even close, with lots of hallucinations) to be discarded. Unfortunately, the audio overview can’t be regenerated from the same sources; creating a new notebook is the only way to remediate.
  • Most recently, a “Studio” tab was also added to NotebookLM, which assists with organizing content by providing a quick overview of each notebook’s content.

Adobe Podcast

Audio overviews that needed only minor edits were eventually turned into podcasts with Adobe Podcast. Instead of working with the audio waveform file, I chose transcription-based editing, currently available in beta from Adobe Podcast. It allows for further improvement in sound quality, even though the original recording was sufficient. A trained mix engineer may opt for other Adobe apps (PremierPro, Audition) or Audacity and waveform, but I chose the convenience of a single app.

The simple workflow was as follows.

  • After uploading the audio file, Adobe transcribed it into a semi-editable text file. By that, I mean limited text editing, i.e., deleting fillers (such as uhm, mm-hmm, like, you know, right, absolutely, etc.) so prevalent in podcasts (but sounding creepy if you remember that these are robots making small talk). I got rid of quite a few irrelevant and unnecessary words or sentences while listening to the audio, which resulted in the final version of the podcast. I wish individual words could be replaced in the transcript by typing rather than just being removed, which is an option with similar, subscription-based AI software applications such as WondercraftAI.
  • To make it sound more podcast-like, I opted to use, albeit sparingly, the various music options offered by the software: intro, outro, background, and transition, fading in and out. The end product is an engaging, “eerily realistic” convo between two millennial voices, a male and female, under the series title “Deep Dive,” provided by NotebookLM. Since that part can’t be changed, I added the “treasure hunt” component as a new note as a source for each topic.

Sources for training AI on ATD

I used only two sources for this podcast, both blog posts recently published on the Alcohol Studies Archives page, however, I also took advantage of the AI-generated FAQ created in NotebookLM, see full text at the end.

Opportunities with NotebookLM

Like anything else with generative AI, this is up for discussion. Here are a few ideas on how creating podcasts with the above workflow can add value to our jobs and help educators and scientists reach untapped audiences.

  • Organizing content in a more structured, more meaningful way to meet the needs of target audiences by selecting content and taking control over the sources
  • Creating timelines, outlines, and FAQs from selected resources and sharing them in various formats, such as handouts, web content, or social media
  • Translating research: comparing sources, analyzing content, and interpreting data at the appropriate level in more engaging ways to meet diverse information-seeking behaviors of diverse audiences
  • Using podcasts as conversation starters on sensitive topics, as a way of “scholarly bibliotherapy,” i.e., allowing people to talk about their own issues through texts they can relate to and encouraging them to find out more

Challenges

AI-generated texts remind us of the old “Garbage In, Garbage Out” principle—on steroids. Even perfectly accurate training data can result in gross errors, often called “hallucinations.” When generative AI is prompted to fulfill a query, any Large Language Model (LLM) will rely on patterns learned from the corpus to make predictions, not factual knowledge. That’s exactly what generative AI is good at: making predictions; AI doesn’t understand the emphasis; AI can’t separate facts and fabrications—AI hallucinates.

To improve the quality of AI-generated content, NotebookLM uses an AI framework called RAG (Retrieval-Augmented Generation), more precisely, a RAG-locked model, customizable by the end user: AI can access a specific information set as a source during the retrieval phase (i.e., the two documents I uploaded) and can’t change that information during the content generation phase (creating notes, summaries, and FAQs, answering questions, and generating the audio overview (the source of the podcast). Training AI on user-provided content ensures that the LLM will draw upon this narrower set in the first phase. At the same time, the user-initiated note generation, source selection from notes, and adding prompts will create a more accurate representation of the original content. However, when in doubt, tracking down anything that feels suspicious is still the way to go, along with manual fact-checking (Jellinek had several wives, but he never married one in Italy).

Responding to the more significant challenge, i.e., ethical considerations, is a personal choice. The jury is still out: scholars, educators, librarians, and information specialists disagree on intellectual and copyright issues related to AI use when uploading documents into a system to be used as resources. I have trusted friends who are uncomfortable with passing on intellectual and interpretive work to AI (anyone who grades students’ papers, obviously) or outsourcing that work to a predictive text model. I have heard using resources called borderline immoral, and most of us polyglots agree on hating AI translations (or at least often laughing at them), even though machine translation has come a long way since the proverbial example of “out of sight, out of mind” ending up “invisible idiot” through multiple translations.

Alcoholism Treatment Digest FAQ (AI-generated)

What is the Alcoholism Treatment Digest?

The Alcoholism Treatment Digest was a periodical published from 1950 to 1973 by the Center of Alcohol Studies, first at Yale University and then at Rutgers University. It featured short review articles summarizing emerging topics in alcohol literature, particularly focusing on treatment and rehabilitation.

What was unique about the Alcoholism Treatment Digest?

The Digest was innovative for its time. It served as a “content provider” by offering concise summaries of multiple articles on specific topics, written by experts in the field. This model allowed busy physicians and other healthcare professionals to stay informed about the latest advancements in alcohol treatment.

What topics were covered in the Alcoholism Treatment Digest?

The Digest covered a broad range of topics related to alcoholism treatment, including:

  • Medical management of alcohol withdrawal
  • Treatment of alcoholics in mental hospitals
  • The role of probation in managing problem drinkers
  • The effectiveness of compulsory versus voluntary treatment
  • Special populations, such as women and Native Americans
  • Adjuvant therapies like music therapy and pastoral counseling

Who wrote the reviews for the Alcoholism Treatment Digest?

The reviews were initially written by the editorial staff of the Quarterly Journal of Studies on Alcohol, with Mark Keller, the founding editor, being a frequent contributor. However, the names of the authors were only listed starting in 1964.

How was the Alcoholism Treatment Digest distributed?

The Digest was a subscription-based publication. Issues were typed and mimeographed, with instructions for further distribution to relevant audiences. Subscribers were encouraged to share the reviews while acknowledging the copyright of the Journal of Studies on Alcohol.

Why is the Alcoholism Treatment Digest important today?

The Digest provides valuable insights into the history of alcoholism treatment and the evolution of thinking about addiction. It also highlights the importance of disseminating scientific information to healthcare providers and the public.

Where can I access the Alcoholism Treatment Digest?

All issues of the Alcoholism Treatment Digest owned by Rutgers University Libraries have been digitized and are available online in RUcore, the Rutgers University Community Repository.

What other resources are available at the Alcohol Studies Archives?

The Alcohol Studies Archives at Rutgers University houses a vast collection of materials related to alcohol research, including the Journal of Studies on Alcohol and Drugs (formerly the Quarterly Journal of Studies on Alcohol), the Classified Abstract Archive of the Alcohol Literature (CAAAL), and the papers of prominent alcohol researchers.