How I Used AI to Support This Keynote

• I first explored the Quadruple Aim in a WordPress discussion piece written while I was learning to use generative AI and publish documents online.
• The original article was fully referenced. I read each paper, added it to Mendeley, checked its content, and ensured that each source supported how I used it.
• When invited to present, I initially planned to discuss the palliative dimensions of falls and supported decision-making. An organising committee member encouraged me to return to the Quadruple Aim as a stronger keynote topic.
• In March 2026, I asked ChatGPT and Google Gemini to:
  • remap the Quadruple Aim literature;
  • compare it with my original article;
  • propose a 40-minute presentation structure; and
  • estimate an appropriate script length.
• I completed approximately ten iterations with each tool, comparing their outputs rather than accepting either as authoritative. Their estimated script length was 4,300–5,100 words.
• I initially developed about 400 words per proposed slide. At this stage, supported decision-making and my PhD remained prominent, while the eventual theme of competing tensions within each aim had not fully emerged.
• I recorded five rehearsals. Plaud.ai produced transcripts, which I returned to ChatGPT and Gemini for critique.
  • Gemini was generally stronger at developing ideas and connections.
  • ChatGPT was effective at identifying repetition and improving clarity, but often shortened the prose until it no longer sounded like my natural speaking voice. ChatGPT had a habit of reducing the speech to odd, small snippets or soundbites that made no sense.
  • The recordings showed that I speak quickly, rush presentations and accelerate further when nervous.
• I then set the talk aside while completing my first PhD progress review, presenting in Germany and Singapore, progressing delayed-discharge research and submitting two papers.
• In June, I rebuilt the keynote using Perplexity, ChatGPT and Gemini. I supplied:
  • the earlier PowerPoint;
  • previous scripts;
  • the original WordPress material; and
  • My Mendeley bibliography of approximately 1,700 reviewed articles. The BIB file from Mendeley contains the references and their abstracts, not the full articles. Full article review remained part of the human review process.
• I downloaded and manually reviewed the 37 articles forming the keynote’s central evidence base. I revised the presentation so every substantive point had supporting evidence and each section flowed logically into the next.
• During this reconstruction, the direct PhD focus was reduced, and the stronger organising idea emerged: each part of the Quadruple Aim contains tensions between legitimate but competing priorities.
• Two experienced colleagues recommended fewer text-heavy slides, more images and different ways for audience members to engage. This led to companion WordPress pages where people could explore the evidence independently.

My Chain of Verification

• Human-curated Mendeley library. This is the full PhD Mendeley library that I’ve amassed via Scoping Review work and Progress Review preparations.
• Comparison across three generative AI systems. Each output was compared to maximise the usefulness of the language structure that best matched my own.
• Retrieval and manual review of all 37 central papers.
• Repeated recording, transcription and rehearsal analysis.
• Feedback from experienced colleagues.
• Manual review and editing of every AI-generated output.
• Final responsibility for every claim, interpretation, slide and word remained mine.

This keynote was AI-supported, evidence-verified and human-authored.