Is your placenta diabetic?

As a creator of any sort, when creating clips to promote your content is simply “the done thing”, you spend an awful lot of time reviewing the tools to make this chore less of a chore. The landscape for this has evolved over the years, and every now and then a new app will pop up on my feed or I’ll be invited to trial a new service. Always with mixed, but instructive, results.

Everyone seems to have a novel “interesting detector” and these emerged en masse once AI started becoming accessible. What is most interesting about this is how “interesting” is defined.

In 2022, I tested a few services against my first STEAM Powered episode with Dr Rebecca Lim on stem cell therapies and regenerative medicine. We spoke about gestational diabetes as well as anti-depressant use during pregnancy in one small section of the conversation, and stem cell use in the context of organ transplants in another part.

Spelling issues aside, this one service’s attempt at topic extraction placed a lot of focus on placentas as though this encompassed the whole conversation, used it in entirely incorrect contexts, and phoned it in on the last one, “Growth in the Area”, by lifting it straight out of my chapter titles.

It was trying to extract what the episode was syntactically about, and still failed to provide reasonable or accurate suggestions that could be used as topics headings, highlights, or notable moments. That said, in isolation, they would certainly be attention-grabbing if not a complete misrepresentation of what the listener is actually going to get.

This was the most egregious example, but in all cases—including some recent tools—where it sort of worked, what I got was a statistically reasonable idea of what was relevant. But is that interesting? And to whom?

AI book summary services have also grown in popularity over the last few years, but many of them give me a niggling feeling that something isn’t quite right. They are for the most part accurate, and certainly tell me the key points of what I need to know about a book, but they are also rather sterile, even from a non-fiction perspective.

When reading human-authored book summaries, the sterility problem is replaced with a relevance problem. You can’t read just anyone’s review of a book. There are times where I’ve encountered a summary from an unknown reviewer that’s given me just enough information to want to read the source, only to find that they’ve left whole swathes of value on the table, entirely unmentioned or misinterpreted in their own analysis. Something very relevant to me was completely irrelevant or read in an entirely different way because their context is not my context.

On the other hand, I’ve often found the best reviews come from someone whose writing I trust, or whose views or background reflects mine. These may not be the same person. Someone whose writing I trust but with completely opposing perspectives will give me just as much value as someone who shares the same perspectives I do.

This is what is missing in programmatic tools for clip selection. These tools optimise for syntactic importance—what is mentioned the most, what sounds like a conclusion, or what the key topic is. But a podcaster who knows their audience is optimising for resonance. They want to share what is going to land for this listener. They’re going to look for what their audience cares about and what they will connect with enough to want to listen to the full conversation.

This isn’t going to be something we can easily engineer our way out of. If we could, everyone would be going viral with a 100% hit rate. The ‘trick’ is traversing the gap between what’s syntactically extractable and what will actually resonate. That is a craft that requires the creator’s hand—the gap isn’t going to close just because the tools get better.

That gap sits within a larger picture, one I’ve been working through as a creator and a communicator navigating the frameworks that support us and constrain us in equal measure. We’re bound by the listens that get counted in fragmented ecosystems. The platforms that rise, fall, and consolidate with their users having to ride the wave or wipe out entirely outside of their control. Tools and features that solve for a different friction point or that create entirely new ones. Extraction instead of context, syntactic importance instead of audience resonance.

Automated clip generation isn’t bad — surfacing key topics is still important — but it is not enough on its own because what’s worth sharing depends on knowing who you’re sharing with.

That’s the shape of the problem as I’ve lived it. Next, I’ll be getting into how I navigate the space between.

Stay curious,

— Michele

Cover photo by Steve A Johnson on Pexels