So-called "AI" disclosure

More "content creators" should disclose how they use so-called "AI", so here we go. Here's how I use it.

A picture of a woman in her studio, but a robot emoji is poorly superimposed over her face.
Me, probably

I believe "content creators" here in the roaring 2K20s should be transparent when it comes to how they use so-called "AI" tools in their work. Increasingly, many of my contemporaries are letting computers do their homework for them, and while I don't think all machine learning is evil, I certainly don't want it near most of my creative processes.

This page attempts to catalog how I use machine learning tools. If it's not listed here, you should probably assume that I'm doing it much in the same way I did it in the 2010s, before large language models like Copilot and ChatGPT became commonplace.

I will keep this page up-to-date as things change and questions come up.

What I don't do

  • Use LLMs in the scriptwriting process (my words are mine and I wrote them myself).
  • Knowingly¹ use images or video footage generated by so-called "generative AI."

Ways I use large language models (LLMs)

  • I use a locally hosted instance of OpenAI's Whisper speech-to-text project to generate a first draft of captions for many videos. I then fine tune this by hand, often rewriting a significant portion of it. I feel justified in this use, as it is baked into my video editing software (Kdenlive) and helps me improve the accessibility of my videos. The captions automatically generated by YouTube are terrible, and using a locally hosted LLM in conjunction with manual review offers a significant improvement.
  • I occasionally use various public and locally hosted LLMs in order to see "what they're saying" about me and my videos. I feel justified in this use as it's similar to looking yourself up in a search engine- considering how many people use language models as though they're search engines, this feels like a good practice.
  • I use a locally hosted instance of Meta's SAM2 object segmentation project as part of my video editing software (Kdenlive). This is used for background removal in some shots. I do not consider this to be "generative AI" in the sense that the term is frequently used, as it's subtractive and not additive- it's removing footage, not creating it.
  • I use a locally hosted instance of Open Source Vision Foundation's OpenCV computer vision project as part of my video editing software (Kdenlive). This is used for motion tracking- creating keyframes matching an object in motion. This is very useful for image stabilization (reducing camera shake) and setting one layer to move with another (for example, when superimposed text sits next to an object in motion). I do not consider this to be "generative AI" in the sense that the term is frequently used, as it's performing actions against real footage that I created the "traditional" way (in camera or manually in screen).
  • I use Immich's machine learning tools for cataloging the footage I capture, in order to make reuse easier. For example, I can search my Immich instance for "Commodore 64" and get b-roll I'd captured in a previous video, thanks to the model which scans frames of footage and generates keywords. This replaced an earlier system involving lots of READMEs in folders with loads of duplicate footage.

Footnotes

¹ It's impossible to be entirely certain these images do not appear in my work. For example, if a software project uses a machine-generated logo, and they do not disclose this fact, it may end up in my work inadvertently.