There are a lot of YouTubers who release daily videos with long rants about their grievances with the video game industry. As a developer, it can be hard to take these rants at face value.
- They often feature a lot of misconceptions about game development
- They are often emotionally charged and sometimes just plain hateful.
- The objectives of the content creator are dubious because they have a financial interest in creating content that provokes strong emotional reactions.
- The videos are often long, rambling, and steeped in culture war lore.
Is there a way to strip all that away, get to the true grievances, and work out what the actual claims are? How far away are they from the bad I see and believe is leading to more churn and less authentic art?
Let’s try using TECHNOLOGY to get to the core, build better communication, and save the world (or just make another tool for more hate)… Most likely, the AI will not do a great job, but let’s try.
Let’s mash a few tools together:
- yt-dlp: Used for pulling the video or audio file from YouTube.
- Whisper: OpenAI’s audio-to-text model.
- AI APIs: Used to tag the transcript with a few simple media literacy and rhetorical frameworks.
Read The Prompt
You are a text analysis assistant. Tag the following transcript with inline markers for:
- [FC] Factual Claim – Objective, verifiable statements.
- [OP] Opinion – Subjective assertions.
- [EL] Emotional Language – Language evoking emotions.
- [RT] Rhetorical Technique – Sarcasm, irony, or rhetorical flourish.
- [SP] Speculation – Uncertainty, guesswork, or predictions.
**Instructions:**
- Segment text into meaningful parts.
- Apply inline tags without overlapping.
- Ensure readability. Keep original wording.
**Example:**
Input: "Last night, there was a big announcement..."
Output: [FC]Last night, there was a big announcement...[/FC] [EL]—everyone said...[/EL]
**Post-Processing:**
- Provide an "Emotional Rating" (0–10).
- Summarize main claims, techniques, and biases.
Now, process the following text:
{transcript}
Then, we pass the tagged text into a simple HTML page that highlights the words for us.
Factual Claims, Opinions ect has been highlighted
I want to be very clear: this tool does not perform fact-checking. A sentence tagged as a “Factual Claim” simply means it is presented as a fact—it does not verify whether it is actually true. It’s up to you to do the research and determine its accuracy.
Does it work?
Well… yes-ish. The results vary wildly between models, but I think the tagging done by the fancy reasoning models is pretty accurate. That said, a traditional fleshy person would still do a much better job if they took their time.
But as a reader trying to understand the claims, I think the tags and highlights do help me quickly parse the text. And the summary is consistently good.
Where does it fall apart
Tagging 5 minutes of transcripts seems to go pretty well, but after that the AI seems to get a more ’lazy’ and less detailed in it’s tagging. Since most of the videos I was interested in lookint at are >10 minutes, some form of segmented call might be necessary to get a proper tagging done.
Also, As of Feb 22, 2025, making an API call to use the fancy ChatGTO-01 model for a 10-minute transcript costs about $1, which can quickly add up when exploring the tool.
Use the HTML highlighter tool here. You can grab YouTube’s own transcript, then drop that along with the prompt into any chatbot to try it out.
Get the full toolset and instructions here. I found that Whisper does a better job transcribing and also doesn’t censor bad language.