Thoughts About AI Slop

I’ve seen several videos on YouTube complaining about how AI slop is ruining videos in their field of interest. This seems to be a particular problem with history videos. Fake info didn’t start with the internet, but it has become more widely visible since it is easier to produce. I haven’t been a paragon of virtue in this regard myself, but I am trying to improve.

What is AI Slop?

I see AI Slop as not so much a problem with AI (Artificial Intelligence) technology, as the misuse of those tools by people who either don’t care about the quality of their work, or who lack the knowledge to recognize problems with their work.

There is a distinction between slop and deception. Slop is created when people don’t care. Deception is a deliberate effort to mislead.

The Cause of AI Slop

I have used AI tools in my own work and found them useful. Most of my experience has been with text to voice systems. This has given me some understanding of the short comings and problems with current AI technology.

I have used text to voice systems extensively. There are some patterns I have noticed. One is getting odd gaps in the dialogue. I usually edit these gaps out later or play around with punctuation to fix the problem.

These systems will often mispronounce uncommon words. In my own work I fix these problems by altering the spellings of these words to force the pronunciation to fit. Another common give away of AI is with words like “lead,” which pronounced differently depending on the meaning in context. Some newer systems are better at it.

When I catch these kinds of issues, it tells me that the “creator” hasn’t bothered to listen to the recording or fix it if they did.

In my own work I’ve sometimes put in gaps, mispronunciations, and “ums” to give the voice more character.

Some creators use AI to create their written material and scripts. AI writing is often repetitious, contain contradictory information, or contain incorrect information. Again, this is not purely an AI problem. In my experience I have never had AI generated writing that I didn’t need to rewrite.

My Hands Are Not Completely Clean

My hands are not completely clean, since I have done some videos and posts, in jest, which some people might see as slop. I now try to avoid creating and accepting fake information.

My film “Who Shot the President,” https://www.youtube.com/watch?v=3lSTEhH0d4E is a fake documentary. Some viewers still took it as a serious proposal. Some even sent threats because of the strong feelings about their own pet theories.

I worked on another project I called “The First Person in Space.” In it a character claims that in 1943, Werner Von Braun used a V-2 to launch a female test pilot into a mission like Shepard’s 1961 Mercury test flight. After my experience with “Who shot the President,” I worried that the idea was too plausible, and some people might think it was real. So I never posted it.

I have also made the mistake of not reviewing my videos thoroughly before I post them. Most recently, I posted a sped-up version of “The Barrier” with out rewatching it. It was only after someone pointed it out that I realised the technique I used adjusted the video and audio by different amounts. I was able to fix the problem, but it still shows how important it is to review your own work closely.

Spotting AI Slop

I have developed my own ways of spotting AI slop, and there are many sources online that have advice on how to do that. However, all these techniques take some effort. How many people will spend two hours checking if they can trust a five-minute video they just watched?

Why is AI Writing Repetitious?

After I wrote a draft of this post, I chatted with a friend. He also noticed that AI generated writing tends to repeat itself. That got me to wonder why.

I recalled seeing an interview with Marshall McLuhan where he said that he would only read every other page of an academic or technical paper. He said this was possible because if it were well written, then there would be enough redundancy in the writing that he could follow what the paper was saying. He contrasted that with fiction and other writing meant for entertainment, which wouldn’t have that redundancy.

In my own experience with technical writing, most of the writing is not intended to read. That is the material is meant as background and the writer would expect that the reader would only skim the report to confirm that the work had been done or consulted when reproducing the work. I found switching from that mindset to writing for entertainment a challenge.

I suspect that most of the writing used to train AI falls into the technical/academic style. That could be one reason why AI writing would be more redundant.

Again, it does mean that creators who don’t bother to review what AI gives them is the source of AI slop.

This post is a mirror from my main blog http://www.dynamiclethargyfilms.ca/blog