Normalization in AI

Why does your AI not sound human? Let's go back to how AI works for a moment. AI as an LLM takes in vast amounts of written words and decides how those words are related to each other in meaning and use and then makes predictions based off 100s of 1000s of 1000000s of previous examples. So if I am writing a sentence, AI can guess what will come next and gives a probability of the next set of words in the sentence. LLMs do this on an even larger scale by predicting what you want to hear.

To go through this process, LLMs need to decide the most likely scenario. As a side effect, we get something called normalization or a flattening effect, where the language and creative patterning is centralized along a bell curve. This means that language becomes less spread out or creative. AI will always choose something in the middle of the curve even when you turn the temperature (an AI term to mean the randomness of a selected likely word choice) higher.

In human language we don't naturally do this. You can tell someone's likely location based off of word choice, you can make estimations about their education level, culture, age, and gender based off the words they use and how they use them. The lexicon varies depending on the person.

AI smushes all of us together into one universal lexicon and then only uses a selection of the most common things in that lexicon which in turn cycles through and flattens itself more.

Language is based off variation, spread, and evolving phrases. AI needs retraining every so often just to keep up with the new ways that language is ever adapting already but if we start re-using only AI outputs in our language our lexicon is at risk of shrinking and losing dimensions.

How can we fix this issue?

AI researchers are tackling this issue every day but what can a daily user do to help mitigate the “AI slop”?

1. Use anti-patterns. Force models outside of those normative standards by banning the middle of the curve. Tell AI it can't use those go-to phrases so it is forced to choose something outside the common choices.
2. Create custom /skills. Feed AI samples of what you want to sound like, samples of your natural, human writing. Favorite authors, speakers, bloggers, etc… and create a tool for your LLM to refer to when writing like you.
3. HITL, or human in the loop. AI should really only build the architecture of written content and not fill in all the details. For one, AI cannot add personal anecdotes which helps users connect on an emotional level.

These tactics work best when explicitly banning words and phrases that are overused or common and helps to add in lexical style back into your work. For example, I am from the Southwest United States but I have strong familial traits that span up the Pacific Northwest. All of these factors give my style subtle but noticeable character. I use these distinctive vocal elements in my speech and writing and have my AI skills emulate these as well to add more character back into my AI dialect.

A few years ago, before generative AI took off, Amazon's Alexa had a feature where you could replace the Alexa voice with a famous actor and have that personality become your personal assistant. If you haven't tried this with your AI before, I highly encourage you to try it and see how the vocabulary changes. In class, one of my professors had an LLM that spoke like Red Forman from That '70s Show and it changed the way the information that AI was giving was perceived.

The key is to create personality and avoid the commonality of the center of the curve in your AI writing. Yes, it takes time and energy and thought, but the value of preserving written culture and expressing yourself in a way that connects is a valuable asset.

If your team is putting AI output in front of customers and it all sounds the same, OPZET can help you build the constraints that fix it.