Marc Cracco
Chief Technology Officer @ Wondersauce
Chief Information Officer @ Project Worldwide
After a conversation with some people at work on the news around Grok I decided to put pen to paper on the topic of LLMs and objectivity. There’s this persistent idea that LLMs are neutral. That they reflect the internet as-is, translating raw data into informed answers through predictive math. But anyone who’s worked with these models knows that’s not how it works.
Every LLM starts with a dataset, and that’s where the shaping begins. The data isn’t pulled in randomly from internet scrapes or other sources. Instead it’s selected, cleaned, filtered. Some sources are prioritized because they rank high in search, others are added because they look authoritative and a lot is excluded entirely. That could be content behind paywalls, niche and hyper focused communities, smaller languages, or anything that introduces legal risk or doesn’t scale cleanly. Even when a company shares its dataset, it rarely shares how different types of content are weighted or what is filtered out. In most cases frequency becomes a proxy for importance and worth a proxy for facts and truth. Visibility becomes a stand-in for credibility. These are human decisions baked into the system from the start.
Then after training comes fine-tuning. This is where the model is taught how to respond. Not just what to say, but how to say it. It’s trained on preferred formats, rewarded for certain tones, penalized for others. This step reinforces values, filters perspective, and sets the boundaries of what’s considered acceptable. Before this its knowledge was biased based on data selecting and now after this it has lost any aspect of neutrality.
A model at this point is deeply opinionated, even if it’s masked in generic language.
Lastly, system prompts and guardrails lock in behavior. They define what the model assumes, what it avoids, and what kind of voice it uses. You can have two models trained on the same dataset and get completely different results, depending on what was adjusted during and after fine-tuning. Based on this even access to the raw dataset wouldn’t give anyone a clear picture of a model. The reality is the training data is just the starting point. What happens after is where most of the shaping really takes place.
Curation Is Not Objectivity
Now I mentioned that the original training data isn’t random. It’s curated. Fine-tuning is just another word for editing. And prompt engineering, which most users never even see, controls how the model thinks about tone, authority, and correctness. So when a model responds to a question, it’s not just summarizing the world. It’s reflecting the choices of the team that built it, the constraints they imposed, and the outcomes they wanted to push.
When Transparency Disappears, Trust Should Too
Unfortunately the illusion of objectivity becomes dangerous when companies present these outputs as fact while hiding the mechanics. To add to this lack of transparency, some of the major players out there have gutted their ethics team, and their tuning is driven by brand safety over truth. Add datasets and rules that are locked behind closed doors, and yet the outputted model is shared as a neutral, factual and unbiased representation. The fact remains and should be clear:
They built something with a perspective. They just didn’t say what it was.
Where This Leaves Us
LLMs are already shaping public opinion, buying decisions, and the way people work. Treating them like passive tools while they are actively curated and controlled is a problem. Until there’s transparency around how these models are trained, fine-tuned, and governed, we should stop pretending they’re anything other than editorial products with invisible bylines.