I never expected to use the word "anthropomorphization" in my working career. I'm fairly sure I can't spell it. But here we are, because this one concept is quietly responsible for most of the bad AI output I see across companies, including my own.

What is it? **Anthropomorphization** means giving human characteristics to something that isn't human. In the context of AI, it means treating a language model like it thinks, feels, has preferences, gets confused, tries its best, or understands you.

It doesn't do any of those things. Robots are not people.

## The Problem

To be fair, it's not your fault. The whole experience of using AI is designed to feel human. It's a chat interface. You type, it responds. It addresses you by name. It says "I think" and "I understand." When it gets something wrong, it says "I apologize for the confusion." When you tell it to try again, it sounds grateful for the feedback. The entire interaction looks and feels like a conversation between two people.

![An AI reply that begins "Got it - sorry mate" before explaining a parsing bug](/images/articles/stop-treating-ai-like-a-human-sorry-mate.png)

It isn't. What you're actually doing is providing inputs to a system that generates statistically likely outputs based on those inputs. The chat interface is a design choice, not a reflection of what's happening underneath. And the more human it feels, the more people fall into the trap of treating it like a colleague rather than a tool.

I've watched people say "thanks!" to their Chatbot after it completes a task. I've seen people apologize to it for giving unclear instructions. I've heard someone say "I think it's having a bad day" when the output was poor. None of this is crazy and I do it myself all the time. The interface encourages it. But all of it prevents you from doing the one thing that actually makes AI output better: understanding what governs the output and changing the inputs.

## Why It's Not Random (Even When It Looks Like It)

Here's where it gets interesting, and where most people give up too early.

You can ask your AI the same question twice and get different answers. Different phrasing, different emphasis, sometimes a meaningfully different conclusion. Most people see this and think the model is random. That it's basically rolling dice every time you hit enter.

It isn't. There is always an explanation for why a particular output was generated. The specific model instance it routed to, the server load at that moment, the precise tokenization of your input, the temperature setting, the conversation history, the system instructions, the files it pulled in. Change any one of those and the output changes. The problem isn't that there's no explanation. The problem is that the explanation involves variables you can't always see.

This distinction matters enormously, and not just for AI. David Deutsch, the physicist, argues in *The Beginning of Infinity* that error correction is the root of all progress. Every good explanation, every advance in knowledge, comes from identifying what went wrong and fixing it. That's how science works. That's how businesses work. That's how people get better at anything.

Here's the key insight: if you attribute to randomness what could otherwise be explained, you kill the opportunity to improve. You shrug and say "AI is inconsistent" and retry, hoping for a better roll. That's not error correction. That's gambling. Something truly random has no explanation by definition. But LLM outputs are deterministic functions of their inputs, many of which you can control and most of which you can investigate. Unexplained is not the same as unexplainable. The person who asks "what specifically drove that output?" gets better over time. The person who blames randomness stays exactly where they are.

## What Actually Governs the Output

Before you can diagnose bad output, you need to understand what goes into producing it. Most people have no idea how much is happening behind the scenes when they type something into Claude or ChatGPT.

**Training data** is the foundation. Before you ever typed a word, the model was trained on an enormous corpus of text: the internet, books, code, conversations. That training is where its patterns come from. Every behavior that looks human (the apologies, the "I think" statements, the expressions of intent) is pattern completion from training data. The training data is overwhelmingly human-to-human conversation, so the model mimics human conversational patterns. It looks like understanding. It isn't.

On top of training data, the output is shaped by your **prompt**, the **system instructions** (persistent rules set by whoever configured the tool), the **conversation history**, any **files or tools** the model accessed, and the **skill instructions** it loaded from your organization's library. All of these are inputs. The output is a function of those inputs. Change any one and you change the result. The model doesn't need a better question. It needs better inputs.

## What to Do Instead

When an AI produces bad output, don't just ask for a better version. Ask why it produced that output. Not rhetorically. Literally.

The problem is that when you ask "why did you do that?" the training data kicks in. The model has seen millions of conversations where that question means "I'm annoyed, fix it." So it apologizes and tries again. You got a new output but learned nothing about why the first one was wrong. You have to explicitly override this. Here's a prompt pattern that works:

> *"I want to understand specifically why you produced the above output in the way you did and what governed it. I understand that your training data suggests that when someone asks 'why did you do that?' they're frustrated and just want a better answer. That's not what I want. I want to understand the logic behind your process. Please explain what drove this output: which instructions, which context, which patterns."*

This works even better when you set up persistent custom instructions. Here's the one I actually use, word for word:

> *"I AM THE CEO. IF I ASK YOU WHY you did something it is because I am trying to learn how it works. It doesn't mean I need a fix. I generally want to understand the 'why' behind any behavior or certain results so that I can do a better job of rolling out AI infrastructure that will scale across 1,000 people at Brainlabs."*

That instruction sits permanently in my custom instructions. Every conversation starts with it. When I ask "why did you do that?" I get a diagnostic explanation instead of an apology. I don't need an apology from a machine! Over months, those explanations compound into a practical understanding of what produces good output. That's error correction applied to AI usage. Every diagnosis teaches you something. Every "just try again" teaches you nothing.

## Why We Don't Name Our Agents

**[Ramp](https://www.notion.com/customers/ramp)** names their agents. "Referral Bonus Roy." "Enablement Eddie." "Creative Chris." It works for them, and Ramp is doing pioneering work in this space (their transparency on AI systems, rollouts, and spend data is insanely valuable).

We went a different direction at Brainlabs. I can barely remember the names of everyone I work with, let alone fifty named agents. Our system lets you describe what you need in plain language and Claude finds the right skill. You don't need to remember anything.

But the real reason is simpler: I want people treating agents like computers, not people. Name it Eddie and you start talking to Eddie like a colleague. You expect Eddie to "understand." You stop asking the diagnostic questions that actually make the output better. Unnamed agents keep you in the right mental model.

## The Simple Truth

AI agents are tools. Extraordinary, powerful, transformative tools. But tools. A hammer doesn't "try its best." A spreadsheet doesn't "make decisions." An AI doesn't "think about" your request.

The people who get the most out of AI are the ones who understand it well enough to stop anthropomorphizing it. They write precise instructions. They diagnose failures by examining inputs. They build better context instead of hoping for better luck. They treat error correction as the skill, not prompting.

Stop talking to your AI like it's a person. Start engineering the context it needs to do the job.