It's Not AI That's Confusing, It's the Models (and Their Names)

The hype around AI is everywhere. Tools like ChatGPT, Perplexity, and Gemini dominate headlines, sales meetings, and strategy decks. But for many people trying to put these tools to use, the biggest challenge isn’t using AI—it’s understanding which AI to use, and why.

It’s not resistance holding people back.It’s confusion.

Most AI tools are front ends. What powers them are models—large language models, multimodal models, reasoning agents, and more. These models determine how the AI behaves. But ask any average business user what model they’re working with, and you’ll likely get a blank stare. Even those of us with a reasonably technical background have to double-check: What’s the difference between GPT-4 and GPT-4o? Is Gemini 2.5 Flash better than Gemini 2.5 Pro, or the other way around? What’s “Perplexity Sora” actually good at?

We don’t just have a tooling problem.We have a model literacy problem.

And the model names? They're making things worse.

The naming conventions for AI models feel like they were designed to confuse. Instead of clarity, we get alphanumeric codes, obscure codenames, and random versioning.

Take OpenAI’s latest releases, GPT-4o, and o4.The “o” stands for “omni,” apparently.But which is best? And for what?Is it different in its training?Is it a subset? A superset? A side project?

And let’s not forget OpenAI’s other ambiguous labels: “o4-mini-high” and “o4-mini”High = better right? Is it slower? Is it smarter?

These names give us nothing about what the model does well. And for people outside of AI research circles, they actively create barriers.

Now layer on the tools themselves.ChatGPT lets you choose a model—but doesn’t always explain what that choice affects.Perplexity doesn’t tell you which model they’re using at all.Anthropic’s Claude models hide behind poetic codenames like “Opus,” “Sonnet,” and “Haiku,” without much context about strengths or ideal use cases.

So a sales enablement lead tries to draft a playbook.A customer success manager tries to summarise a call.A learning designer tries to generate course content.

They test a tool, get unexpected results, and walk away unimpressed.The AI isn’t broken.The behaviour was just mismatched to the task.And no one told them what to expect.

The real conversation we should be having is about AI behaviours, not AI tools.

If we start classifying models by what they do well, rather than what they’re called, we create a much more usable mental model for non-technical users.

Imagine a world where model names weren’t cryptic—they were descriptive.

These are not just friendlier names—they’re functional. They give users a shortcut to understand what a model might be good for, and where it might fall short.

If AI companies want to accelerate adoption, and retain customers, or even just stand out in a crows they need to stop thinking like research labs and start thinking like product teams.

The current approach feels like watching the early internet unfold again—technical brilliance, hidden behind awful naming, inconsistent UX, and limited user education.

We need naming conventions that reflect behaviours, not architectures.We need onboarding that explains use cases, not token limits.And we need tooling interfaces that show you what model is in use, and what to expect from it.

Until we get that, we’ll keep seeing people abandon powerful tools out of frustration, not failure.

Most AI resistance isn’t rooted in fear. It’s rooted in confusion and friction.

By making models more understandable, by naming them in ways that reflect their strengths, and by educating users on behaviour instead of backend architecture, we can remove the most frustrating barriers to AI adoption.

The tools are ready.The models are powerful.Now it’s time to make the experience usable.