Nobody Knows What They're Banning

The debate about AI keeps snagging on the same problem. People are arguing about a word they haven't defined.

The word is AI. Or sometimes LLM. They get used interchangeably, as if they mean the same thing. They don't. Banning "AI" without specifying which kind is like banning "vehicles" and hoping nobody notices you've also banned wheelchairs.

This matters because the accessibility argument, the one I've been making, lives in a completely different category from the one most critics are actually worried about. When Iris van Rooij, whose work I take seriously, says prohibit the use of LLMs for grant writing, she's being more precise than most. LLM is a real term with a real meaning. But even that is getting stretched to cover things it was never designed to describe.

I'm not making this argument from the outside. I've spent 25 years in the technology industry, a significant portion of it working closely with AI systems at large scale. I understand what these tools are, how they're built, & what they actually do in enterprise environments. I'm also a husband, a father, a writer, & someone diagnosed dyslexic in 1983. Those two things, the practitioner & the person asking for accommodation, are not in conflict. But the debate keeps acting as if they are.

Here are the five main categories of AI technology, in plain language. They are not the same thing. Policy that treats them as one thing will get the wrong answers.

1. Large Language Models (LLMs)

This is what most people picture when they say "AI." GPT-4, Claude, Gemini. Systems trained on enormous text datasets to generate, summarise, translate, or respond to natural language. When critics talk about AI replacing writers, ghost-writing essays, or flooding the internet with synthetic content, this is the category they mean. The critique is legitimate, at least in part. These are powerful tools with real potential for misuse, particularly when the output is presented as wholly human-authored without disclosure.

2. Assistive & Scaffolding Tools

This category includes Grammarly, text-to-speech, speech-to-text, autocomplete, screen readers, and AI-assisted layout tools. Many of these tools are not LLMs. Some use older ML models. Some use rule-based systems that aren't "AI" in any meaningful contemporary sense. These are the tools that make written communication possible for people with dyslexia, ADHD, motor impairments, processing disorders, and a range of other conditions. They are the category that disability advocates have been using for decades. Policy written about category one, if drafted carelessly, will land on category two. That is not a theoretical risk. It is already happening.

3. Image, Video & Audio Generation

Midjourney, DALL-E, Sora, Suno. These are diffusion models, not language models. The mechanics are fundamentally different. The copyright, consent, & labour concerns in this category are real & distinct: the training data questions, the artist consent issues, the synthetic media risks. These deserve serious policy attention. But they are not the same problem as an LLM writing a grant application, & conflating them produces bad law.

4. Predictive & Recommendation ML

The Netflix algorithm. Spotify's Discover Weekly. Credit scoring models. Insurance risk assessment. The fraud detection system on your bank account. This is the oldest category of AI in widespread use, & almost nobody calls it AI in casual conversation. It has been shaping what you see, what you're offered, & what you're denied for twenty years. The harms here, particularly in financial & legal contexts, are well-documented & ongoing. Somehow this category rarely enters the AI ban conversation, possibly because banning it would require confronting systems that benefit powerful institutions.

5. Agentic & Autonomous Systems

Self-driving vehicles, robotics, AI systems that can take real-world actions with minimal human oversight. The safety, liability, & autonomy questions here are serious & genuinely novel. This is the category most likely to produce harms that no existing legal framework is equipped to handle. It is also, again, not an LLM.

These five categories have different architectures, different training methods, different risk profiles, & different relationships to human creativity, labour, & agency. Treating them as one thing because they all get called "AI" is not a policy position. It's a category error.

The accessibility argument sits almost entirely in category two. When I use AI assistance to write, I am using a scaffolding tool to bridge the gap between the ideas in my head & the words on the page. The cognitive architecture is there. The perplexity score is there. The documented reasoning is there. What the tool does is reduce the friction that dyslexia & processing speed differences create between thought & output. That is not ghostwriting. It is a prosthetic.

6. The Physical Cost Nobody Is Counting

There is a separate argument that rarely enters this conversation, & it should.

The infrastructure that runs large-scale AI, data centres, cooling systems, power grids, is a material problem with a measurable footprint. A 2023 study by Li et al. estimated that training GPT-3 consumed roughly 700,000 litres of fresh water for cooling. Inference costs, the resources used every time someone runs a query, compound this at scale. The environmental cost of AI is real, significant, & almost entirely absent from mainstream AI ethics debates, which tend to focus on content & labour rather than physical infrastructure.

This cost is not evenly distributed. Data centres are frequently located in regions with existing water stress. The communities bearing the environmental cost are rarely the communities driving demand.

7. The Argument Nobody Is Having

If the environmental cost of AI is a genuine concern, & it should be, then the conversation needs to include not just whether AI should be used, but how it is used.

A single careless query, a vague prompt that forces a model to run multiple inference passes before producing something useful, has a higher resource cost than a well-structured, targeted prompt that gets the right output on the first pass. Responsible prompting is not just about accuracy. It is an environmental practice.

Nobody is having this argument. The discourse is locked between "ban it" & "use it freely." The more useful question is: if we accept that some uses of AI are legitimate, what does responsible use look like? What does it cost, & who pays that cost?

The answer depends entirely on which category of AI you're talking about.

Which is why defining the word matters.

Carl Freeman is a writer, husband, inventor, father, LLMego enthusiast (dad AI joke) & the author of The Faculty of Matter, a dreampunk novel in progress. He has spent 25 years in the technology industry working closely with AI at enterprise scale. He was diagnosed dyslexic in 1983 & uses AI as an accessibility scaffold. Full methodology & provenance documentation available on request.