AI slop comes from
bad communication.
By 2030 that excuse runs out.

Every week I see it. Someone posts a screenshot of an AI-generated response that misses the point, hallucinates a fact, or produces something generic and limp. "AI is overrated." "It can't think." "It's just autocomplete." The post gets engagement. People nod along.

The same week, GPT-6 Astra scored 100% on ExploitBench — a cybersecurity benchmark designed to represent real-world attack surfaces. The same week, Anthropic's Fable 5.1 and Mythos 5.1 launched as the most capable AI systems ever made publicly available. The same week, I had a conversation with Claude Sonnet — not Opus, not the frontier model, the mid-tier — and it helped me build and iterate on a production web system that real users are using right now.

The doubters are not wrong that AI produces bad output. They are wrong about why.

AI slop is not an AI problem. It is a communication problem. The relationship between a person and an AI is exactly like any other working relationship — it produces results proportional to the clarity of the brief, the quality of the context provided, and the effort put into directing and reviewing the output. Give a vague brief to the best developer in the world and you get vague software. Give a clear, detailed, well-structured brief to a mid-tier AI model and you get something that works.

What I actually built — and what I built it with

TripSpace Global — tripspaceglobal.com — is a commission-free travel marketplace. Accommodation listings, a host onboarding system, booking flows, a partner programme, schema markup for GEO indexing, affiliate integration. It is a real platform, in production, handling real enquiries.

I built it using Claude Sonnet at medium tokens. Not Opus. Not an unlimited context window. Not the frontier model with special access. Sonnet, at the tier most people reading this have access to, with considered token management. I am writing this article on Sonnet at medium right now.

Forge Vertical itself — the agency, the articles hub, the onboarding system, the Cloud Functions, the Firestore integration, the GEO schema — all of it directed through AI, built in production, indexed by Google, ranking for real search terms within hours of publishing. The entire article infrastructure you are reading right now came from the same model tier that people are calling overrated.

The difference between what I build and what someone else describes as "AI slop" is not the model. It is the brief. It is knowing what you want before you ask for it. It is providing context, constraints, and examples. It is reviewing the output, identifying what is wrong, and communicating the correction precisely. It is treating the AI as a capable collaborator that needs good direction — not a magic button that produces results without input.

The communication parallel — why this is exactly like human relationships

Think about the worst working relationship you have had. The one where deliverables came back wrong, where expectations were not met, where both parties ended up frustrated. Almost certainly, the root cause was communication — unclear briefs, assumptions that were never stated, feedback that was vague rather than specific, context that was never shared.

Now think about the best working relationship. Clear expectations. Specific feedback. Context shared upfront. A shared understanding of what success looks like before the work begins. The relationship that produces good work is the one where communication is precise, generous, and continuous.

The relationship between a person and an AI is structurally identical. The AI cannot read your mind. It cannot fill in context you did not provide. It cannot know that the client has a specific formatting preference you forgot to mention, or that the previous version of this feature broke in a specific way, or that "simple" means something particular in your context. You have to tell it. Precisely. With examples where useful. With constraints stated explicitly.

When people post screenshots of AI producing garbage output, the screenshot almost always includes a two-sentence prompt that could mean three different things. The AI made a reasonable interpretation. It was not the interpretation the person intended. That is not the AI failing — that is a brief that would fail any collaborator, human or otherwise.

What 2026 looks like — and what 2030 will look like from here

The trajectory from GPT-3.5 in 2022 to GPT-6 Astra in 2026 is four years. In those four years the capability went from "impressive party trick" to "100% on a real-world cybersecurity exploitation benchmark." Whatever the next four years produce will not be a smaller leap. The underlying compute, the training methodologies, the architecture improvements — all of it is accelerating, not plateauing.

The people who are most confident that AI has hit a ceiling are the same people who were most confident that GPT-3.5 represented the limit of what the technology could do. The pattern is consistent and the error is the same: judging the trajectory by the current position rather than the direction and rate of change.

Dario was right — and it is worth sitting with what that means

He was not talking about discomfort for the people who adapt. He was talking about structural discomfort — the kind that happens when an economic system built around human labour in specific forms encounters a technology that can perform those forms better, faster, and cheaper. That discomfort is not hypothetical. It is already present in the 165,000 AI-related layoffs in the first seven months of 2026, in the 491 tech jobs that ceased to exist every day of that period.

The comfort in this is that the discomfort is navigable. Not easy. Navigable. The people who are building skills in AI communication — who are learning to brief well, to review critically, to direct and iterate rather than passively receive — are building something that compounds. The skill of communicating precisely with AI is the skill of thinking precisely about what you want. That skill transfers to every domain. It makes you a better manager, a better writer, a better developer, a better founder.

The people who are waiting for AI to become easy enough to use without effort are waiting for a thing that will not arrive. The AI will get more capable. The communication requirement will not go away — it will change in character. Better models can handle more ambiguity and more implicit context. But the person who provides clear context and precise direction will always get better output than the person who provides neither. That gap does not close.

The practical conclusion — what to do with this

I am not writing this from a theory position. I write it from a production position. Forge Vertical is live. TripSpace Global is live. The articles you are reading were researched, structured, and written with AI collaboration. The Cloud Functions that handle lead submissions from the onboard form, the Firestore database that stores them, the Zoho CRM integration that syncs them — all directed through AI, all in production, all working.

None of this required Opus. None of it required frontier model access. It required knowing what I was building before I started building it. It required clear prompts. It required reviewing output and communicating corrections precisely. It required treating AI as a capable collaborator rather than a search engine that writes paragraphs.

If you are in the tech space doubting AI's capability — spend a week building something with it before you form that opinion. Not asking it questions. Building something. A tool, a system, a product. Give it real context. Review its output seriously. Communicate your corrections clearly. See what it produces when it is given what it needs to produce well.

If the output is slop, ask yourself honestly: was the brief clear? Was the context complete? Was the feedback specific? In my experience, the answer to at least one of those is no. Fix that first. The model is probably not the problem.