AI Isn't Replacing You. It's Waiting for Instructions.
The fight isn't AI vs. you. It's whether you're willing to teach it how to do your work. Why both AI doomers and AI maximalists miss the point.
Last week a friend in production told me he wanted to keep creative work in human hands, not AI. I told him that's the wrong argument. The fight isn't AI versus you. It's whether you're willing to teach it how to do your work the way you do.
400 million people, one chat box
Two facts about May 2026.
ChatGPT just crossed 400 million weekly active users. The fastest-growing developer tool of the decade — Cursor — has around one million. That's a 400-to-1 ratio. Almost everyone using AI is using it the same way: typing into a chat box and reading what comes back.
Meanwhile, the word "agentic" appears on more SaaS homepages this quarter than the number of products that have autonomous agents. A line going around AI Twitter sums it up: "Startup: 'we built agentic AI.' Reality: a cron job that hits GPT-4 and calls Zapier."
Two crises at once. Most users have plateaued at the chat box and can't see what's past it. And the language we'd need to describe what's past it has been abused into meaninglessness.
That's the gap I want to talk about.
What the critics get right
The "keep creativity human" argument deserves a steelman, because the people making it aren't fools.
In May 2024, Scarlett Johansson accused OpenAI of using a voice for GPT-4o that sounded indistinguishable from hers — after she'd explicitly declined to license it. "I was shocked, angered and in disbelief that Mr. Altman would pursue a voice that sounded so eerily similar to mine," she said. OpenAI pulled the voice within days.
Stephen King wrote in The Atlantic the same year that "there's something missing in AI-generated fiction, some mysterious human quality that can't be reduced to pattern-matching." Tom Hanks publicly disowned a deepfake of himself in a dental ad. Drake and The Weeknd had a fake song impersonating both of them rack up millions of streams before labels yanked it.
These aren't paranoid. They describe a real thing. When you treat AI as a replacement — a substitute voice, a substitute author, a substitute likeness — it strips the consent, the lived experience, and the economic stake out of the work. The critics are correct about that.
But they're describing one failure mode and calling it the whole technology.
What they don't see
Replacement isn't the only way to use AI. It's the loudest one. The high-performance pattern looks different.
You research a topic. You form a take. You hand it to the model with specific instructions about voice, structure, constraints. It drafts. You push back, redirect, edit. The work that comes out has your stake in it — because you put your stake in throughout the process. The art is the back-and-forth.
There's hard evidence this matters. In 2025, researchers at MIT, Wellesley, and MassArt measured brain activity in students writing essays. Students who used ChatGPT without limits showed lower neural integration, narrower brain activity, and — the most striking finding — 83% could not quote their own essays back from memory. The control group, writing without AI, came in at 11%.
The viral headline was "AI rots your brain." The conclusion the researchers wrote was different: "Strategic timing of AI tool introduction following initial self-driven effort may enhance engagement and neural integration."
Translated: write first, then bring in the AI. The collaboration model wins. The replacement model loses.
The loudest critics on both sides miss this. The doomer position ("AI is destroying creative work") and the maximalist position ("AI does everything for you now") share an assumption: that the human is supposed to step out of the loop. They argue about whether that's good or bad. Neither asks whether stepping out of the loop was ever the right move.
It isn't. The loop is the point.
The word "agentic" is broken. Here's how to fix it.
If we're keeping the human in the loop, we need a working definition of what the AI is for.
The technical definition of "agentic" is reasonable: a system with goals, memory, planning, and tools that can take real action. The SaaS-marketing definition has decayed to: "any LLM with a chat interface inside an app." Bot that answers your PTO questions? Agentic. Cron job that triggers an email sequence? Agentic. Chatbot with a database connection? Agentic.
The version operators should use:
An AI is agentic when you've taught it how to plug into your specific life and workflows, with persistent context and the ability to do things — not just answer.
It centers you, not the model. It makes the bar concrete — persistent context, real action. And it makes the work the agent does yours, because you taught it.
A chat session that resets every time you close the tab isn't an agent. It's a search engine you can argue with. Useful. But it's the floor of what AI can do, not the ceiling.
What's past the chat box
For 400 million people, the chat box IS AI. They've never seen past it because past it requires looking at a ~/.claude/skills/ directory once, and that sounds technical. It isn't. It's a learning curve that takes a weekend.
What's there:
- Skills — folders of markdown that teach the AI specific workflows you use: how you write a quote, triage a support ticket, draft a newsletter, edit a podcast. The AI invokes them automatically.
- MCP servers — let the AI read your Notion, post to your Slack, query your database, navigate your browser. Not metaphorically. Actually.
- CLI + agent mode — Claude in your terminal, with permission to read your files, run scripts, edit code. No copy-paste.
- Hooks — automation that fires on AI events. Lint after every edit. Block dangerous commands. Inject context at session start.
None of this requires being a developer. It requires giving the AI durable instructions about your work — once — instead of re-typing the same context into a chat box every day.
That's the frontier. Not bigger models. Not "agentic" rebranding. Teaching the tool how to live in your workflow.
What to actually do
If you've only ever used ChatGPT or Claude in a browser:
Install Claude Code. Open a folder with real work in it. Ask it to draft a project brief, audit a spreadsheet, or write an outreach email in your voice. Then teach it one thing about how you work — a template you reuse, a tone you prefer, a workflow you run weekly. Save it as a skill. Use the skill next week. That's the unlock.
If you've already crossed that line:
Look at how often you're still copy-pasting context into a chat. Every instance is a skill you haven't written yet. Each one is a five-minute fix that compounds for the rest of the year.
Close
The critics are right about one thing: if you outsource your thinking to AI, you lose your work. That's not a prediction. That's the MIT study.
But the answer isn't to keep AI out of your craft. It's to keep yourself in it. Teach the tool. Direct it. Give it the instructions it's been waiting for.
AI isn't replacing you. It's waiting. And most people never give it anything to do.