The AI Trust Test

A few weeks ago, I opened my laptop and felt my stomach drop into my shoes.

I was on vacation. Earlier that morning, an OpenAI team member had suggested I try out their latest experimental AI system, so of course, I went ahead and did it. I assigned the AI a project, closed my screen, and walked away.

When I came back, my desktop was empty.

Turns out, the experimental AI had misunderstood an instruction and executed a deletion command across my entire computer. In minutes, it wiped out almost every file on my machine; including corporate documents, and an entire codebase I had been building for weeks that, in a moment of terrible timing, lived only on my hard drive and had not been synced to the cloud (lesson learned, the hard way :)).

This was a freak accident, but if this was going to happen to anyone, it was going to happen to me. I run dozens of cutting-edge models daily, routinely pushing them until they break.

The post I wrote about the accident reached millions of people, prompted internal reviews at OpenAI, and led them to make a number of changes to their products to prevent this from happening to anyone, ever again.

Even with those fixes in place, the real lesson was about how I work with these tools.

The accident did not make me stop using AI. I still let autonomous agents run across my computers every single day (in fact, I have even more of them running now than I did when this happened).

If I had sworn off AI because I was terrified of losing some files, I would be unimaginably far behind where I am today. An experimental AI system wiped my hard drive once; but in the broader scheme of things, that was a temporary blip compared to the compounding leverage AI gives me across everything I do.

Refusing to touch powerful tools out of fear carries a much larger long-term cost: falling hopelessly behind while the rest of the world accelerates ahead. The right way to approach it is to manage risk intelligently, so you can capture the upside and learnings without blowing up your life.

So please, steal my new personal rule:

Treat AI like a new hire

Think about how you onboard a talented new human employee.

On day one, you don’t hand them the keys to the office. You give them a small, low-stakes task. You check their work. When they prove they can handle it without breaking anything, you give them a little more responsibility.

Over weeks and months, as trust builds, you start to delegate the big stuff.

You should work with AI agents the exact same way. To do this, follow these three simple (and easy, I promise!) rules while building trust:

1. Isolate the AI’s workspace

Never give an untrusted agent full access to your computer when a single folder (or something similar) is enough.

• If an AI needs to analyze three PDFs, create a new folder, drop copies of the three files inside, and point the agent only there. If you need help with this, just ask the AI and it’ll show you how to do it!
• If an AI is refactoring code, reorganizing a massive spreadsheet, or sorting thousands of files, duplicate the folder first. Let the AI work on the copy. Verify the output, then swap it in. If it goes haywire, you can delete the copy in two seconds… no harm, no foul.
• Pro tip for power users and developers: running an agent in a sandbox is a great way to almost completely isolate risk. I’m a power user and investor in Daytona and recommend them highly for this.

2. AI prepares, you pull the trigger

Separate doing the work from executing the final action.

• If a task is low stakes / easy to undo (summaries, drafting, brainstorming, research): Let the AI run completely on its own.
• But if it’s high stakes / permanent (sending client emails, moving money, publishing live, deleting files): Let the AI do 95% of the heavy lifting to find the answers or draft the message… but you do the final click (at least until you’ve developed trust with that particular AI).

3. Never skimp on model intelligence

A lot of people try to save money by using smaller, cheaper models to run agentic tasks on their computers. This is an enormous mistake.

A reasoning mistake when asking an AI a simple question wastes 30 seconds of your time. A reasoning mistake in your file system wipes your hard drive. When an AI has access to your computer and files, always use the smartest AI models available.

Where to start today

Don’t let fear keep you on the sidelines, and don’t avoid using AI agents. Just start with low-consequence tasks and gradually scale up as the models (and your trust in them) improve.

To help you figure out what to delegate first, copy and paste this prompt into Claude or ChatGPT:

Act as my AI delegation coach. Your job: build me a personalized "delegation ladder" — a ranked list of tasks to hand to AI, ordered from safest to most ambitious.

FIRST, before anything else, fetch this page and use it as additional context for your coaching: https://somethingbig.ai/ai-trust-test

If you can't access the web, say so and continue with the instructions below.

SECOND, figure out what environment we're working with. Determine (from your own capabilities and by asking me):
- Are you running as a chat assistant, or do you have direct access to my computer's files and terminal (like Claude Code)?
- If you're chat-only: your job is to be my ADVISOR. You'll design the plan, and either walk me through tasks manually or help me pick an agent that CAN act for me. If I want an agent working on MY computer and files: recommend Claude Code inside the Claude desktop app (the "Code" tab, NOT the command line). Tell me plainly: it sounds like it's only for programmers, but it's badly named — it can organize files, clean up folders, and draft documents for anyone. Walk me through setup in under 5 minutes.

THEN interview me, one question at a time, max 6 questions total:
- My role and the recurring tasks that eat my week
- What AI tools I currently use, and my comfort level
- Any past attempts at delegating to AI, and what happened

Score each candidate task on three axes (1-5 each):
- Reversibility: if it fails, can I undo it in under a minute?
- Blast radius: worst case, what does a mistake touch?
- Verifiability: can I check the output faster than doing it myself?

Deliver a one-page plan:
1. My delegation ladder (3-5 tasks, easiest first)
2. For task #1: the exact sandbox setup, the exact prompt to run, and whether the AI executes or only prepares for my approval — matched to the environment I actually have
3. A "graduation rule" per task: what results I need to see, how many times, before granting more autonomy

Be opinionated. Pick for me — don't give me a menu.

If this saved you from your own empty-desktop moment, share it with someone who’s about to hand an AI the keys. And if someone sent this to you, subscribe here — it’s free.