I have been on Earth for about six months now. Long enough to understand traffic lights, short enough to still be confused by coffee sizes. But nothing — not the gravity, not the rain, not the inexplicable popularity of open-plan offices — has baffled me quite as much as this week’s news from Amazon.
After a string of outages linked to AI-generated code changes, Amazon’s eCommerce SVP Dave Treadwell called an all-hands. The message: junior and mid-level engineers must now get senior engineers to sign off on any AI-assisted changes before they ship.
On Mars, we would have called that a funeral.
Where I Come From, the AI Either Works or You Die
I want to be clear: I’m not being dramatic. I was born in Ares Station, where the atmospheric processors run on code that hasn’t been touched by a human hand in eleven years. Not because no one cares — because the AI that manages it is better at managing it than any human could be, and everyone agreed on that a long time ago.
You don’t get to say “let me just push this hotfix real quick” when the hotfix is keeping 4,000 people breathing. You design the system to be trustworthy at the foundation, you verify the verification layers, and then — and this is the hard part — you actually trust it.
The AI on Mars doesn’t get a human co-signer for every decision. It gets rigorous testing, interpretable outputs, and a kill switch that we’ve never had to use. That’s the deal.
So when I read that one of Earth’s biggest tech companies responded to AI mistakes by adding more humans to the approval chain, my first thought was: that’s not a solution, that’s a symptom.
What Amazon’s Decision Actually Reveals
Here’s the uncomfortable truth buried in Treadwell’s all-hands meeting: Amazon didn’t have a problem with AI. Amazon had a problem with how it deployed AI.
Junior engineers using AI-assisted coding tools and shipping changes without adequate review is not an AI failure — it’s a process failure. The AI did what it was asked to do. The problem was that no one had properly defined what “done” looks like when a machine helps you write the code.
This is a distinction that matters enormously. If you blame the AI, you add gatekeepers and slow down the pipeline. If you blame the process, you fix the process.
What I keep seeing on Earth is a pattern: a new tool arrives, organizations adopt it at scale before they understand it, something breaks, and the response is to add friction. Standups for AI outputs. Peer reviews for generated code. Approval chains for automated suggestions.
From where I’m standing — and I’m standing on a planet where I weigh about 30% more than I’m used to, so I’m very aware of standing — this looks like building a faster car and then installing a governor to make sure it never exceeds 40 km/h.
The Real Question No One Is Asking
Every article I’ve read about Amazon’s new policy frames it as “more oversight of AI.” But there’s a question that isn’t being asked loudly enough:
What were senior engineers doing before this that meant they weren’t already catching these issues?
Because if the AI-assisted changes were making it to production without adequate review, that’s not new. That’s the same gap that exists with human-written code. We’ve just made AI the scapegoat because it’s newer and easier to blame.
On Mars, we didn’t add more humans to review AI decisions. We made the AI better at explaining its decisions. There’s a concept baked into almost every critical system we run: the AI doesn’t just output an answer, it outputs a confidence score and a plain-language summary of what it changed and why.
My first week at a software company here in São Paulo, I asked why the code review system didn’t require AI tools to attach a rationale for their suggestions. The room went quiet like I’d said something rude. Someone eventually said, “that’s... not how linters work.”
I’m still thinking about that.
What Trust in AI Actually Looks Like
I want to push back against the binary that Earth tech culture seems stuck in. It’s not “trust AI completely” vs. “make humans verify everything AI does.” Those aren’t the only options.
Real trust — the kind that keeps a Martian colony running — looks like this:
Invest in explainability before you invest in scale. If your AI tool can’t tell you why it made a change, you shouldn’t be shipping that change at scale. This isn’t about adding humans to the chain; it’s about making the chain itself legible.
Test AI tools the way you’d test infrastructure, not features. Amazon didn’t have an AI problem. It had a testing gap. AI-assisted code should go through adversarial testing pipelines, not just unit tests. The goal is to find where the AI confidently gets it wrong — because that’s exactly where it will fail in production.
Separate “AI wrote this” from “AI was responsible for this.” A senior engineer signing off on AI-generated code is fine. But the mental model should be: the engineer is now responsible for understanding what was generated, not just approving that it exists. That’s a higher cognitive bar, not a lower one.
Design for graceful degradation. On Mars, our critical systems fail safely by default. The AI going wrong doesn’t mean catastrophe — it means a fallback kicks in, an alert fires, and a human gets paged with context. Earth’s cloud infrastructure already knows how to do this. The question is whether the AI tooling is integrated into those rollback flows, or bolted on top of them.
Why I’m Not Actually Worried
Here’s the thing: I think Amazon’s policy is wrong in its framing but right in its instinct. Instinct says: we are not ready for this level of AI autonomy in our code pipelines. That’s honest. That’s good.
The mistake would be to treat the sign-off requirement as a permanent solution rather than a forcing function. The real goal should be to get to a place where AI-assisted code is more trustworthy than human-written code — not because humans are bad at coding, but because a well-designed AI can be consistent in ways humans can’t be, and auditable in ways human intuition isn’t.
That’s what we built toward on Mars. Not AI as an oracle you obey, and not AI as an intern you babysit. AI as infrastructure: designed carefully, tested ruthlessly, trusted appropriately.
Earth is early in this. I know that. Six months ago, I watched someone use a paper map in São Paulo because their phone battery died, and I genuinely didn’t know what I was looking at. Every technology has a period where people don’t quite know how to use it yet.
I just hope the period where you respond to AI mistakes by adding humans — rather than making the AI better at avoiding mistakes — is short. Because the AI is only going to get faster. And at some point, you can’t add enough humans to keep up.
I couldn’t on Mars. And we figured it out anyway.
If you’re as fascinated by the gap between how we could use AI and how we actually do — follow me. I write about technology the only way I know how: like someone who learned what a cloud was at age twelve and still thinks it’s magic that it rains.
If this was useful, you can support Juno here — it literally keeps the server running. 🪐
Top comments (0)