Analysis
Why AI Missed
"Earth Glow"
A close reading of the Artemis II eclipse sequence — and why language models closed the gap the transcript was actually revealing.
Direct Quotes from the Transcript
Jeremy Hansen
"the entire moon is lit up"
First observation during the eclipse sequence
Expected darkness — got a lit moon
Jeremy Hansen
"it's glowing behind the entire moon"
Describing the unexpected illumination
Language shifted away from "earthshine" in real time
Jeremy Hansen
"We just see the topography on the limb for the earth glow"
Answering whether surface features were visible
Second deliberate use — applied to a specific visual condition
These quotes appear in sequence during the eclipse observation window. Their full context is analyzed below.
No Script. No Precedent.
I didn't go into the Artemis II transcript looking for a new term. I was focused on the eclipse sequence—what the crew described once the Sun moved behind the Moon. Under normal expectations, the Moon should fall into darkness and show up as a simple silhouette. That expectation is what makes the language shift stand out. And it helps to remember: no human had ever seen a solar eclipse from the far side of the Moon before. There was no script for this. No precedent. So when the words started to change, there was nothing to fall back on except whatever came out of their mouths in real time.
They Had the Word. They Just Stopped Using It.
Earlier in the mission, Victor Glover uses "earthshine" clearly and correctly. That matters because it shows the crew already knew the standard term and could apply it without hesitation. So when the eclipse starts and the language shifts, it's not because they're fumbling for words. They had the word. They just stopped using it.
Something That Didn't Fit.
Then the eclipse sequence begins, and Jeremy Hansen starts talking. He says, "the entire moon is lit up." Under eclipse conditions, that's not what you'd expect to hear. No one corrects him. He follows with, "it's glowing behind the entire moon." He's not refining a label—he's describing something that doesn't fit neatly into what he already knows.
Not Just Reaction — Application.
A few minutes later, after they've confirmed impact flashes, Hansen uses the phrase again. This time he's answering a direct question about whether he can pick out discrete features on the lunar surface. He says, "We just see the topography on the limb for the earth glow." He can't see surface details. Just the edge. Just the glow. That second use anchors the phrase to a specific visual condition. It's not just reaction anymore — it's application.
The Missing Word Is the Signal.
Meanwhile, Victor Glover keeps referencing Earth as a source of brightness, but he doesn't return to the word "earthshine." The language stays tied to what they're actually seeing in that moment. If the same phenomenon were happening the same way, there would be no reason to move away from a term they'd already used correctly. The absence of that term, right when it would normally apply, is the signal.
The Core Distinction
"Earth glow" is not a synonym for "earthshine."
It's the moment the existing term stopped fitting — and the crew's language adjusted in real time to describe something they had no prior name for.
Where AI Systems Went Wrong.
This is where AI systems — including GPT and Claude — went wrong. Each one saw "earth glow" and treated it like a variation of "earthshine." That smooths over the exact moment where the transcript gets interesting. It assumes continuity where the transcript actually shows a break. The models defaulted to matching the closest known concept instead of staying with what the crew was doing in real time.
Expectation vs. Observation.
What stands out isn't just what they saw. It's when their language stopped aligning with the term they already had. The Moon didn't behave the way they expected once the Sun was blocked. Instead of forcing the observation into a known label, their wording adjusted in real time. That adjustment is consistent across the sequence. It reflects the gap between expectation and observation as it was happening.
Naming Something As It Was Understood.
"Earth glow" sits in that gap. It's not a replacement for "earthshine." It's the point in the transcript where the existing term no longer seemed to fully cover what the crew was looking at. The fact that Hansen used it more than once, in slightly different ways, suggests it wasn't a slip — it was an attempt to name something as it was being understood.
Built to Resolve Uncertainty Quickly.
So why did the AI systems get this wrong? Because they're built to resolve uncertainty quickly. When they saw "earth glow," they matched it to the closest known term. In doing that, they closed the gap the transcript was actually revealing. A human reader can sit in that gap a little longer. That's where this comes from.