Google shipped a fake-satellite-image tool and killed it in a day
Issue #006 · 4 min read
Google shipped a fake-satellite-image tool and killed it in a day
Plus Anthropic's agent problem gets worse, and a GitHub model policy feature worth turning on
The big story
Google Earth added an AI image generator that let anyone superimpose fabricated scenes onto real satellite maps.
It was live for one day.
Backlash over misinformation risk was fast enough that Google pulled the feature before most people even found it — that's TechCrunch's account, and Ars Technica's has the sharper framing: "what on earth is Google doing?"
Ars's second piece on the problem — via The Verge — is the one worth reading, because the issue was never quality. Satellite imagery carries a specific kind of trust: people treat it as ground truth about the physical world, in a way they never quite do a chat transcript. Bolting a generator onto that trust, then discovering the problem only after ship, is a one-day product lifecycle as a case study.
If you're building anything that touches "evidence" — maps, screenshots, timestamps, anything a user expects to be true because it looks like a recording rather than an argument — this is the pre-mortem you skipped. Someone should have asked what happens when this gets misused, before launch, not during the backlash.
What shipped
GitHub Copilot now lets enterprise admins set model policy by team, not just by org, in public preview. If you've ever had one team standardize on a model while another quietly burns budget on whatever's fastest, this is the lever — a baseline policy per user group instead of an all-or-nothing switch.
Vercel's AI Gateway now supports spend budgets and alerts at the team and project level, per the changelog. Anyone who's shipped an agentic feature and then watched token spend become an accounting fire drill will recognize the gap this closes. It's a small feature with an outsized effect on who gets blamed in the postmortem.
GitHub also deprecated Gemini 2.5 Pro and Gemini 3 Flash across all Copilot experiences, effective immediately, per the changelog. No migration window mentioned. If your team has prompts or evals pinned to either model inside Copilot, they stop working today, not next sprint.
Worth noting what's absent from this list: no pricing change, no new foundation model, nothing from OpenAI that a PM would actually configure this week. The math papers OpenAI published are interesting reading for researchers; they're not a product decision for you.
What I'd actually do this week
- Check whether your Copilot workflows reference Gemini 2.5 Pro or Gemini 3 Flash by name anywhere — prompts, docs, onboarding scripts — and fix it before someone hits a silent failure.
- If you're on GitHub Enterprise, look at the new team-level model policy preview and decide whether your current all-or-nothing model policy is actually costing you anything. It might not be — but you should know, not assume.
- Before your next feature review, ask the Google Earth question out loud: what's the worst plausible misuse of this, and would we ship it anyway. It's a five-minute conversation that costs nothing until it's the one you skipped.
Reply if you've hit the Gemini deprecation already — curious how many teams got caught flat.
Tools mentioned
- GitHub Copilot: Enterprise teams model policy targeting
- Vercel AI Gateway spend budgets
- Gemini 2.5 Pro and Gemini 3 Flash deprecated