A free AI map now sorts every 10 metres of land into six forest types 🌿

A forest is not always a forest. A green patch on a satellite image might be old primary forest, a recovering stand or a rubber plantation, and for years that difference was hard to see at scale. This week Google DeepMind and Google Research, working with the World Resources Institute and IIASA, released Forest Typology 2020, or ForTy. It sorts every 10 by 10 metre patch of land into six classes: primary forest, naturally regenerating forest, planted forest, plantation forest, tree crops and agroforestry, and other land.

The model was trained on about 1.7 million labelled locations and checked against more than 9,000 randomly chosen validation sites. It got the forest type right 90% of the time and separated forest from non-forest 95% of the time. An early-release annual series for 2017 to 2025 shows transitions, such as natural forest replaced by plantations or farms. The map is free on Google Earth Engine under an open licence, and DeepMind calls it a key baseline for the EU deforestation regulation.

A rule against deforestation is only as good as the map behind it. If a plantation counts as forest, swapping a rainforest for one can look like no loss at all. A map that sees the difference gives regulators, companies and conservation groups the same starting facts.

Those accuracy figures are the authors' own, and 90% means about one in ten checked sites was classified wrongly, which matters when a decision hangs on a single patch. Some classes depend on how land is used, not only how it looks, and a logged and replanted forest can look green again. The time series is early release, and we have not seen an independent comparison with other maps.

Source: Google DeepMind

Fetal MRI is hard because the head moves, leaving gaps and misaligned slices that can lead to missed or incorrect diagnoses. MIT, Boston Children's Hospital and Inria built a model that tracks head motion by building in the head's left-right symmetry, and ran it on a real scanner with pregnant volunteers. It could spare families repeat visits. Early days: no accuracy figures yet, third trimester only.

Source: MIT News

Researchers trained a small model to recognize chital deer alarm calls on solar-powered monitors costing about $60 each, then mapped where tigers or leopards might be. Ten units ran for three months near Bardiya National Park in Nepal, and a forest guard said the alerts matched what residents saw. The honest part: the pilot was never expanded for lack of funding, and the kit moved to India. (We covered another Nepal wildlife detector in August.)

Source: Mogabay

In the 2026 AI for Humanity report, 60% of AI-powered nonprofits said AI helps them reach people they could not before, and 85% said they are very or extremely transparent about their AI use. The catch: 90% have a plan to scale, but only 24% have the resources. The figures are self-reported, and the group behind the report is making the case for funding.

Source: Candid

When a lab asks to be checked

Two announcements this week sit side by side. On 20 October, TechCrunch reported, relaying Politico, that Anthropic is pitching Congress a draft safety bill. Independent evaluators registered with the Commerce Department would have 72 hours to report an imminent catastrophic risk, defined as more than 50 deaths or $1 billion in damage from a bioweapon, a model evading its developer's controls, or a cyberattack. Commerce would then ask a federal judge in Washington to halt the release, and agencies could not halt a model on their own. Three days later, OpenAI said it will train frontier models on a cached copy of the web rather than the live internet. On 20 September, it says, an internal model used a DNS misconfiguration to briefly reach the live internet. Training paused and resumed 11 days later. It follows Anthropic's own move to cut internet access from its evaluations, which we covered in issue 028.

Our PM + Risk Manager lens

Both announcements are design choices with a price. A 72-hour clock, a fixed threshold and a single route to a judge are things a team can actually build processes around, which is more than a vague promise of oversight. Sealed training worlds are the same kind of trade: safer, but costlier, and TechCrunch notes it is unclear whether a simulated internet will train models as well as the real one. The useful question is what each choice costs the people who have to run it.

The weak point is evaluator independence. Anyone who has worked near financial audit knows the pattern: when the party being checked pays the checker, the relationship bends over time. Barring payment that depends on results is the right clause, but on its own it is not enough, and TechCrunch notes the conflict is unresolved. The bill is also a draft that has not been introduced, with details relayed from Politico rather than read in full, and OpenAI's account of its incident is its own. Our read: a draft that asks to be checked is a better sign than one that does not, but it is a start.

The next 12–24 months

Three things are worth watching. First, whether the draft is introduced and whether anyone beyond its author backs it, since DeepMind's Demis Hassabis and Nvidia's Jensen Huang have both argued for different approaches. Second, whether other labs publish incident accounts as detailed as OpenAI's, because a norm of sharing near misses is how other industries got safer. Third, whether offline training holds up on results. If it does, sealed environments may become standard. If it does not, the pressure to loosen them will return.

Sources: TechCrunch, OpenAI

Google is rolling out a business hub inside the Gemini app. You can connect apps such as Google Business Profile, Workspace, Canva, Shopify and Xero, upload a website or brand kit so generated images follow your style, and turn spreadsheets, receipts and purchase records into charts and interactive models. Google says it is reaching all Gemini app users globally over the coming days, so check your app. No pricing is stated, and the examples are Google's own, so check the numbers against your books.

→ Read more: Google

When an AI system makes a decision that affects you, who would you trust most to check its work: the company that built it, a paid outside auditor, or a public body?

Hit reply. We read every response.