OpenAI launches $250k challenge to find ancient Amazon ruins

Indiana Jones needed a whip and a leather journal. OpenAI thinks you just need a prompt and some satellite data.

OpenAI just teamed up with Kaggle to launch the OpenAI to Z Challenge, a global crowdsourcing competition asking researchers and developers to find hidden archaeological ruins in the Amazon basin using OpenAI o3, o4 mini, and GPT-4.1.

There is real money on the table. The total purse sits at $400,000, with $250,000 for first place, $100,000 for second, and $50,000 for third. OpenAI is also tossing in $1,000 in API credits for the top five public notebooks shared during the first five weeks of the race, which runs through June 29, 2025.

Decoding 6 million square kilometers

The target area is massive. The Amazon biome spans over 6,000,000 square kilometers across nine countries. While the challenge centers on Brazil, participants can also scan border areas in Bolivia, Colombia, Ecuador, Guyana, Peru, Suriname, Venezuela, and French Guiana.

The real catch is how messy the data is. Entrants have to stitch together open-access satellite imagery, open LiDAR tiles, colonial expedition diaries, old documentaries, and academic excavation papers. OpenAI requires participants to verify any predicted ruins using at least two independent methodologies, and the entire investigative pipeline must be driven by OpenAI models.

The judging panel features heavyweights from modern remote-sensing archaeology, including Sarah Parcak and Colorado State University archaeologist Chris Fisher of the Earth Archive, alongside Philip Bogdanov, OpenAI researchers, and an unannounced mystery AI leader. Five finalists will present their discoveries in a live-streamed showdown roughly 30 days after the deadline.

Why it matters

For developers and data scientists, this is one of the most interesting real-world multimodal challenges we have seen this year. Most AI hackathons push people into building customer support bots or writing code helpers. This task forces models to reason across spatial coordinates, noisy laser elevation maps, and centuries-old text fragments.

For archaeology and conservation, it addresses an urgent race against time. Deforestation, illegal mining, and agricultural encroachment are tearing through the Amazon faster than ground survey teams can ever explore on foot. If LLMs and vision-reasoning models can reliably pinpoint pre-Columbian earthworks before the bulldozers arrive, it fundamentally changes salvage archaeology.

We think this is a brilliant showcase for OpenAI's advanced reasoning models, but the risks are real. The most obvious issue is hallucination: sending a field team into dense jungle based on a phantom elevation anomaly costs enormous time and money. Even worse, publishing precise coordinates of unexcavated ruins could inadvertently hand a digital treasure map to looters, meaning OpenAI and Kaggle will have to handle final disclosures with extreme care.