Data Sovereignty vs. Digital Prospecting

Aerial of the Amazon rainforest in the Para’ state, Brazil (October 6, 2020) | European Union, Copernicus Sentinel-3 imagery / Contains modified Copernicus Sentinel data 2020

“Now, for the first time in history, anyone can conduct archaeological research,” or so promised artificial intelligence giant OpenAI in a recent promotion-via-public competition. Marketed as an Indiana Jones–style adventure, the “OpenAI to Z Challenge” invited users to digitally scour the Amazon rainforest for archaeological treasure. More than 2,500 teams joined. Many entrants were data scientists or hobbyist coders drawn from the global Kaggle platform, where machine-learning enthusiasts compete in public challenges. The competition’s framing, complete with livestreamed finals and a “mystery AI leader,” was designed to appeal to a tech-savvy online audience rather than to archaeologists or local stakeholders. The winners received $250,000 in cash and Application Programming Interface credits, along with “funding to continue their work in collaboration with archaeologists.”

The “finds” were not verified archaeological sites but algorithmically generated “hotspots,” places that the contestants’ models predicted might contain buried structures. Submissions were judged on how convincingly they combined different types of open-source data (satellite imagery, lidar, historical maps, and even Indigenous oral maps) into a plausible case. Contestants were not excavating or confirming sites themselves; their role was to produce digital predictions. OpenAI has said that the winners will have opportunities to collaborate with archaeologists and potentially extend their methods into field projects, though it has not specified whether these would involve the Amazon “hotspots” flagged in the competition or unrelated future work.

A three-person team, known only by the collective moniker “Black Bean,” took first prize, identifying 67 one-square-mile “hotspots” across the Amazon that AI models flagged as potentially rich in undiscovered archaeological artifacts. Yet beyond their onscreen usernames (“Yao,” “Yingjie Zhang,” and “tianmiao11”), little is publicly known about who these participants actually are. The fact that the competition’s winners remain effectively anonymous highlights a deeper issue: OpenAI’s claim that “anyone” can be an archaeologist is borne out not by trained researchers or local experts but by screen names detached from identifiable people or communities.

The event was designed for spectacle. OpenAI let the public vote. The scoring rubric emphasized “impact” and “ingenuity” but offered no clear standards for ethics, community engagement, or transparency. Instead, entries were graded on presentation design, novelty, and the cleverness of the data overlays. Points were awarded for smooth visuals, pacing, and reproducibility, while “rules compliance” was limited to checking that links opened without paywalls and that at least two public data sources were cited. Historical texts and oral maps appeared in the same grab-bag category as lidar tiles and satellite imagery, with no criteria for evaluating how responsibly such sources were used. In other words, the contest rewarded technical flair and rhetorical polish, not accountability to the people or places its data purported to represent.

Despite the fact that OpenAI’s large language models have been found to produce and echo biases and stereotypes both in the underlying data used and in pattern recognition, there was no requirement to include information on error rates, bias audits, or the ways models were trained. If the models are trained primarily on regions, sources, or site types that have been well studied in Western or colonial archaeological traditions, they will be better at “finding” the kinds of structures those traditions valorize (stone foundations, monumental architecture) and worse at detecting Indigenous or less “visible” sites (earthen mounds, seasonal encampments, sacred landscapes). This reproduces long-standing hierarchies of what counts as “archaeology.”

The contest leaned heavily on colonial archives and Indigenous oral histories, repackaged as “open data” ready for algorithmic mining. But what does it mean to “discover” cultural heritage on others’ lands through code? And where were the safeguards to ensure that communities’ rights and cultural heritage were protected?

Contestants, branded as “digital explorers,” mined satellite imagery, lidar scans, and Indigenous maps to locate sites. Such digital “prospecting” evokes a long history of outsiders extracting knowledge and artifacts without consent. Many of the so-called “open” datasets used in the competition were created by governments or corporations without consulting affected communities. While no Indigenous groups have yet publicly objected to this specific contest—perhaps because no physical interventions have taken place—the absence of consultation itself is part of the problem. Such broad data usage ignores Indigenous Data Sovereignty principles, which call for community control over data about their lands and heritage.

Allowing AI developers to scrape, analyze, and publicize sensitive cultural information legitimizes unauthorized extraction, potentially violates national heritage protection laws, and perpetuates historic imbalances. OpenAI also made no mention of Free, Prior, and Informed Consent (FPIC), a global ethical standard for research involving Indigenous communities.

Despite marketing itself as democratizing archaeology, the OpenAI challenge favored those with access to high-speed computing, technical infrastructure, and AI expertise—resources that many local and descendant communities simply don’t have. While anyone theoretically could have participated, in practice the field was stacked in favor of well-funded developers far removed from the Amazon. The winning team, Black Bean, remains effectively anonymous. That very opacity makes it impossible to know whether they are professional data scientists, hobbyists, or something else entirely. What’s clear is that they are not identifiable community stakeholders. One reason Indigenous peoples may not have raised objections is that they are rarely present in these highly technical competition circuits to begin with, a structural exclusion that makes the absence of protest an unreliable signal of consent.

The looting of archaeological sites is a concern worldwide. Brazil, where the contest was focused, has strict heritage laws and a federal agency, IPHAN, that oversees archaeological research. Exact site locations are often withheld to prevent looting, which can fund organized crime and extremist groups. While the contest did not authorize or encourage excavation, it nonetheless amplified the visibility of potential “hotspots.” For now, the “hotspots” remain digital predictions. Any excavation in Brazil would still require permits from IPHAN. But the contest nonetheless increased the visibility of potential site locations by circulating them in a high-profile global forum. Even if no one digs tomorrow, publishing coordinates or maps of plausible sites can attract attention from actors outside legal frameworks.

None of this is to say that technology cannot serve archaeology responsibly. AI applications are becoming more widespread in archaeology. Among other things, they have been used to reconstruct submerged landscapes and shipwrecks, to monitor looting, and to catalog and reconstruct artifacts. Unlike OpenAI’s competition, however, these projects are collaborations between trained archaeologists knowledgeable in local and national cultural heritage laws and computer scientists; they typically receive significant external review; and they are directed with specific research goals in mind rather than the online challenge’s needle-in-a-haystack approach. Responsible archaeological research is slow, collaborative work that requires permits, community involvement, and transparency. By bypassing these norms, OpenAI normalized a “move fast” approach that prioritizes publicity over trust and rigor.

Now that the winners have been crowned, OpenAI and the tech community at large must decide whether to treat archaeology as a branding opportunity or a serious discipline deserving of ethical and scientific engagement. Codesigning projects with archaeologists and local stakeholders, seeking FPIC before publishing findings, creating meaningful review processes, and committing to transparency in data sourcing and model training should all be minimum standards. Just as importantly, AI companies should affirm that governments and communities, not corporations, retain authority over their cultural data and how it is used.

Only with such reforms could a contest like this be anything more than a high-profile marketing scheme. As it was, the OpenAI to Z Challenge was just the latest chapter of a very old story: outsiders exploiting local, descendent, and Indigenous knowledge and lands in the name of “progress.” Only this time, the conquistadors came bearing neural nets.