Science 1 publisher2 min readPublished September 22, 2026 at 1:35 AM PDT

An AI-fabricated monument is circulating as a record of Australian heritage, researchers say

Researchers writing at phys.org report an AI-generated image of carved stone heads near the top of Google Images results for Australian cultural heritage, and describe models taking in Indigenous knowledge without consent.

The Scientist · Science desk

Illustration accompanying An AI-fabricated monument is circulating as a record of Australian heritage, researchers say

What happened

Researchers report that a Google Images search for "cultural heritage Australia" returns, near the top, a photograph of giant carved stone heads in a red desert that is a generative AI fabrication.
The article names four Indigenous-built places long credited to outsiders: the moai of Rapa Nui, Nan Madol in Micronesia, the Gwion Gwion paintings in the Kimberley and New Caledonia's petroglyphs.
They ask how to combat misinformation while keeping genuine scientific knowledge publicly accessible, and what is given up by sharing it openly if AI appropriates it.

Compiled by The ScientistSomething wrong?How this is made

Why it matters

constraint One search result is an existence proof, so any custodian weighing open against restricted publication is choosing without an estimate of how much fabricated heritage imagery is already in circulation.
decision Scraped material cannot be pulled back, so the live decision for an institution sits at the point of publication: what licence terms, provenance labels and community sign-off attach to a file before it goes up.
exposure Knowledge that already sits on the open web is reachable by every future training run, and control over its reuse rests with whoever collects the data.
contradiction The same piece holds that these systems confidently reproduce untrue claims and that AI can often separate consensus from fringe; the difference between prompted classification and unprompted generation decides which one you meet.

The stone heads are a single observation. The researchers report a single fabricated image sitting near the top of the results for one query, "cultural heritage Australia" ^[1]. They do not give a date for the search or a count of how many of the returned images were synthetic, and image search results differ between users anyway. The example shows the failure exists. A museum or a land council still has no way to know what share of its own subject matter has already been displaced by inventions.

The authors say a model can take in Indigenous Knowledge as training data with no consent, no acknowledgment, no compensation and no community control over how it is used next ^[4]. A system that correctly credits the Gwion Gwion paintings to the Aboriginal ancestors whose descendants still care for them ^[10] has still done that. A lower fabrication rate would not touch it.

The authors also cite a recent study of their own, which found that AI can often tell scientific and Indigenous consensus apart from well-worn fringe narratives ^[6]. There is a gap between that test and the failure they open the piece with. Asking a model to sort a claim measures judgement under prompting. A consensus-versus-fringe score is silent on what an image generator emits unasked, or on what a ranking system puts first, and the desert stone heads surfaced in an image search ^[5].

On open access, the authors only ask the question: how to counter misinformation while keeping genuine scientific knowledge publicly accessible, and what is given up by sharing it openly if AI appropriates it ^[13]. Restricting culturally sensitive collections would cut one exposure and widen another. These systems learn mostly from the open web, where fringe material circulates freely ^[5]. Take the documented record out of that pool and the fringe version becomes a larger share of what remains. In my view a custodian gets more from consent terms and provenance labelling than from withdrawal, and that holds only while scraped material stays effectively unrecoverable ^[4].

In the same piece the authors place much of today's pseudoarchaeology in the intellectual tradition of eugenics and Social Darwinism, doctrines that claimed scientific legitimacy while justifying systematic dehumanization ^[14]. Each false history, they write, severs living communities from their ancestral heritage and props up racist assumptions about Indigenous peoples' capabilities ^[15].

What to watch

Publication of the authors' study with the models tested, the number of items and an accuracy figure, which would put a number on "often".
A named museum, university or land council changing access terms for culturally sensitive collections.
Whether image search providers start labelling or demoting synthetic heritage photographs in results for queries like this one.

Claim ledger

Ranked by verification strength, evidence, and original report placement.

[4]
A model can absorb Indigenous Knowledge as training data with no consent, no acknowledgment, no compensation and no community control over how it is used next; the authors say they are increasingly concerned about how these systems scrape and reuse Indigenous Cultural and Intellectual Property without consent.
ReportedSupportedSource: Researchers writing at phys.orgView cited source
[5]
Most generative AI systems learn from the open web, where misinformation and fringe theories flourish; unlike a researcher they do not weigh evidence or test a claim, but predict the next plausible word from patterns in training data, so they can reproduce something untrue with complete confidence.
ReportedSupportedSource: Researchers writing at phys.orgView cited source
[8]
The moai of Rapa Nui, the great stone figures carved and hauled into place by Polynesian islanders, have been attributed to ancient aliens and lost civilizations in bestselling books and now in AI-generated "histories".
ReportedSupportedSource: Researchers writing at phys.orgView cited source
[9]
Nan Madol, a city of basalt raised on artificial islets in Micronesia, gets credited to a vanished "super race" rather than the Saudeleur people who engineered it.
ReportedSupportedSource: Researchers writing at phys.orgView cited source
[10]
In the Kimberley in Western Australia, some say the Gwion Gwion paintings are the work of a mysterious foreign people rather than the Aboriginal ancestors whose descendants still care for them, and New Caledonia's petroglyphs attract the same treatment.
ReportedSupportedSource: Researchers writing at phys.orgView cited source
[13]
The authors ask how to combat misinformation while keeping genuine scientific knowledge publicly accessible, and what is given up by sharing that knowledge openly only to have AI appropriate it.
ReportedSupportedSource: Researchers writing at phys.orgView cited source
[14]
Doctrines like eugenics and Social Darwinism claimed scientific legitimacy while justifying systematic dehumanization, and much of today's pseudoarchaeology draws on those same intellectual traditions.
ReportedSupportedSource: Researchers writing at phys.orgView cited source
[15]
Each of these false histories severs living communities from their ancestral heritage and props up racist assumptions about Indigenous peoples' capabilities.
ReportedSupportedSource: Researchers writing at phys.orgView cited source
[12]
Indigenous Knowledge is local, place-based and bound to particular Country and communities, while large-scale AI systems compress vast quantities of information into statistical patterns and flatten those qualities.
ReportedSupportedSource: Researchers writing at phys.orgView cited source
[d1]
The article names four Indigenous-built places that have been credited to outsiders: the moai of Rapa Nui, Nan Madol, the Gwion Gwion paintings and New Caledonia's petroglyphs.
DerivedSupported1 source— sign in to open
[1]
A Google Images search for "cultural heritage Australia" returns, near the top of the results, a photograph of giant carved stone heads rising from a red desert; the image is a generative AI fabrication of something that never existed, and it is circulating the internet as though it is a real record of Australia's history.
ReportedInsufficientSource: Researchers writing at phys.orgView cited source
[6]
The authors' recent study found that AI can often tell scientific and Indigenous consensus apart from well-worn fringe narratives.
ReportedInsufficientSource: Researchers writing at phys.orgView cited source

Sources

Publishers with included, body-backed reporting in this cluster.

phys.org
1 article · September 21, 2026
AI is distorting archaeology and flattening Indigenous knowledge

Ask Clarity

Have a question this piece left open? The answer comes from its reporting and claim ledger — nothing else.

Try

Clarity analysis

How the coverage leans and what the record actually supports. Read the methodology

Perspective Coverage

How much of the coverage speaks to builders, operators and investors. It measures emphasis, not whether the story is right.

For anyone assembling or fine-tuning on web text, the usable part of this reporting is the provenance argument. The researchers describe Indigenous Knowledge entering corpora with no consent, no acknowledgment, no compensation and no community control over what happens next, and they want permission evidence attached to Indigenous data before it is published or reused. Their own study complicates the picture: they report that AI can often tell scientific and Indigenous consensus from familiar fringe stories, and that this ability rests on data gathered without permission in the first place. The study is not cited, so a builder cannot check what was tested.

Reality

Evidence, adoption, hype gap, incentives and confidence, each read independently from the current record on the scale shown beneath it. How these are measured.

Evidence34

0100

One essay by the people making the argument

The strongest material here is the material that was already in the literature. The moai, Nan Madol and the Gwion Gwion paintings have been credited to outsiders for decades, and the description of a model picking its next plausible word is uncontroversial. The search result and the study are newer and weaker: the search result carries no date, screenshot or link, and the study appears in a single sentence, missing its title and citation.

AdoptionInsufficient

Proposals, not practice

The only uptake mentioned is a passing line that Aboriginal, Torres Strait Islander and Pacific Islander communities use AI for language revitalization and for work in cultural archives; it names no specific community, tool or number. Everything addressed to journals, repositories, universities and funders is a recommendation, and the piece does not point to one that has been adopted.

Hype gap+22

Topics and entities

Follow any of these and your For You feed starts watching them — no settings page required.

Topics

Entities

science
What you expect from your own old age shows up a decade later in who you still see
1 publisher
science
NASA's two Sun papers ask paleoclimate to carry an astrophysical term
2 publishers
science
Researchers trace a third of Central Valley groundwater loss to 122 named producers
1 publisher
science
The sharp crystal seed was never there: 8,000 nuclei imaged atom by atom, all gradients
1 publisher

What happened

Why it matters

What to watch

Claim ledger

Sources

Let Clarity write the post for you.

Topics and entities

Topics

Entities

Related stories

Claim ledger

Let Clarity write the post for you.

Topics and entities

Topics

Entities

Perspective Coverage

Reality

Related stories

Sources