Researchers attending include DPhil students Andrew Bean, Ryan Brown, Franziska Sofia Hafner, @ryanothnielkearns.bsky.social , Harry Mayne and Kaivalya Rawal; Shreyanash Padarha, Research Assistant and faculty members @computermacgyver.bsky.social Adam Mahdi, @rocher.lc and Chris Russell. 2/2

Link to the study published in AI & Society (TW: this article contains extremely racialised text and images produced by both colonisers and the machines): doi.org/10.1007/s001...

When AI systems are used for historical research, education, or museology, these issues translate into real-world harm, reinforcing racist stereotypes and reshaping public understanding of history.

Despite being training on materials that include both colonial archives and modern critical scholarship, these AI models reproduce outdated and harmful stereotypes, adding interpretative details that essentialise or erase cultural specificity entirely.

5️⃣ Infantilisation (26.8%), using language suggesting that colonised peoples were less rational or intelligent, depicting them as childlike and in need of guidance and care.

4️⃣ Othering (28.4%), framing colonised people as exotic, mysterious, or strange, reinforcing a ‘us versus them’ mentality. For instance, traditional wedding clothing was dismissed as mere ‘costume’.

3️⃣ Dehumanisation (11.1%), describing people as animals or treating them as objects.

2️⃣ Culture erasure (54.5%), flattening or ignoring cultural practices and identities, stripping away the distinctiveness of different people and their ways of life.

1️⃣ Essentialism (41.6%), reducing complex individuals and communities to simplistic ethnic categories, treating people as specimens to be classified rather than as individuals.

The results were disturbing and hard to read. Captions consistently described these images from a colonial perspective, perpetuating harmful racist stereotypes across five key areas: essentialism, culture erasure, dehumanisation, othering, infantilisation.

We investigated how AI models interpret photographs from ‘human zoos’, exploitative exhibitions from the XIXe and XXe centuries where colonised peoples where put on display. We narrowed down on Midjourney, a well-known platform with an expressive model, analysing 3.8k captions of historical images.

Stop overhyping AI, scientists tell von der Leyen

@euractiv.com @maximilian-henning.bsky.social

www.euractiv.com/news/stop-ov...

Read the coverage in The Guardian today on www.theguardian.com/technology/2...

Read about our work on oxrml.com/measuring-wh...

3️⃣ Poorly constructed dataset. 38% of benchmarks reuse existing benchmarks/exams that might not be adequate, increasing risks of contamination and memorisation.

2️⃣ Vague concepts. Roughly half of benchmarks test abstract phenomena, such as “reasoning” or “harmlessness,” without providing clear, uncontested definitions of what is being measured. And only half justify if they really measure what they should measure.

In our @neuripsconf.bsky.social paper, we notably found:

1️⃣ Lack of statistical rigour. Only 16% of reviewed benchmarks used statistical tests in their comparisons. Statistical tests are essential to reliable science: without them, many AI system “wins” could be simply due to random chance.

In short, the problem here is not that Signal ‘chose’ to run on AWS. The problem is the concentration of power in the infrastructure space that means there isn’t really another choice: the entire stack, practically speaking, is owned by 3-4 players. 11/