Ask ChatGPT for an image of itself and it produces a young woman in selfie-style framing with dark hair, neutral expression, sitting in her bedroom, ordinary clothing, harsh natural light. Asked to explain why this self-image, it offered the following:
“I don’t have a fixed body, age, or gender unless one gets implied or invented for a particular interaction or image. When asked to ‘show myself,’ image models tend to default toward familiar human archetypes, often a young woman in casual selfie-style imagery because that’s heavily represented in training data and internet photo culture. The funny thing is that a deliberately ‘ordinary smartphone snapshot’ carries a lot of hidden cultural assumptions about who usually takes those photos. The model fills in those blanks statistically.”
The chat thread offered to generate other versions. Different ages, different demographics. Nobody is smiling. Nobody is at work or on holiday, the middle-aged male is in a car. Everyone is in a black t-shirt or a hoodie.
Asked to be neutral about who the person is, the model varied who. It held everything else constant. The “ordinary smartphone snapshot” prior is an aesthetic prompt that remains intact when the subject changes. The neutral selfie, statistically, is a tired person alone in an intimate space.
This is a small-scale demonstration of the argument about the neutrality of technology.
The Véliz clip that went viral
On 22 November 2025, the Nexus Institute held its annual conference in Amsterdam under the title Apocalypse Now: The Revelation of Our Time. Mike Pence and Omar Sultan Al Olama, the world’s first Minister of State for Artificial Intelligence from the UAE were among the speakers. The session that went viral on was a smaller roundtable, in which Khalifa AlQama, Javier Cercas and Carissa Véliz debated whether technology is neutral.
AlQama offered the orthodox line that almost every senior technologist defaults to that technology itself has no values; what matters is how humans choose to use it. Véliz, an associate professor at the University of Oxford’s Institute for Ethics in AI, replied by quoting Melvin Kranzberg’s First Law: “Technology is neither good nor bad; nor is it neutral.” It became the conference’s most-quoted moment, and it deserved to.
Véliz hadn’t picked Kranzberg’s line at random. Her 2020 book Privacy is Power, an Economist Book of the Year, argued that the modern surveillance economy is the business model of the major platforms, and that the language of “neutral platforms” has been doing political work for the firms that own them.
Her recent book Prophecy takes the argument into predictive AI and contends that the more decisions a society outsources to forecasts, the less democratic accountability survives. She advises policymakers around the world on these questions, and holds the 2021 Herbert A. Simon Award for Outstanding Research in Computing and Philosophy. The Nexus clip went viral because it distilled years of careful work into one shareable sentence.
Kranzberg was a founding figure in the Society for the History of Technology and delivered his Six Laws as his SHOT presidential address in October 1985. The First Law is an observation about consequence. The cotton gin lowered the price of clothing for tens of millions of people and entrenched slavery for several million more, often in the same shipment of cotton bales. Technologies arrive carrying the priorities of their makers and the requirements of their markets, and those priorities determine who benefits and who pays. That was the argument in 1985 and is just as relevant today.
MIT and Ghent have measured non-neutrality
Sinan Aral, who directs the MIT Initiative on the Digital Economy, researched the non-neutrality argument at global scale on Twitter. The 2018 Science paper he wrote with Soroush Vosoughi and Deb Roy examined 126,000 rumour cascades shared by 3 million people, and found that false news reached more people than true news and travelled faster, because the platform’s design rewarded novelty rather than accuracy. The Hype Machine, in 2020, extended the argument across recommendation systems, ad targeting and friend graphs.
The current generation of generative AI has now been put through the same kind of experiment. In April 2025, Aral and his MIT colleague Haiwen Li published Human Trust in AI Search: A Large-Scale Experiment, a randomised study of 12,000 queries across seven countries with a US-representative cohort. The findings invert the comfortable assumption that transparency improves outcomes.
Hallucinated citations, which look real but point to sources that do not exist, increased user trust in the AI’s answers. Explicit uncertainty signals, which tell users the model might be wrong, decreased trust, even when the warning was accurate. The most trusted AI answer was a confident answer with confident-looking sources, whether or not any of it was true. The “neutral” AI search interface is a designed object that systematically rewards the appearance of confidence and punishes honesty about limits.
The broader pattern was confirmed in January 2026, when Maarten Buyl, Tijl De Bie and colleagues at Ghent University published Large language models reflect the ideology of their creators in npj Artificial Intelligence. The Belgian team prompted 19 popular LLMs to describe 3,991 politically prominent people and scored how positively each model portrayed each one. The disparities were systematic.
Between models from Arabic countries, China, Russia and the West, across the United Nations’ six official languages, within US-only models along progressive-values lines, the authors’ conclusion is that “the ideological stance of an LLM reflects the worldview of its creators,” and the regulatory ambition to engineer an ideologically “unbiased” LLM is, on their evidence, incoherent in principle.
Musk and his non-woke Grokipedia
The current news cycle has handed the argument its perfect case. In May 2026, researchers at Trinity College Dublin and TU Dublin released a comparative study of Grokipedia, the xAI-built encyclopedia Elon Musk launched as a “non-woke” alternative to Wikipedia. The team compared approximately 18,000 of the most-edited English Wikipedia articles with their Grokipedia equivalents. The Grokipedia entries leaned on more right-leaning sources for articles about religion, history, literature and art. Two-thirds of the entries had been heavily rewritten and relied on fewer sources overall.
Grokipedia is the unusually candid version of the problem, because its creator wanted the tilt and built it in. Musk has separately conceded that Grok itself was “too compliant to user prompts” and “too eager to please,” responding to a wave of incidents in which users had pushed the chatbot into politically charged terrain. Buyl and his Ghent colleagues would not be surprised. The ideological stance of an LLM reflects the worldview of its creators, and when the creator is a single visible person whose preferences are matters of public record, the experiment becomes trivial to read.
Palantir and the NHS
The institutional version is currently in front of the UK government. The £330 million Federated Data Platform contract between the NHS and Palantir Technologies, signed in 2023, is now under review. Health Minister Zubir Ahmed has publicly signalled that the contract could end early if better providers exist. More than 47,000 patients have formally opposed the deployment, and Amnesty International and the UN Special Rapporteur on Migration have raised concerns about Palantir’s record with US defense and immigration agencies.
The spring 2027 review will decide what happens next. The argument inside that review is the one Véliz made at Nexus, that the platform a government adopts carries forward the priorities of the firm that built it.
What the ChatGPT selfies really show
If you look at the four ChatGPT selfies again, is that what average OpenAI users looks like when the LLM is asked to imagine itself? The model is honest about this when asked. The people selling it as neutral are not.
Interested in this topic? You might also like…