Hi everyone. Welcome to the first instalment of my newsletter on algorithmic governance. Here, you can find my selection and commentary on some of the latest developments in the interactions between artificial intelligence, digital transparency, and human rights. Enjoy.
Governance
https://www.bbc.co.uk/news/articles/c0m2wjlkzplo
In the first prosecution of Chinese spies in British history, two men of dual British and Hong Kong nationality were found guilty on Thursday under the National Security Act of assisting a foreign intelligence service. One of them, Chi Leung "Peter" Wai, was working for the UK Border Force and used his access to government immigration databases to track and even attempt to arrest Hong Kong dissidents, feeding this information to Hong Kong authorities.
Wai’s misuse of such databases since 2013 represents not only gross negligence by the Home Office and Metropolitan Police, but the culmination of the gradual erosion of restrictions on the government’s handling of sensitive data. In its pursuit of economic growth and digital IDs, the government has introduced successive legislation — most notably the recent Data Protection (Use and Access) Act 2025 — that relaxes critical compliance mandates and introduced new “legitimate interests” under which sensitive data can be accessed while circumventing the checks-and-balances mandated by the GDPR.
This weakening exposes citizens’ and political refugees’ data not only to misuse by the domestic government, but also opens the door wide to the transnational repression exemplified in this story. Securing this data while enabling necessary access to it is possible, but only with sufficient and lasting political will.
Business and markets
Perspective: AI demand is inflated, and only Anthropic is being realistic
https://www.cnbc.com/2026/04/17/ai-tokens-anthropic-openai-nvidia.html
Tokens, the words and characters that constitute user inputs and model outputs, are the basic unit of AI usage and are used to model demand. Chatting with a LLM consumes a few hundred tokens per paragraph. But the increasingly popular agentic AIs that write code, trawl the web, and execute multi-step workflows, consume 50-100x more tokens. Given this disparity, extracting the actual demand underlying record levels of token consumption is critical for AI companies to judge infrastructural build-out.
Anthropic’s latest pricing model directly responds to this increased token consumption. While it hasn’t changed the price it charges per token, the latest Opus 4.7 model increases the amount of tokens used for processing input categories like code and structured datasets. It also increases the amount used for generating outputs, and output tokens cost 5x more than input! Staff have stated that the new model can use up to 35% more tokens than before. Anthropic has also cut off third-party agentic tools that were large token consumers, alongside tools like OpenClaw through which users would route their commands to bypass their plans’ limits.
By increasing token count and removing these third-party tools, Anthropic can simultaneously increase its token revenue and better estimate demand. As OpenAI pursues its aggressive scaling strategy at a cost of $1.4tn, Anthropic has in contrast invested just $50bn in infrastructure. But it’s projected to break even by 2028, while OpenAI projects losses into the 2030’s. Correctly estimating and monetising token consumption as a proxy of demand is critical for profitability. As Anthropic CEO Dario Amodei said in a recent podcast, “If you’re off by a couple years, that can be ruinous”.
Regulation
What the EU AI Omnibus Deal Changes for the AI Act and What Lies Ahead
https://www.techpolicy.press/what-the-eu-ai-omnibus-deal-changes-for-the-ai-act-and-what-lies-ahead/
One of the most contentious and far-reaching components of the Omnibus is Article 4(a). It will operate as lex specialis alongside existing GDPR to allow the use of sensitive data for bias detection and correction in narrowly defined cases for high-risk AI systems. These systems can be private or public sector and include those use for biometrics and identification, employment and worker management, and migration and border control. Article 4(a)’s application is contingent upon strict assessments of necessity and proportionality, security measures, access controls, and deletion requirements, and doesn’t mandate such bias detection and correction.
I welcome this change. From facial recognition CCTV systems to welfare assessments, there’s extensive evidence that algorithms do bias against ethnicity, health data, and religious beliefs. Processing such sensitive data is necessary to assess for bias, but is prohibited under Article 9 of GDPR. Article 4(a) finally creates a legal route for governments and researchers to do so.
Opponents to Article 4(a) warn of privacy breaches, mission creep, and data leaks, but sensitive data can be anonymised and processed by governmental and intergovernmental bodies like the EU AI Office in narrowly defined contexts and with tightly regulated access/retention durations. I think the biggest risk is data leaks, as swathes of recent examples like the Biobank leaks attest. But there are still various technical methods of storing and processing this data in maximally secure digital environments.
Algorithmic bias is rife and algorithmic governance is necessary. Only governmental bodies can possess the legal mandates and accountability to do it.
Research
LLMorphism: When humans come to see themselves as language models
https://arxiv.org/abs/2605.05419
Dr. Capraro presents a lucid exploration into how the characteristics of LLMs and the vocabulary used to describe them may be projected onto humans. I think some of the most compelling cases are in psychiatry.
LLMs are already being rapidly incorporated into psychiatry, from extracting underlying themes from unstructured data like assessment transcripts to monitoring patients’ progress through analysing their interactions with chatbots. All instances of such use involve making people’s self-descriptions machine-readable by distilling them into statistical distributions of classifiable variables etc. etc. etc. It might seem that the question is how to ensure that psychiatrists still see their patients as more than such data while working with it.
But it isn’t the individual psychiatrist’s perspective that should be considered. Once you place LLM use in psychiatry within an administrative context and recognise the accompanying pressures towards cost-efficiency, scaling, and interoperability, it seems unavoidable that patients must be represented in increasingly abstract and formalised variables to enable them.
Now, the question shifts from what a psychiatrist thinks to how psychiatrists collectively function. In a future digitised healthcare system that so strongly tends towards this interoperability, what room is left for them to even conduct psychiatry in ways whose purposes aren’t to make patients machine-readable? And what of the tragedy of those psychiatrists who do see their patients as more than this, but who can’t act on it?