When a task is delegated to an AI system, something is usually left over. Someone has to check what the machine has written, sustain a capacity that is no longer being exercised, assess models whose safety results have not been made public, or build the infrastructure the system takes for granted.

This week, that remainder appears in four settings rarely considered together. A review of clinical guidance written in Birmingham, an experiment involving participants recruited in the United States, an audit of Chinese AI models and a World Bank report on Africa. Each describes a different kind of work that delegation does not eliminate. Each also invites us to ask who is expected to take it on.

In brief

  • Birmingham, 6 Oct. Samuel Oluwatobi Atiku publishes an analysis of 74 official documents on AI scribes in medical consultations in PLOS Digital Health.
  • United States, April–October. A preprint involving 1,222 participants examines what happens when an AI assistant is withdrawn after ten minutes of use.
  • China, 8 Oct. A SemiAnalysis report finds published safety results for 31 out of 857 releases by nine Chinese AI developers.
  • Africa, 6 Oct. The World Bank devotes its Africa Economic Update to AI readiness and recommends small-scale, affordable applications.

Birmingham — The Drafting Fiction

Samuel Oluwatobi Atiku, of Aston University’s Research Services, published a qualitative analysis of 74 English-language official documents on AI scribes in PLOS Digital Health on 6 October. These systems listen to medical consultations and produce draft clinical notes.[1] The corpus brings together guidance from the NHS, Canadian privacy commissioners, Australian bodies and the Medical Council of New Zealand.

Atiku develops a concept he calls the ‘drafting fiction’. The documents treat AI-generated text as provisional until a healthcare professional reviews and signs it. That convention allows patient reassurance, the integrity of the medical record and professional accountability to rest on a single act.[1] Yet almost none of the documents specifies how much time such a review requires or how that time should be protected.

The author also observes that consent is meaningful only if someone is available to write a clinical note without the system when a patient declines its use. He cites a warning from Ontario’s privacy commissioner that voice recordings are virtually impossible to de-identify.[1]

Automation does not eliminate clinical work. It turns part of that work into the scrutiny of someone else’s text, signed under one’s own name.

The scope of the study matters. Its corpus was purposively selected, restricted to English-language documents and closed on 10 May 2026. It does not measure whether organisations comply with the guidance or whether any particular product is safe.

United States — Ten Minutes

A team from Carnegie Mellon, UC Berkeley, Oxford, MIT and UCLA, led by Grace Liu, published a preprint in April, updated on 3 October, reporting three randomised experiments involving 1,222 participants recruited in the United States.[2] After roughly ten minutes solving fraction problems or reading-comprehension exercises with an assistant, participants who lost access performed worse and abandoned more tasks than those who had never used it.

The study is particularly interesting for the distinction it draws. In one experiment, participants who reported asking for direct answers showed the greatest decline, while those who asked for hints did not. The authors themselves caution that this comparison is not necessarily causal.[2] Even so, it distinguishes two practices we often conflate — receiving help to solve a problem and handing over the problem-solving itself.

The study measures immediate effects under experimental conditions and does not demonstrate any lasting deterioration. But it brings empirical evidence to the concern we raised in Learning Not to Delegate — which capacities do we want to keep exercising ourselves?

China — 31 out of 857

The US technology analysis firm SemiAnalysis reviewed 857 releases by nine Chinese AI developers — Alibaba, ByteDance, Tencent, Baidu, DeepSeek, Moonshot, Zhipu Z.ai, MiniMax and StepFun — between 2021 and 15 September 2026. Only 31, or 3.6%, had ever been accompanied by safety results published by the company. Just nine had such results available at launch.[3]

The report measures public documentation, not actual safety, since companies may have carried out internal testing. Its own standpoint also matters. Titled Beijing Will Not Pace the Frontier, it frames Chinese AI policy as a race for speed and, according to Reuters, provides no comparable figures for US developers.[4] The lack of published assessments it documents is verifiable. So is the absence of a basis for comparison.

This shortage of documentation leaves those deploying the models — from hospitals and public authorities to small laboratories such as ours — having to carry out checks with incomplete information. In July, we wrote that releasing model weights is not enough. Nor is downloading them. Openness provides autonomy, but it cannot replace documentation of risks, data provenance or transparency about limitations.

Africa — The 0.6%

On 6 October, the World Bank presented its Africa Economic Update, subtitled Building AI Readiness.[5] According to the report, Africa accounts for nearly 18% of the world’s population, around 2% of AI training data and just 0.6% of global data-centre capacity — a measure that is not equivalent to the specialised computing power needed to train AI models.

Activity is concentrated in Kenya, Nigeria and South Africa. The report’s central recommendation is to focus on small, affordable, locally adapted applications rather than pursuing frontier models.[6]

The proposal is sensible, but it is also revealing. ‘Small AI’ can be understood either as an intelligent adaptation to existing conditions or as the allocation of a subordinate position within an international division of computing power.

The difference depends on who controls the data, shared data centres and the data-protection laws the report itself discusses. It is the same concern we explored in ‘Africa at Scale’, now with figures from the World Bank.

What Remains

In all four cases, the system takes something for granted. A verified clinical note, a capacity kept in use, an evaluated model, an operational infrastructure. Producing what is taken for granted is work, and it reappears elsewhere, seldom accounted for.

Tomorrow, in Thursday’s article, we follow that thread to a remainder on another scale — what data extraction leaves behind in the communities whose languages are to be digitised.

References

[1] Atiku, Samuel Oluwatobi. Regulating the drafting fiction: A qualitative content analysis of official guidance and regulator documents on AI scribe adoption and use in healthcare. PLOS Digital Health 5(10), e0001776. 6 October 2026. https://doi.org/10.1371/journal.pdig.0001776

[2] Liu, Grace; Christian, Brian; Dumbalska, Tsvetomira; Bakker, Michiel A.; Dubey, Rachit. AI Assistance Reduces Persistence and Hurts Independent Performance. arXiv:2604.04721, v5, 3 October 2026. https://arxiv.org/abs/2604.04721v5

[3] Chen, Mark; O’Laughlin, Doug; Patel, Dylan. Beijing Will Not Pace the Frontier: China’s Speed-First AI Safety Regime. SemiAnalysis, 8 October 2026. https://newsletter.semianalysis.com/p/beijing-will-not-pace-the-frontier

[4] Baptista, Eduardo. China AI developers publish safety tests for just 3.6% of model releases, report finds. Reuters, 9 October 2026. https://www.reuters.com/legal/litigation/china-ai-developers-publish-safety-tests-just-36-model-releases-report-finds-2026-10-09/

[5] World Bank. Africa Economic Update, October 2026: Building AI Readiness. https://www.worldbank.org/en/region/afr/publication/africa-economic-update

[6] World Bank. Economic Growth in Sub-Saharan Africa Gains Momentum Despite Global Uncertainty. Press release, 6 October 2026. https://www.worldbank.org/en/news/press-release/2026/10/06/economic-growth-in-sub-saharan-africa-gains-momentum-despite-global-uncertainty