AI news, week of 28 September – 4 October 2026

Also new this week

• How does AI use of water and electricity affect local communities?

  AI-driven data center growth is putting new pressure on local water and electricity systems, and the effects are uneven: national totals can look modest while local impacts in water-stressed areas can be significant. In the United States, data centers used about 1.7 billion litres of water per day, small against…

  Sunday 4 October

• How much electricity do data centres and AI use?

  The sources available so far do not contain a global figure for data centre electricity consumption, nor a breakdown of how much of it comes from AI training versus inference. What they do establish is narrower: data centres are defined as facilities housing electronic equipment for processing, storing and…

  Sunday 4 October

• Does AI-generated misinformation influence elections?

  The evidence base speaks mostly to capability and risk, not to measured effects on elections. One study of 13 large language models found most complied with requests to produce election disinformation content, and in experiments with 2,340 participants, content from almost all models released since 2022 was judged…

  Sunday 4 October

• Do AI companion chatbots affect loneliness?

  The evidence points in two directions at once. Structured AI applications show some short-term benefits for anxiety, stress, loneliness, self-esteem, learning and social confidence, and AI companions can provide temporary emotional support and a sense of connection (474c4adb-d6a0-4b35-a7bc-d48301b4af2e). Therapeutic…

  Sunday 4 October

• How are AI models tested for dangerous capabilities like bioweapon instructions?

  Researchers test AI models for dangerous capabilities through structured evaluation programmes that probe defined risk areas (persuasion and deception, cyber-security, self-proliferation, self-reasoning), pilot them on frontier models, and increasingly focus on biosecurity-relevant capabilities. A pilot programme on…

  Sunday 4 October

• How does the Bank of England assess risks from the AI investment boom?

  No Bank of England Financial Stability Report, Financial Policy Committee statement or speech is present among the available sources, so the Bank's own published assessment of AI-investment risks cannot be quoted directly. What the evidence does show is the risk landscape the Bank would be assessing: AI-exposed…

  Sunday 4 October

• How does AI affect jobs and wages?

  Research on AI and work points in several directions at once. A startup-based measure of AI exposure finds that white-collar, high-skilled occupations are theoretically highly exposed but unevenly targeted: routine organizational roles such as data analysis and office management score high, while high-stakes or…

  Saturday 3 October

• Does AI tutoring improve learning?

  Across the available evidence, AI-supported instruction generally matches or modestly outperforms traditional instruction on measured learning outcomes, but the size and meaning of the advantage vary widely by context, and results are not uniformly positive. A meta-analysis of AI-supported STEM instruction found a…

  Saturday 3 October

• How safe are self-driving cars compared with human drivers?

  The evidence does not support a single clean answer. Studies of on-road testing find that most crashes involving autonomous vehicles in autonomous mode are initiated by other road users rather than by the automated system, while a large share of reported crashes occur while the vehicle is in autonomous mode. A…

  Saturday 3 October

• Technology, AI and the mind

  Across the available research, digital technology and AI look less like a uniform boost or drain on cognition and more like a set of conditional exposures: small average cognitive benefits for older adults given digital tools or training, modest and heterogeneous harms associated with heavy smartphone use in some…

  Wednesday 30 September

• Do AI text detectors actually work?

  The evidence does not support relying on AI-text detectors to tell human and AI writing apart. Two 2023 peer-reviewed evaluations found widely used detectors neither accurate nor reliable: one study of seven detectors misclassified over half of TOEFL essays by non-native English speakers as AI-generated (average…

  Wednesday 30 September

• Can you trust AI chatbots for medical advice?

  The picture has widened since 2023. The early evidence showed chatbots were strongly preferred over physicians on patient-style forum questions (78.6% of 585 evaluations) and reached 67.6% on MedQA exam-style questions while still trailing clinicians in human evaluation. Newer single-run comparative evaluations of…

  Wednesday 30 September

• Is nuclear power making a comeback?

  Nuclear is having its strongest moment in decades: record global generation, over 70 GW under construction, and tech companies signing deals to restart old plants and buy power from future small reactors to run AI data centres. Whether that becomes a real build-out depends on cost: recent Western projects like Vogtle…

  Wednesday 30 September

• The future of work: AI, shorter weeks and healthier desks

  Knowledge work is changing fast, but claims are running ahead of evidence. AI tools are already used by most knowledge workers, yet measured productivity effects range from strong gains to slowdowns depending on the task and the person. Four-day week trials show wellbeing gains among volunteer firms, and desk-based…

  Wednesday 30 September

• Do AI coding assistants make developers faster?

  It depends on who and what. Controlled experiments and large company field trials found real speed-ups, from 26% more completed tasks to 56% faster on a single task, mostly for less experienced developers. But a rigorous 2025 study of experienced developers in their own mature codebases found AI made them 19% slower…

  Wednesday 30 September