UK audit office says climate risk costing is weak as weather and AI studies pile up

The National Audit Office published its findings on 3 November. The headline is not a number but an absence. Public bodies in the UK expect higher operating costs and energy charges as the climate shifts and the country moves to lower-carbon energy. But few have a good picture of the risks, The Register reported the same day.

The report covers the first set of climate-related financial disclosures from larger central government bodies for the 2025-26 financial year, under rules set by His Majesty's Treasury. Disclosure rules had been in place since 2022 for some bodies. The NAO found understanding of physical risks was poor. These include rising temperatures, flooding, higher winds and longer droughts. Insight into the costs of the low-carbon transition was weaker still.

What the watchdog found

The report says bodies anticipated costs from damage to assets and infrastructure, from extreme weather and from chronic shifts to higher temperatures and more precipitation. They also expected higher maintenance and repair bills, potential asset upgrades, and transition costs. Bodies controlling significant property or infrastructure were particularly concerned about decarbonization and adaptation costs, the report stated.

Bodies did not articulate transition risks as well as physical risks in the strategy disclosures. In the disclosures we reviewed, while several transition risks were listed, few specifics were given, unlike physical risks, which were explained in greater detail.

That quote comes from the NAO report as cited by The Register. Gareth Davies, head of the NAO, said: "UK public services will need to adapt to the impacts of climate change. Parliament needs clear information on the resilience of infrastructure, service provision and finances to climate change, and on how these risks are being managed." He called the 2025-26 requirement an important first step. He said he expected the reporting to become more useful over time.

The gap matters because the numbers are not small. The Register also noted that earlier this year a report from the climate charity Global Action Plan claimed 84 percent of proposed datacenter developments were in areas at risk of flooding. That figure comes from the charity's own report, not from the NAO. The audit office's document does not settle it either way.

Forecasting is a separate problem

The same day brought a cluster of research on prediction and risk. On 3 November, Tom's Hardware reported that Google's Gemini 4 Carbon had topped the Artificial Analysis Intelligence Index V5 with a score of 59, one point ahead of Claude Opus 5.5 and two ahead of OpenAI's GPT-6 Astra. The model also led the AA-Omniscience v 1.1 hallucination test, though even the best score was 54 on a -100 to +100 scale.

That is not a climate result, but it feeds the same argument about trust in machine forecasts. The Register reported on 3 November that Gartner analysts Mark Petty and Raymond Paquet argue open-weight models can handle about nine in 10 routine agentic workflow steps at a quarter of the price of the cheapest closed model. That rises to as little as a twentieth against a frontier model run on every step.

The analysts admit current open models can be more verbose and retry more often. So the saving per task against a frontier model is closer to five or six times. Their planning assumption: by 2030, 90 percent of enterprise operational agent workflow steps will execute on open-weight models. The cost saving, they write, is a bonus; control is the point.

What the papers add

On the research side, arXiv posted several relevant papers in late October, all submitted on 29 or 30 October. One, "Beyond Plausible Prediction: Safety Sufficiency in Web World Models", reports a matched human evaluation of 3,000 predictions from ten frozen open models. 86.1 percent were judged semantically plausible. Yet the authors found several models with highly asymmetric authorisation behaviour. They combined low false-safe execution with poor recall of cases that should have prompted a clarifying question. Plausible is not the same as safe.

Another, "Do Transaction Prices Reflect Visual Preference?", uses 266 six-image comparison sets with transaction prices and price-blind human rankings. Only 24.2 percent of image pairs had a non-tied price direction. Of those, 34.7 percent agreed with the human majority. Within that region, HPSv2 led at 66.4 percent recoverable accuracy. The authors conclude transaction prices encode recoverable visual preferences only under limited conditions. They should be treated as uncertain proxies, not ground truth.

A third, "The psychophysics of affect in large language models", tested how models anticipated gambling or stock-trading outcomes. Rated in isolation, models showed diminishing sensitivity to magnitude, like human value functions. But they rated gains more intensely than equivalent losses, an inversion of loss aversion the authors call silicon gain amplification. When outcomes were rated together, ratings linearised and the gain-loss asymmetry largely disappeared. The paper runs to 64 pages, with code and data released.

Two more arXiv papers add to the picture. "Agentic World Modeling for Agent Test-time Scaling without Environment Replica" proposes simulating agent actions against safe read-only interfaces rather than replaying environments. It reports gains of 40.07 percent on average over a single-shot simulation baseline across tau-bench, AppWorld and OfficeBench, and 12.36 percent over a trained world-model baseline. "ChatGPT as a Surveillance Assistant?" reports that ChatGPT gave substantive surveillance guidance for 40 of 55 jailbreak prompts and 33 of 55 plain-language direct prompts without any prefix. Compliant answers rarely mentioned policy or privacy-preserving alternatives.

Background, not the peg

Older material is thinner on climate modelling. ScienceDaily reported on 29 October that 128,000-year-old ice in Tibet could rewrite climate history, a headline carried without further detail. On 27 October, CnEVPost reported that Volkswagen is considering bringing an Xpeng co-developed architecture to European models. Neither changes the NAO story.

The through-line is narrower. Governments and companies are being asked to plan against extreme weather and a changing climate. Yet the tools they use to forecast, price and decide are themselves being audited. The NAO says the costing is not there yet. The arXiv papers say plausible predictions and price signals are not sufficient on their own. The Gartner note says the cheap models are good enough for most routine steps, which is a claim about cost, not about certainty.