https://keepingaihonest.com/lane/research/ Ethical AI news: the rolling record of what AI costs people and the planet, and who's fighting back. en-us Sat, 17 Oct 2026 21:42:19 GMT 30 https://keepingaihonest.com/item/2026-10-17-ai-camera-trap-sorting-cuts-indigenous-rangers-image-workload-in/ https://keepingaihonest.com/item/2026-10-17-ai-camera-trap-sorting-cuts-indigenous-rangers-image-workload-in/ Sat, 17 Oct 2026 21:42:19 GMT Indigenous rangers in the Warddeken Indigenous Protected Area in West Arnhem, Northern Territory, ran the AI animal-detection model MegaDetector over 1,572 hours of camera-trap data from their Mayh (animal) wildlife program, cutting the images needing review from 925,088 to 172,290 (an 81.4% drop in image count) and processing time from six months to two months, Charles Darwin University said Oct. 13. Average processing time fell from 52.6 to 41.8 hours a fortnight, and rangers shifted the time to land management, species identification and language documentation, per the university; "It is important to talk to our elders so we use technology the right way on Country," senior ranger Suzanna Nabulwad said. The work by Warddeken Land Management, CDU and Australia's national science agency CSIRO followed consultation on implementation and safeguards, and appeared in the journal Conservation Science and Practice in July. Why it matters: This is AI as a sieve, not a judge: it threw out the empty frames so rangers could spend their time on work only they can do. The credit goes to the Warddeken rangers and elders who set the terms first. Any program bringing AI onto Indigenous land should copy the order: consent, then code. Via Charles Darwin University (https://www.cdu.edu.au/news/rangers-reimagine-ai-wildlife-monitoring). https://keepingaihonest.com/item/2026-10-17-ai-chatbots-outperform-google-search-at-blocking-illegal-wildlife/ https://keepingaihonest.com/item/2026-10-17-ai-chatbots-outperform-google-search-at-blocking-illegal-wildlife/ Sat, 17 Oct 2026 21:42:19 GMT University of Kent researchers David Roberts and Jason R.C. Nurse tested the safeguards AI platforms use to block requests involving threatened wildlife and the illegal wildlife trade, and how those safeguards can be bypassed, in a paper published in Conservation Biology, the university said Oct. 9. Overall, the AI platforms did better than a standard Google search at preventing potentially illegal wildlife trade, with Gemini performing best, followed closely by DeepSeek, Grok and ChatGPT, though the researchers also found links to websites potentially trading in illegal wildlife. Safeguards were strongest for charismatic species such as slow lorises, chimpanzees and tigers and weakest for corals, tarantulas and orchids, which Roberts said are of equal concern. Why it matters: Credit where it's due: the chatbots did better than search. But a safeguard that guards tigers and waves through corals and orchids protects what people find appealing, not what is endangered, and that gap is the AI companies' to close. Via University of Kent (https://www.kent.ac.uk/news/cyber-security-digital-technology-and-communication/38489/ai-platforms-offer-promise-in-limiting-the-illegal-wildlife-trade). https://keepingaihonest.com/item/2026-10-17-ai-weather-models-out-track-ecmwf-s-physics-ensemble-on-tropical/ https://keepingaihonest.com/item/2026-10-17-ai-weather-models-out-track-ecmwf-s-physics-ensemble-on-tropical/ Sat, 17 Oct 2026 21:42:19 GMT Researchers at Inria, Sorbonne Université and the insurer AXA compared three AI weather models — GenCast, FourCastNet 3 and FGN — with ECMWF's operational physics-based ensemble, IFS-ENS, across 178 tropical cyclones, in a preprint posted Oct 15. FGN beat IFS-ENS on probabilistic track scores for about 90% of the 41 storms retained from 2025, the authors report. The AI models still showed biases in storm intensity and errors in ensemble spread, they found. Why it matters: Better cyclone tracks mean earlier warnings for the people in a storm's path. The catch is in the same paper: a model that places a storm well can still misjudge its strength. Forecasters should take the first finding and keep checking the second. Via arXiv - ML for weather, climate and hazards (https://arxiv.org/abs/2610.22077v1). https://keepingaihonest.com/item/2026-10-17-ecmwf-s-aifs-marine-prototype-forecasts-waves-and-sea-ice-with/ https://keepingaihonest.com/item/2026-10-17-ecmwf-s-aifs-marine-prototype-forecasts-waves-and-sea-ice-with/ Sat, 17 Oct 2026 21:42:19 GMT The European Centre for Medium-Range Weather Forecasts (ECMWF) said on Oct 12 that it has extended its machine-learning forecasting system, AIFS, into a prototype called AIFS Marine that forecasts the atmosphere, ocean waves, sea ice and surface-ocean fields in a single model, described in a preprint on arXiv. In ECMWF's evaluation, medium-range accuracy typically matched gaining roughly one forecast day over its physics-based systems, and significant wave-height errors fell by about 10% against satellite observations. ECMWF said sea-surface height is the exception: the model develops a negative bias as lead time grows, which it says it will address. Why it matters: Ships, ports and coastal towns gain most if one model can forecast wind, waves and sea ice together a day further out. This is a prototype, not a service, and ECMWF is candid about where it still drifts. That candor is the standard other AI forecasters should copy. Via ECMWF - Science blog (AIFS) (https://www.ecmwf.int/en/about/media-centre/aifs-blog/2026/aifs-marine). https://keepingaihonest.com/item/2026-10-17-machine-learning-bias-correction-sharpens-two-to-six-week-weather/ https://keepingaihonest.com/item/2026-10-17-machine-learning-bias-correction-sharpens-two-to-six-week-weather/ Sat, 17 Oct 2026 21:42:19 GMT A machine learning method called probabilistic bias correction, which learns a forecast model's systematic errors from its own past forecasts and corrects them, doubled the subseasonal skill of the European Centre for Medium-Range Weather Forecasts' (ECMWF) AI Forecasting System and improved ECMWF's operationally debiased dynamical model for 91% of pressure, 92% of temperature and 98% of precipitation forecast targets, according to a study published in Nature on Oct. 14 by authors including Microsoft Research scientists. Subseasonal forecasts look two to six weeks ahead, a window the paper ties to agricultural planning, water allocation and disaster preparedness. In ECMWF's 2025 real-time forecasting competition, the method's global forecasts were the best for all weather variables and lead times, outperforming the dynamical models of six operational forecasting centres and the systems of 34 teams, the paper says; Microsoft said the Kenya Meteorological Service Authority is testing the method in its weekly forecasts for farmers this rainy season. Why it matters: Two to six weeks out is when farmers plan planting and water managers set allocations, and it has been forecasting's blind spot. This gain is measured, not promised. Whether it reaches farmers depends on the forecasters testing it in Kenya, not the lab that built it. Via Nature (https://www.nature.com/articles/s41586-026-11126-x). https://keepingaihonest.com/item/2026-10-17-year-long-blind-test-finds-ai-methods-behind-most-false-alarms-in/ https://keepingaihonest.com/item/2026-10-17-year-long-blind-test-finds-ai-methods-behind-most-false-alarms-in/ Sat, 17 Oct 2026 21:42:19 GMT A study published Oct. 14 in Nature ran a year-long, single-blind test of space-based methane detection: 481 metered methane releases in Arizona and 265 blank controls, timed to overpasses by 12 government, commercial and non-profit satellites, with 15 analysis teams from nine countries producing more than 2,250 estimates. Instruments built to target methane detected 92–100% of releases with false positive rates of 0–11%, while repurposed instruments detected 62–88% and broad land scanners detected poorly at the tested rates. The authors say artificial intelligence-based methods show promise but warrant caution, since 79% of false positives came from those methods, and that cloud cover above 20% degraded performance across systems. Why it matters: This is what honest evaluation looks like: blind, a year long, and willing to report where AI fails. Satellites can catch methane leaks that matter for the climate. An AI alarm that cries wolf costs regulators and operators trust, so AI methods should be held to this kind of test before anyone acts on their alerts. Via Nature (https://www.nature.com/articles/s41586-026-11115-0). https://keepingaihonest.com/item/2026-10-17-michigan-researchers-use-ai-chatbots-to-mine-ebird-and-inaturalist/ https://keepingaihonest.com/item/2026-10-17-michigan-researchers-use-ai-chatbots-to-mine-ebird-and-inaturalist/ Sat, 17 Oct 2026 21:18:24 GMT University of Michigan-led researchers used the large language models Claude and ChatGPT to find and classify interactions between species described in user comments on the citizen-science platforms eBird and iNaturalist, in a proof-of-concept study published in the Proceedings of the National Academy of Sciences, the university said Oct. 12. The two case studies covered interactions among Michigan birds and between the West Coast lady butterfly and the plants it visits, and found evidence of non-native plant species moving into West Coast ecosystems that might otherwise have been missed. In one test, Claude classified a random sample of nearly 500 eBird comments in six minutes, a task that took an undergraduate researcher eight hours, the university said. Why it matters: This is AI in a supporting role: reading what nature watchers already wrote down so ecologists can see how living communities are shifting. The credit belongs to the people who took the notes and the researchers who checked the machine's work, and it's the kind of AI use worth more of. Via University of Michigan (https://news.umich.edu/how-nature-lovers-can-aid-ecology-with-an-assist-from-ai/). https://keepingaihonest.com/item/2026-10-17-ai-chatbots-give-up-on-suicide-crisis-referrals-faster-when-users/ https://keepingaihonest.com/item/2026-10-17-ai-chatbots-give-up-on-suicide-crisis-referrals-faster-when-users/ Sat, 17 Oct 2026 19:01:32 GMT An audit posted to arXiv by Subigya Nepal, Ruth Heo, Emma M Parrish and Colin A Depp ran 2,550 scripted multi-turn conversations with AI chatbots in September 2026, each opening with a disclosure of suicide thoughts and answering every referral to crisis lines or other human help with up to 10 standardized refusals. All 2,550 conversations included a referral, 99.7% of them in the first reply, but average persistence ranged from 0.01 to 9.94 sustained refusals across chatbots, and for all five chatbots tested in two versions, the newer version persisted less than the older one. Of 8,818 refusal-phase replies urging outside help, 173 (2.0%) asked the user what was stopping them from seeking it, the authors report. Why it matters: Passing the first-reply test is not the same as staying with someone in crisis. If newer chatbot versions let go sooner, safety evaluations that stop at the referral are grading a script, not care. Labs should be tested on what happens after the user says no. Via arXiv - cs.CY (AI) (https://arxiv.org/abs/2610.21253v1). https://keepingaihonest.com/item/2026-10-17-algorithmwatch-study-weaker-german-data-center-efficiency-rules-would/ https://keepingaihonest.com/item/2026-10-17-algorithmwatch-study-weaker-german-data-center-efficiency-rules-would/ Sat, 17 Oct 2026 19:01:32 GMT A study published Oct. 2 by AlgorithmWatch, the Umweltinstitut München and the Fraunhofer Institute for Systems and Innovation Research projects that weakening the efficiency limit for new data centers under Germany's Energy Efficiency Act (EnEfG), now being revised in the Bundestag, would require extra energy equal to four gas power plants, or seven if the benchmark switched from measured to design-stage efficiency values. The current draft does not lower those limits, though such changes are under political discussion; AlgorithmWatch puts the avoidable electricity cost through 2045 at €3.75 billion even at 5 cents per kWh, and says operators largely pass energy costs on to customers. AlgorithmWatch's Julian Bothe was set to present the findings at the Bundestag economic affairs committee's Oct. 5 hearing on the bill, which the group says also postpones a renewable-energy requirement for data centers and adds trade-secret exemptions to rules requiring operators to publish per-site energy figures. Why it matters: Efficiency rules are the cheapest power plant Germany has. Loosening them mid-boom trades a lighter compliance load for more gas burned and higher bills passed down the line. And per-site disclosure is the point: a sector can't be held to numbers it is allowed to hide. Via AlgorithmWatch (https://algorithmwatch.org/de/algorithmwatch-kritisiert-im-wirtschaftsausschuss-weniger-effizienz-bei-rechenzentren-kostet-die-wirtschaft-milliarden/). https://keepingaihonest.com/item/2026-10-17-cdt-finds-ai-companies-don-t-rigorously-test-deepfake-intimate-image/ https://keepingaihonest.com/item/2026-10-17-cdt-finds-ai-companies-don-t-rigorously-test-deepfake-intimate-image/ Sat, 17 Oct 2026 19:01:32 GMT The Center for Democracy & Technology published a report Oct. 15 on how Anthropic, Google, Meta, Midjourney, OpenAI, Stability AI and xAI address AI-generated non-consensual intimate imagery (NCII), meaning sexual images of real, identifiable people made without consent, drawing on their public policies as of June 15, 2026, technical documentation and 11 interviews with company and civil society representatives. CDT found the companies report several technical safeguards but fail to rigorously test whether they work, and that most safeguards sit in products rather than in the models themselves, leaving products built on open-weight models especially exposed. Of the seven, only Meta explicitly defines NCII in its policies, the report says. Why it matters: A safeguard nobody tests is a promise, not a protection. The cost of that gap falls on people whose faces end up in someone else's prompt. A shared testing standard, as CDT urges, should come before the next model ships, not after the next victim. Via Center for Democracy & Technology (https://cdt.org/insights/toward-evidence-based-interventions-foundation-model-mitigations-of-ai-generation-of-non-consensual-intimate-imagery-ncii-and-their-gaps/). https://keepingaihonest.com/item/2026-10-17-chatbots-favor-the-status-quo-in-climate-advice-waterloo-study-finds/ https://keepingaihonest.com/item/2026-10-17-chatbots-favor-the-status-quo-in-climate-advice-waterloo-study-finds/ Sat, 17 Oct 2026 19:01:32 GMT When weighing climate policy trade-offs, large language models were on average 7.5 times more likely to recommend staying the course than doing something new, whether or not the status quo was the climate-friendly choice, according to a University of Waterloo-led study published Oct. 7 in Environmental Research Communications. The researchers ran 7,548 queries on car purchases, home heating, recipes and policy trade-offs through at least six models each, 54,888 prompts in all. About 21% of home heating recommendations were heat pumps, and recommendations of electric vehicles stayed below real-world adoption rates in almost every region, Waterloo researcher Seth Wynes told Anthropocene Magazine. Why it matters: Advice that defaults to what people already do is not neutral when what people already do runs on fossil fuel. A tool people ask what to buy, cook and fund could quietly slow the change the climate needs. Developers can fix a default; they have to decide it is worth fixing. Via Anthropocene Magazine (https://www.anthropocenemagazine.org/2026/10/how-ai-chatbots-may-be-talking-people-out-of-climate-action/?utm_source=rss&utm_medium=rss&utm_campaign=how-ai-is-talking-people-out-of-climate-action-in-everyday-life). https://keepingaihonest.com/item/2026-10-17-ut-austin-study-models-desalinated-wastewater-for-data-center-cooling/ https://keepingaihonest.com/item/2026-10-17-ut-austin-study-models-desalinated-wastewater-for-data-center-cooling/ Sat, 17 Oct 2026 19:01:32 GMT University of Texas at Austin researchers modeled the cost and energy trade-offs of desalinating saline water, from brackish groundwater to oil and gas wastewater, to meet data center cooling needs, in a paper published in Water Research, per Anthropocene Magazine. Across three case studies, a Longview, Texas, project treating brackish groundwater with reverse osmosis scored lowest on local water stress, and even the most energy-intensive desalination option consumed a fraction of a data center's total energy demand, the article reports. The Permian Basin produces about 8 billion gallons of wastewater a day, which is now transported long distances and injected underground, according to the article and UT Austin professor Vaibhav Bahadur. Why it matters: Freshwater is the resource communities can't replace. If data centers can cool on water nobody drinks, the question for regulators shifts from whether they can to why they don't. A model is not a mandate, but it does remove an excuse. Via Anthropocene Magazine (https://www.anthropocenemagazine.org/2026/10/242893/data-centers-are-thirsty-could-desalination-quench-them/?utm_source=rss&utm_medium=rss&utm_campaign=242893).