Harm and benefit, counted side by side and never summed.
The balance · a SIREN view · does not feed the main number
How does SIREN count AI harm and benefit side by side?
Newsroom window
· registers checked by hand as of 2026-09-28
· 6 of 6 counters live
The utopia case, as far as it has receipts: what AI has measurably done for people,
beside what it has measurably done to them. Three layers, each on its own terms — a count of two word lists
in the same headlines, six counters read from public registries, and two registers of dated, sourced
results — and one refusal. Nothing here is added up, netted or divided into a verdict, because the two
sides share no unit.
“The world is awful. The world is much better. The world can be much better.”
Roser’s three sentences are three readings of one measure, child mortality, and none of
them is subtracted from another; his point is that they hold at the same time. This page borrows the posture
and not the arithmetic. Its two sides have no common measure, so they are printed apart, and the drawing below
is allowed to weigh exactly one like-for-like pair of counts.
What did the newsroom’s headlines carry?The beam
24 of 500 stories in the newsroom's current window matched the harm list and 5 matched the benefit list, as of 05:24 UTC on Thu 15 Oct 2026. A count of words, not of outcomes.
The beam tilts on one thing: stories that matched one list and not the other, in this window. Tilt = 8° × |H − B| ÷ (H + B), heavier pan down, never past 8°; level when the two are within 2 of each other or fewer than 5 matched. The difference is drawn, never printed as a number.
Every scanned headline lands in exactly one row. A story is a
cluster of headlines about one event, so syndicated coverage cannot fill a pan three times; the beam
tilts on the first two rows of the stories column and on nothing else.
Matched
Stories
Headlines
On the drawing
Harm list only
24
36
the right pan
Benefit list only
5
5
the left pan
Both lists
0
0
at the fulcrum; moves neither pan
Neither list
471
568
under the stand, every time
Scanned
500
609
every headline of a scanned kind
The window. 800 items, published from
2026-10-12 11:22:24 UTC to 2026-10-15 05:15:00 UTC, a span of
65.9 hours. The file caps the window at
800 items, nominally 7 days, so its
real span moves with how busy the news is, and any series built from these counts has to use shares, not raw counts.
Not scanned, by kind: lab 38, paper 139, release 14, 191 in all, because vendor blogs never report their own harms and arXiv abstracts discuss attacks as subject matter.
62 scanned headlines tripped the hedge rule.
What does the beam never tilt on?
the counters
the registers in data/ledger.json
SIREN against BLISS
severity tiers, item scores, recency, corroboration or engagement
counts from different windows or source sets
any language-model or sentiment judgement
What is printed. the two counts, large; never their difference or their ratio as a number.
Which headlines matched a word list?The matches
Newest first. The marked words are the ones the rule counted; a word with “no”, “not” or
“never” just before it was ignored and is not marked. Each headline links to its scored item page on this
site where one exists, and to the original otherwise. A headline is reproduced exactly as published.
Which headlines matched Benefit language?5 stories
These carried a word from one of the lists and still counted on neither side. They are listed
so the rule can be checked in the direction that does not flatter it.
one distinct summary hit; the rule needs 2 · harm words found: attacked
What documented results sit on both sides?The registers
Two registers kept side by side: what AI has measurably delivered or demonstrated, and what harm it has been documented, measured or alleged to cause. Each row separates what was measured from what is claimed about it, with the sources for both.
Checked by hand against every linked source as of 2026-09-28,
in the register’s own order. Each row prints what was measured beside what is claimed about it; recording a
claim is not endorsing it. The number of rows on each side is a choice made by the compilers, not a measurement. Equal lengths are not a finding that the two sides are equal, and unequal lengths would not be a finding either.
delivered4the output exists at scale and a stranger can check it: a database, a dataset, an authorisation list, a price
in-use4deployed and working for people outside the organisation that built it, by the operator's own report
in-trial2in human clinical trials; no approval, and no outcome beyond what the trials have reported
demonstrated8shown to work in a study, a trial or a test, and not documented as in routine use
claimed0asserted by the organisation behind it with no measurement a stranger can check; listed so the claim is visible, never counted as a benefit
Rentosertib: an AI-found target and AI-designed molecule passed a randomised Phase 2a and is now in Phase 3
in-trialin human clinical trials; no approval, and no outcome beyond what the trials have reported
drug discovery · confidence verified · as of 2026-09-28
71patients randomised in Phase 2a (18/18/18 on drug, 17 on placebo)2025-06-03 · source
+98.4 vs -20.3mL mean change in FVC at 12 weeks, 60 mg QD vs placebo2025-06-03 · source
320patients planned in the 52-week Phase 3 (estimated enrolment, recruiting)2026-09-09 · source
What is claimed
That this is the first drug whose target and molecule were both found by generative AI, and that it can improve lung function in IPF. The efficacy signal came from a secondary endpoint in a small 12-week trial, and the confidence intervals of the 60 mg and placebo arms overlap.
What was measured, in full, and the caveat
What was measured
A double-blind, placebo-controlled Phase 2a in idiopathic pulmonary fibrosis (NCT05938920) randomised 71 patients to three doses or placebo for 12 weeks (Nature Medicine, online 3 June 2025). The primary endpoint was safety. The share of patients with at least one treatment-emergent adverse event was similar across arms: 72.2 to 83.3% on drug and 70.6% on placebo. At 60 mg once daily, forced vital capacity (a secondary endpoint) changed by a mean of +98.4 mL (95% CI 10.9 to 185.9), against -20.3 mL (95% CI -116.1 to 75.6) on placebo. Insilico announced a 320-patient, 52-week, placebo-controlled Phase 3 (NCT07687459) on 7 July 2026. The registry gives an actual start date of 9 September 2026 and lists 47 sites, all in China; the trial is recruiting.
Caveat
The efficacy result is small, short and exploratory, with wide confidence intervals. The most common adverse events leading to discontinuation were liver toxicity and diarrhoea. Insilico sponsored the Phase 2a and employs 11 of its authors. The Phase 3 has produced no results.
AI-discovered molecules passed Phase I unusually often but Phase II at the usual rate; none is approved yet
in-trialin human clinical trials; no approval, and no outcome beyond what the trials have reported
drug discovery · confidence probable · as of 2026-09-28
80-90% Phase I success rate for AI-discovered molecules2024-04-30 · source
~40% Phase II success rate (limited sample, comparable to historic industry average)2024-04-30 · source
75AI-derived molecules entered the clinic since 2015 (secondary report of the paper)2024-06-24 · source
67AI-derived molecules in clinical trials in 2023 (secondary report of the paper)undated · source
Q1 2027FDA target action date for the zasocitinib NDA (Priority Review accepted)2026-09-14 · source
What is claimed
That AI makes drug discovery faster and more successful overall. What has been measured is lower Phase I attrition; no advantage in Phase II efficacy has been shown. Secondary reviews found that the FDA had approved no AI-discovered drug as of July 2026, and no later approval was found.
What was measured, in full, and the caveat
What was measured
The first published analysis of the pipelines of AI-native biotech companies appeared in Drug Discovery Today (online 30 April 2024, by BCG authors). It found that AI-discovered molecules succeeded in Phase I 80 to 90% of the time, well above historic industry averages. In Phase II the rate was about 40%, on a limited sample, which is comparable to historic averages. Secondary reports of the paper's body give 75 AI-derived molecules entering the clinic since 2015, and 67 in clinical trials in 2023. Among computationally designed molecules, zasocitinib (TAK-279) is at the application stage. It is a TYK2 inhibitor that Nimbus Therapeutics designed with Schrödinger's physics-based FEP+ simulations. On 14 September 2026 Takeda announced that the FDA had accepted its NDA for plaque psoriasis under Priority Review, with a target action date in the first quarter of 2027.
Caveat
The percentages come from the PubMed abstract. The paper is open access (CC BY-NC-ND), but ScienceDirect blocked automated reads, so the 75 and 67 counts rest on secondary reports. The authors define which molecules count as 'AI-discovered', and the set is not audited. Zasocitinib came from physics-based free-energy simulation against a target already validated by an approved drug (deucravacitinib), not from generative AI, and Takeda's releases do not call it AI-discovered. The claim that no AI drug is approved depends on how 'AI-discovered' is defined.
FDA lists 1,614 authorised AI-enabled medical devices
deliveredthe output exists at scale and a stranger can check it: a database, a dataset, an authorisation list, a price
medical devices · confidence verified · as of 2026-09-28
1614AI-enabled devices on the FDA list2026-09-04 · source
76% of listed devices in the radiology panel (1,230)2026-09-04 · source
335devices authorised in calendar 2025 (highest year)2026-09-04 · source
What is claimed
That AI is transforming clinical care. The list shows marketing authorisation, not measured patient benefit. Most clearances rest on substantial equivalence to an existing device, not on outcome trials.
What was measured, in full, and the caveat
What was measured
The FDA's AI-Enabled Medical Device List (content current as of 4 September 2026) has 1,614 authorised devices, the most recent decided on 29 June 2026. Radiology accounts for 1,230 (76%) and cardiovascular for 154. The highest yearly total was 335 in 2025, and 181 more are listed for January to June 2026. Of the 1,614, 1,553 were cleared via 510(k), 40 via De Novo and 21 via PMA, counted from the table on the FDA page today.
Caveat
The FDA says the list is not comprehensive. It is built mainly from AI-related terms in summaries, and devices whose decision summaries were not yet published are added in a later update, so the 2026 count is incomplete.
1.55 vs 1.76interval cancers per 1,000 women, AI vs standard (ratio 0.88, 95% CI 0.65-1.18)2026-01-29 · source
80.5 vs 73.8% sensitivity, AI vs standard2026-01-29 · source
44.2% reduction in screen-reading workload (61,248 vs 109,692 readings)2025-02-03 · source
What is claimed
That AI makes breast screening both better and cheaper everywhere. The 12% lower interval-cancer rate was not statistically significant, and the trial was designed to show non-inferiority. It tested one AI system in one national programme and has not reported mortality.
What was measured, in full, and the caveat
What was measured
MASAI randomised 105,934 women in Sweden between 12 April 2021 and 7 December 2022. One arm had AI-supported screening, in which AI sent each exam to single or double reading and supported detection; the other had standard double reading (The Lancet, 29 January 2026; 105,915 analysed). Interval cancers were 1.55 per 1,000 in the AI arm (82 of 53,043) and 1.76 per 1,000 in the control arm (93 of 52,872). The proportion ratio was 0.88 (95% CI 0.65 to 1.18), which met the 20% non-inferiority margin. Sensitivity was 80.5% vs 73.8%, and specificity was 98.5% in both arms. An earlier report from the same trial found 29% more cancers detected (6.4 vs 5.0 per 1,000) and 44.2% fewer screen readings, with no significant rise in false positives.
Caveat
The Lancet page returned 403, so the figures were checked against the PubMed abstracts and The Lancet's press release. The press release rounds sensitivity to 81% vs 74% and gives false-positive rates of 1.5% vs 1.4%. Results from one Swedish double-reading programme may not transfer to single-reader systems such as the US.
AlphaFold: more than 260 million open structure predictions and a Nobel Prize
deliveredthe output exists at scale and a stranger can check it: a database, a dataset, an authorisation list, a price
protein science · confidence verified · as of 2026-09-28
>260 millionprotein and complex structure predictions in AlphaFold DB2026-09-24 · source
>2 millionAlphaFold2 users in 190 countries2024-10-09 · source
What is claimed
That protein folding is 'solved'. These are predictions with per-residue confidence scores, not experimental structures. Accuracy varies, especially in disordered regions and complexes, and a prediction does not by itself produce a drug.
What was measured, in full, and the caveat
What was measured
EMBL reported on 24 September 2026 that the AlphaFold Protein Structure Database holds more than 260 million protein and protein-complex predictions, after adding complexes for more than 2,800 viruses. Half of the 2024 Nobel Prize in Chemistry went jointly to Demis Hassabis and John Jumper 'for protein structure prediction', and the other half to David Baker 'for computational protein design'. The Royal Swedish Academy of Sciences' press release says AlphaFold2 has been used by more than two million people from 190 countries.
Caveat
The EBI site gives two figures. Its headline says 'over 260 million' predictions, while the release note says the latest release has 'over 200 million entries' covering UniProt.
'An order-of-magnitude expansion in stable materials known to humanity.' Cheetham and Seshadri (Chemistry of Materials, April 2024) examined the first 250 entries in the GNoME database. They found 'scant evidence' of compounds that are novel, credible and useful all at once, and many entries relying on metal-ion orderings unlikely to occur in real crystals. They argue these are crystalline inorganic compounds, not 'materials'.
What was measured, in full, and the caveat
What was measured
DeepMind's graph network found 2.2 million crystal structures below the previous convex hull (Nature, 29 November 2023). Because these compete for stability, 381,000 new entries remain on the updated hull, for 421,000 stable crystals in total. The paper reports that 736 of the stable structures 'have already been independently experimentally realized' by other laboratories.
Caveat
Other groups made the 736 independently, not because of GNoME, so they validate the predictions rather than show AI-driven synthesis. DFT stability does not guarantee that a compound can be made or is useful. The companion A-Lab synthesis paper received an Author Correction on 19 January 2026 (DOI 10.1038/s41586-025-09992-y), and its title changed from 'novel materials' to 'inorganic materials'.
AI weather models are operational at ECMWF and slightly beat NHC's official short-range Atlantic hurricane track forecast in 2025
in-usedeployed and working for people outside the organisation that built it, by the operator's own report
weather and climate · confidence verified · as of 2026-09-28
2025-02-25AIFS Single operational at ECMWF2025-02-25 · source
~1000x less energy per forecast than physics-based (ECMWF figure)2025-02-25 · source
66.5 vs 81.4n mi Atlantic 72-h track error, GDMI vs NHC official (homogeneous sample)2026-03-30 · source
97.2% of 1,320 targets where GenCast beat ECMWF ENS2024-12-04 · source
What is claimed
That AI has replaced physics-based forecasting. ML models are trained on physics reanalysis and start from initial conditions produced by physics-based data assimilation. AIFS ENS runs at 31 km, against 9 km for the physics ensemble.
What was measured, in full, and the caveat
What was measured
ECMWF made its AIFS Single model operational on 25 February 2025 and the 51-member AIFS ENS on 1 July 2025, running alongside the physics-based IFS. ECMWF reports gains of up to 20% on some measures and roughly 1,000 times less energy per forecast. NHC's 2025 verification report (30 March 2026) covers the first season in which NHC used AI models in real-time operations. In it, the Google DeepMind ensemble mean (GDMI) slightly outperformed the official Atlantic track forecast at 12 to 72 h; at 72 h its error was 66.5 n mi against 81.4 for the official forecast. GenCast (Nature, December 2024) had greater skill than ECMWF's ensemble on 97.2% of 1,320 targets.
Caveat
GDMI was unavailable early in the season, so the Atlantic comparison leaves out the first two storms. At 120 h the official forecast beat GDMI (168.5 vs 173.1 n mi). For intensity, the official forecast beat all models at 12 and 96 h and matched GDMI at other lead times. In rapid-intensification cases, official forecasts had lower errors than the models at nearly every lead time.
That Flood Hub protects 2 billion people. That figure is the population of covered at-risk areas, not the number of people warned or harms prevented. No count of lives saved or losses avoided has been published.
What was measured, in full, and the caveat
What was measured
In a peer-reviewed study (Nature, 20 March 2024), Google's AI forecasts of extreme river floods in ungauged watersheds were as reliable at up to five days' lead time as the same-day nowcasts of the Copernicus GloFAS system, or more reliable. At publication the system produced free public forecasts in more than 80 countries. Google's Flood Forecasting page now says the forecasts cover 2 billion people in over 150 countries.
Caveat
The skill result is peer-reviewed. The coverage figures are Google's own and carry no date on the page.
Reinforcement learning controlled real tokamak plasmas and steered around tearing instabilities
demonstratedshown to work in a study, a trial or a test, and not documented as in routine use
energy · confidence verified · as of 2026-09-28
2plasmas held simultaneously in TCV by the RL controller2022-02-16 · source
What is claimed
That AI is bringing fusion power closer. These results are control demonstrations on research machines. Neither produced net energy, and neither is in routine use on a reactor-class device.
What was measured, in full, and the caveat
What was measured
On the TCV tokamak (Nature, 16 February 2022), a reinforcement-learning controller built by DeepMind and EPFL learned to command the full set of magnetic control coils. It produced and held conventional, negative-triangularity and 'snowflake' shapes, and kept two separate plasmas ('droplets') in the vessel at once. On DIII-D, the largest magnetic-fusion facility in the US (Nature, 21 February 2024), an RL controller kept the predicted likelihood of tearing instability, the leading cause of disruptions, below a set threshold while maintaining H-mode performance.
Caveat
The results are qualitative control demonstrations, and no energy-yield figure applies. Applicability to ITER is the authors' stated aim, not a measured result.
ML control cut data-centre cooling energy by around 30% in autonomous operation (Google, self-reported, 2016-2018)
demonstratedshown to work in a study, a trial or a test, and not documented as in routine use
energy · confidence probable · as of 2026-09-28
up to 40% reduction in cooling energy (test days, one site)2016-07-20 · source
15% reduction in overall PUE overhead2016-07-20 · source
~30% average energy saving (kW/ton) after nine months of autonomous control2018-08-17 · source
What is claimed
The headline 'cut Google's data-centre cooling bill by 40%'. The 40% is an 'up to' figure for cooling energy on test days at one site, not for the whole facility.
What was measured, in full, and the caveat
What was measured
In 2016 DeepMind reported that its model, deployed on a live Google data centre, cut the energy used for cooling by up to 40% on test days, equal to a 15% reduction in overall PUE overhead. In August 2018 Google reported that the system was running cooling autonomously in multiple data centres. It said savings in kW per ton of cooling rose from 12% at launch to about 30% on average after nine months.
Caveat
The figures come from company blogs and have not been independently audited. Google has not published fleet-wide savings, and no primary source after 2018 confirms the system is still in use, so the status is 'demonstrated' rather than 'in-use'.
42/42points claimed as officially graded for Huawei Celia and Xiaohongshu dots-note-3.0 at IMO 20262026-07-23 · source
7/666human contestants with perfect scores at IMO 20262026-07 · source
What is claimed
That AI now does research-level mathematics. Olympiad problems are timed puzzles with known answers.
What was measured, in full, and the caveat
What was measured
Google DeepMind's Gemini Deep Think scored 35 of 42, solving five of six problems in natural language within the 4.5-hour limit. IMO coordinators graded and certified the score using the criteria applied to students (21 July 2025). That score equalled the gold cutoff, which 72 of 630 human contestants (11.4%) reached. OpenAI also reported 35 of 42, but three former medallists graded it, not the IMO. At IMO 2026 in Shanghai, Huawei and Xiaohongshu said their systems scored 42 of 42 under official grading; 7 of 666 human contestants also scored 42.
Caveat
IMO President Gregor Dolinar has said the IMO 'cannot validate the methods, including the amount of compute used or whether there was any human involvement'. The IMO 2026 AI scores are company claims reported by AFP. Neither IMO 2026 page on imo-official.org mentions AI. DeepMind's statement that 'about 8%' of contestants win gold does not match IMO 2025, where the share was 11.4%.
0.7% of Google's worldwide compute continuously recovered (self-reported)2025-05-14 · source
593outer spheres, new lower bound for kissing number in 11 dimensions2025-05-14 · source
What is claimed
That AI is making new scientific discoveries. The mathematical constructions can be checked independently. The compute and training-time savings are Google's own internal measurements.
What was measured, in full, and the caveat
What was measured
DeepMind reported on 14 May 2025 that its Gemini-powered search agent found a way to multiply 4x4 complex-valued matrices with 48 scalar multiplications. Applying Strassen's 1969 algorithm recursively needs 49. On more than 50 open mathematical problems it rediscovered the state of the art about 75% of the time and improved on the best known solution about 20% of the time. The improvements include a configuration of 593 outer spheres that sets a new lower bound for the kissing number in 11 dimensions. DeepMind also says a scheduling heuristic it found has recovered on average 0.7% of Google's worldwide compute for over a year.
Caveat
The source is a company blog and white paper. The 48 is over the complex numbers; 47 was already known in characteristic 2. Other researchers have since derived a 48-multiplication algorithm with rational coefficients from it (arXiv 2506.13242), which independently supports the result. The infrastructure savings cannot be checked externally.
Co-Scientist hypotheses confirmed in the lab: leukaemia drug candidates and liver-fibrosis drugs
demonstratedshown to work in a study, a trial or a test, and not documented as in routine use
science · confidence verified · as of 2026-09-28
91% reduction in TGF-beta-induced chromatin changes by vorinostat in liver organoids2025-09-14 · source
25drugs tested in human hepatic organoids (peer-reviewed version; the preprint tested 14)2025-09-14 · source
What is claimed
That AI is now an independent scientist. Every validation so far is in cell lines or organoids. The bacterial result reproduced something the authors already knew, in a test run by the same team and Google that was not blinded.
What was measured, in full, and the caveat
What was measured
Google's Co-Scientist is a multi-agent system built on Gemini (Nature, online 19 May 2026). It proposed drug-repurposing candidates and synergistic combinations for acute myeloid leukaemia that were validated in vitro. In a peer-reviewed study (Advanced Science, online 14 September 2025), 25 drugs were tested in human hepatic organoids. Two epigenetic drugs that Co-Scientist recommended showed anti-fibrotic activity, and one of them, vorinostat, reduced TGF-beta-induced chromatin changes by 91%. Separately, Imperial College researchers reported in Cell (online 9 September 2025; 13 November 2025 issue) that its top-ranked hypothesis matched a mechanism of bacterial gene transfer. The mechanism, cf-PICIs hijacking diverse phage tails, was one they had confirmed experimentally but not yet published.
Caveat
No patient data exists. The organoid study has Alphabet co-authors, and the Nature paper was funded by Alphabet.
Over 1.1 million blind and low-vision people use Be My Eyes; NLLB-200 translates 200 languages
in-usedeployed and working for people outside the organisation that built it, by the operator's own report
accessibility and language · confidence probable · as of 2026-09-28
1108235registered blind and low-vision users (live counter)2026-09-28 · source
43 millionAI and video user requests per year (combined, no split published)2025-01-15 · source
1 millionBe My AI uses within two weeks of iOS open beta2023-10-18 · source
200languages in one translation model (NLLB-200)2024-06-05 · source
What is claimed
That AI makes the world accessible. Be My Eyes' numbers are registrations, and they include human volunteer calls as well as AI. NLLB's gain is a benchmark score, not measured real-world use.
What was measured, in full, and the caveat
What was measured
On 28 September 2026 Be My Eyes' live counter read 1,108,235 blind and low-vision users and 11,172,203 volunteers. In January 2025 it reported serving more than 43 million AI and video requests a year. Its GPT-4 image-description feature, Be My AI, reached one million uses within two weeks of its iOS open beta. Be My Eyes reported on 18 October 2023 that it was answering more than 30,000 questions a day. Meta's No Language Left Behind (Nature, 5 June 2024) is a single model covering 200 languages, evaluated over more than 40,000 translation directions and released free for non-commercial use. It averaged a 44% improvement in BLEU over the previous state of the art.
Caveat
The Be My Eyes figures are self-reported and never separate AI from human assistance. NLLB's figure is BLEU benchmark quality on FLORES-200.
Deep learning on satellite imagery showed 72-76% of industrial fishing vessels are not publicly tracked
deliveredthe output exists at scale and a stranger can check it: a database, a dataset, an authorisation list, a price
conservation · confidence verified · as of 2026-09-28
72-76% of industrial fishing vessels not publicly tracked2024-01-03 · source
21-30% of transport and energy vessel activity missing from public tracking2024-01-03 · source
What is claimed
That AI stops illegal fishing. What was measured is visibility. No study has measured a reduction in illegal catch that can be attributed to the tool.
What was measured, in full, and the caveat
What was measured
Global Fishing Watch and partners combined satellite imagery, vessel GPS data and deep-learning models to map industrial vessel activity and offshore infrastructure in the world's coastal waters from 2017 to 2021 (Nature, 3 January 2024). They found that 72 to 76% of the world's industrial fishing vessels are not publicly tracked, and that 21 to 30% of transport and energy vessel activity is missing from public tracking systems. Global Fishing Watch publishes the data free for anyone to use.
Caveat
This is a measurement and transparency result, not a conservation outcome. An author correction was published in February 2024, and the current abstract still gives these figures.
Camera-and-ML targeted spraying covered 5 million acres in 2025 and roughly halved herbicide use (Deere, self-reported)
in-usedeployed and working for people outside the organisation that built it, by the operator's own report
agriculture · confidence probable · as of 2026-09-28
>5 millionacres sprayed with See & Spray in 20252025-11-04 · source
~50% average reduction in non-residual herbicide use vs broadcast spraying2025-11-04 · source
~31 milliongallons of herbicide mix saved2025-11-04 · source
What is claimed
That AI makes farming cleaner and more profitable. The figures come from the manufacturer, are measured against broadcast spraying, and have not been independently audited. Deere charges a per-acre fee ($1 per fallow acre, $5 per in-crop acre) only when the technology delivers measurable savings.
What was measured, in full, and the caveat
What was measured
John Deere reports that farmers used See & Spray on more than 5 million acres in the 2025 season. The system uses cameras on the boom and machine learning to spray only weeds. Deere says it cut non-residual herbicide use by an average of nearly 50% compared with broadcast spraying, saving nearly 31 million gallons of herbicide mix (press release, 4 November 2025).
Caveat
deere.com blocked automated reads. The figures were taken from the press release as reprinted by Oklahoma Farm Report and High Plains Journal, which are copies of one company source, not independent confirmation.
Randomised trials measured learning gains from AI tutors: 0.31 SD in Nigeria, 0.63-1.3 SD at Harvard
demonstratedshown to work in a study, a trial or a test, and not documented as in routine use
education · confidence verified · as of 2026-09-28
0.31SD improvement overall, 6-week AI tutoring RCT, Nigeria2025 · source
0.23SD improvement in English specifically, Nigeria2025 · source
0.73-1.3SD effect size from ceiling-corrected quantile regression, AI tutor vs in-class active learning (N=194; raw 0.63)2025-06-03 · source
What is claimed
That AI tutors will transform education. Both trials were short and at a single site, and both tutors were carefully designed for teaching.
What was measured, in full, and the caveat
What was measured
In a World Bank randomised trial (Policy Research Working Paper 11125, 2025), first-year senior secondary students in Nigeria used GPT-4-based Microsoft Copilot as an after-school English tutor for six weeks in June and July 2024. Scores rose 0.31 SD overall and 0.23 SD in English, which the authors equate to 1.5 to 2 years of business-as-usual schooling. A randomised crossover trial with 194 Harvard physics students (Scientific Reports, 3 June 2025) compared a custom AI tutor with an active-learning class. Median learning gains with the tutor were more than double. The raw effect size was 0.63 SD, and a quantile regression that corrects for a ceiling effect gives 0.73 to 1.3 SD.
Caveat
The Nigeria result is a working paper, not peer-reviewed. A PNAS field experiment (June 2025) with nearly 1,000 high-school maths students found the opposite pattern for an unguarded GPT-4 tutor. It raised practice grades by 48%, but grades were 17% lower once access was removed, and a tutor with guardrails largely avoided that harm. How the tool is designed decides the sign of the effect.
The price of a fixed level of AI performance has fallen about 13x a year since 2023
deliveredthe output exists at scale and a stranger can check it: a database, a dataset, an authorisation list, a price
cost of intelligence · confidence verified · as of 2026-09-28
13x per year decline in cost at fixed performance (47% per quarter), 2023-20262026-09-22 · source
0.30 to 0.0004USD per question for 75% on GPQA Diamond, early 2025 to mid-20262026-09-22 · source
60 to 0.07USD per million tokens at GPT-3-level MMLU, Nov 2021 to Oct 20242025-03-12 · source
9-900x per year price decline range across performance thresholds2025-03-12 · source
What is claimed
That intelligence is becoming 'too cheap to meter'. These are falling list prices at fixed capability for the cheapest qualifying model. They are not total spending or energy use, and not the price of frontier capability.
What was measured, in full, and the caveat
What was measured
Epoch AI (22 September 2026) finds that across five benchmarks the cost of reaching a given level of performance has fallen about 47% per quarter since 2023, or 13x per year. The drop was slower on game puzzles (7 to 10x per year) and faster on maths (16 to 19x per year). Its example is 75% on GPQA Diamond: about $0.30 per question with OpenAI's o3 in early 2025, against $0.0004 with GPT-5.6 Luna just under 18 months later, a 725-fold drop. In March 2025 Epoch measured declines of 9x to 900x per year across six benchmarks. On MMLU, the cheapest price for GPT-3-level performance fell from $60 to $0.07 per million tokens between November 2021 and October 2024.
Caveat
The trend depends on which benchmark is chosen, and it assumes a user who always switches to the cheapest model above the threshold. Epoch itself reports no standard errors. Epoch dates the o3 example to 31 January 2025, the release date of o3-mini; the full o3 model was released on 16 April 2025. This is a measured past trend, not a forecast.
12,550,185USD, AI-related employment-scam losses, 2025 (all AI-tagged employment complaints, not only deepfake interviews)2026-04 · source
What is claimed
AI is driving a fraud epidemic; third-party write-ups say AI-enabled fraud 'grew 1,210% in one year'.
What was measured, in full, and the caveat
What was measured
The 2025 IC3 Annual Report is the first to include a dedicated AI section. It counts 22,364 complaints carrying the 'AI Related' descriptor, with $893,346,472 in adjusted losses, out of 1,008,597 complaints and $20.877B in losses overall. Most of the AI-tagged loss is investment fraud ($632.0M). Other AI-tagged losses include business email compromise ($30.3M), tech/customer support scams ($19.5M), confidence/romance scams ($19.0M) and employment scams ($12.6M). The report describes deepfaked job interviews under employment, but says these appear aimed at getting into company networks rather than at large dollar losses.
Caveat
The IC3 defines 'AI Related' as a report whose information 'contains a reference to artificial intelligence'. Complainants report it themselves; investigators do not establish it. The report sets AI-tagged investment losses ($632M) against more than $8B in all investment-scam losses and says many victims may not realise AI was involved, so the true count could be higher or lower. The report gives no year-on-year AI figure, so the '1,210%' growth claim cannot be recomputed from it. Press coverage of the report had appeared by 7 April 2026. The PDF served now is dated 16 April 2026, and the FBI press release returned 403.
25.6USD million (Arup-confirmed case)2024-05-17 · source
What is claimed
A deepfake CFO stole $25M, proof that video calls can no longer be trusted.
What was measured, in full, and the caveat
What was measured
Hong Kong police say a finance employee made 15 transfers totalling about HK$200M (US$26M per AFP, US$25.6M per CNN) after video conference calls in which everyone except him was an impersonation of senior staff. Police said the scammers built the fakes from public video and audio, and that the deepfake videos were pre-recorded rather than interactive. The report was made on 2024-01-29, and police said no one had been arrested. In May 2024 an Arup spokesperson confirmed the company was the victim and 'that fake voices and images were used'. The spokesperson added that Arup's internal systems were not compromised.
Caveat
The primary statements (the police briefing and Arup's emailed statement) reach us only through AFP/HKFP and CNN. There is no written police release or court record. Arup has not described the technical method. The sources checked report no arrests and no recovered funds.
80-90percent of tactical operations performed by AI, per Anthropic2025-11-13 · source
4-6human critical decision points per campaign2025-11-13 · source
17organisations targeted by GTG-2002 data extortion (at least; the PDF says 'potentially affecting')2025-08-27 · source
What is claimed
The first AI-orchestrated cyberattack; AI now hacks on its own.
What was measured, in full, and the caveat
What was measured
Anthropic's 13 November 2025 report says a group it assesses with high confidence as Chinese state-sponsored (GTG-1002) used Claude Code against roughly thirty organisations. It says its investigation 'validated a handful of successful intrusions'. By Anthropic's account, AI performed 80–90% of tactical operations, and humans made 4–6 critical decisions per campaign. The same report says Claude 'frequently overstated findings and occasionally fabricated data', for example claiming credentials that did not work. Separately, Anthropic's August 2025 report describes GTG-2002, which used Claude Code for data extortion targeting at least 17 organisations, with ransom demands sometimes above $500,000.
Caveat
This is the vendor's own account of misuse of its own product. Victims are not named, Anthropic published no indicators of compromise for GTG-1002, and the attribution is Anthropic's assessment. Security researchers publicly questioned the autonomy claims and the lack of technical evidence (BleepingComputer, 14 November 2025). No independent corroboration has been published.
Anthropic, September 2026: AI-run intrusions stole national ID records and hit about 200 downstream companies
documenteda specific event recorded in an official finding, a court record or a named organisation's own report
cyber · confidence verified · as of 2026-09-28
300,000+national identity records exfiltrated (one intrusion)2026-09-10 · source
~200downstream customer organisations whose data was taken via one SaaS breach2026-09-10 · source
14 of 42targeted entities internally accessed (hacktivist case)2026-09-10 · source
What is claimed
AI is closing the gap between nation-states and lone actors in cyber offence.
What was measured, in full, and the caveat
What was measured
Anthropic's 10 September 2026 report covers operations it disrupted between December 2025 and August 2026. A Russian espionage actor (GTG-20006; Anthropic says its attribution is consistent with public reporting on Midnight Blizzard) 'used AI at every point in their operations'. One of its intrusions, against a North African government technology authority, took more than 300,000 national identity records and registry data on more than half a million companies. A suspected ShinyHunters affiliate breached a SaaS provider and took data belonging to roughly 200 of its customer organisations. The same affiliate then dumped over 2,100 Azure AD token sets from more than 40 corporate tenants in about 34 hours. A single French-speaking hacktivist targeted 42 entities, got internal access to at least 14, and took an estimated 12–26 GB, including political donor records.
Caveat
This is the vendor's own report. Anthropic says the cases are 'the most notable and novel' threat activity it found, not typical misuse, and it gives no count of all misuse. Victims are unnamed, so the figures cannot be checked independently. Unlike the GTG-1002 report, this one publishes indicators of compromise. Press coverage restates the report and adds no independent confirmation.
6000+reports (articles), not incidents2026-09-28 · source
What is claimed
Thousands of AI harms are catalogued (the site's search box says 'over 6000 reports').
What was measured, in full, and the caveat
What was measured
On 2026-09-28 the highest AI Incident Database ID with a live page is 1,714; IDs 1715 and later return an empty page. The site's own API returns 1,703 incident records, because 11 IDs between 90 and 542 are not in use. The database's roundup says it added 148 incident IDs (1471–1618) from May to July 2026. The 'over 6000' figure on the home page counts reports, meaning articles about incidents, not incidents; 6,533 distinct reports are linked to incidents.
Caveat
Editors curate the entries from press reports, and many are allegations ('Reportedly', 'Allegedly' in the titles). IDs are assigned when an incident is processed, often long after the event, so additions per period show editorial throughput, not how often harms occur. The count of 1,703 comes from the site's GraphQL endpoint, which accepts only requests carrying the site's own origin and is not documented as a public API. The incident pages themselves are public.
OECD AI Incidents and Hazards Monitor: about 17,899 incidents and hazards
measureda count or a rate from a published dataset or study, with its denominator and its method
incidents · confidence verified · as of 2026-09-28
17,899AI incidents and hazards (about)2026-09-28 · source
What is claimed
AI incidents are exploding.
What was measured, in full, and the caveat
What was measured
The OECD AIM page read on 2026-09-28 shows 'About 17899 incidents & hazards'. It describes itself as an 'Automated media discourse monitor of AI incidents and hazards (Beta)' built on Event Registry news data. The same page says AI incidents 'have actually gone down as a share of all AI news'.
Caveat
The monitor detects events automatically from news. It merges incidents (harm occurred) with hazards (harm could occur), and one event can draw many articles. It measures media coverage, not verified harm, and its per-event summaries and incident/hazard classifications are labelled '[AI generated]'. The total is the figure the page shows for its default search. The OECD says the monitor should not be reported as representing official OECD views.
491reports with realistic AI CSAM, 2025 (193 in 2024)2025 · source
65percent of AI videos in Category A (43% for non-AI videos)2025 · source
What is claimed
AI is becoming a 'child sexual abuse machine'.
What was measured, in full, and the caveat
What was measured
In 2025 IWF analysts assessed 3,443 photorealistic AI-generated child sexual abuse videos, out of 63,682 criminal videos in total; in 2024 they found 13. 65% of the AI videos were Category A (most severe), against 43% of non-AI videos. Reports containing realistic AI imagery rose from 193 to 491, yielding 8,029 criminal images and videos.
Caveat
The IWF says the totals depend on where its analysts proactively search and on what the public reports, so part of the rise may reflect detection effort. It also names wider availability of open-source video models as a possible cause. The counts are of content found, not content in existence. The IWF's second page compares the 65% to non-AI images (46%); the 43% used here is from its video page, which compares like with like.
158,000+images/videos categorised as GAI CSAM, Jan 2023–Dec 20252025-12 · source
What is claimed
AI-generated CSAM is flooding the internet.
What was measured, in full, and the caveat
What was measured
NCMEC counts more than 400,000 CyberTipline reports in 2025 with a generative-AI link. It says 182,000+ involved offenders possessing, generating or attempting to generate AI CSAM. That total is made up of 7,000+ reports of generating or possessing, 30,000+ of attempting to generate by uploading a file with text prompts, and 145,000+ of using AI to engage with or alter an existing CSAM file. A further 12,000+ were CSAM that companies found in AI training data, 3,000+ involved other uses such as chat-based exploitation, and 200,000+ had an AI link without enough detail to classify. From January 2023 to December 2025, NCMEC staff categorised 158,000+ reported images and videos as AI-generated. NCMEC says more than 275 direct victims of AI CSAM were identified in 2024 and 2025.
Caveat
'GAI link' is broader than 'AI-generated CSAM': half the 2025 reports cannot say how AI was involved. These are reports, not unique images or victims. Five platforms filed over 75% of all 21.3M CyberTipline reports in 2025. The series is not like-for-like. NCMEC's September 2025 post counted 440,419 GAI reports in January–June 2025 alone, more than the full-year figure now given. Bloomberg (as summarised by Techdirt) reported that about 380,000 of the first-half AI reports were Amazon hash matches to known CSAM found in training data, not AI-generated material. The full-year page lists only 12,000+ training-data reports, so the categories changed between releases.
1,000,000USD paid by carrier Lingo Telecom2024-09-26 · source
What is claimed
Deepfakes are subverting elections.
What was measured, in full, and the caveat
What was measured
The FCC found that on 2024-01-21, two days before the New Hampshire primary, calls Kramer commissioned went out with an AI voice clone of President Biden and a spoofed caller ID. Voice Broadcasting initiated 9,581 of them. The FCC adopted a $6,000,000 forfeiture on 2024-09-26 (released 2024-09-30). It set $1,000 for each of 3,000 verified calls and doubled the total for egregiousness. Kramer did not respond to the proposed fine. The originating carrier, Lingo Telecom, paid $1,000,000.
Caveat
The FCC violation is caller-ID spoofing under the Truth in Caller ID Act, not the deepfake itself. No effect on turnout has been measured. On 2025-06-13 a New Hampshire jury acquitted Kramer of all state voter-suppression and candidate-impersonation charges (NHPR, AP). The FCC forfeiture order is separate and was not affected. These sources do not say whether the fine has been collected.
115,608 Flock Safety ALPR cameras mapped in OpenStreetMap (repo's own measurement)
measureda count or a rate from a published dataset or study, with its denominator and its method
surveillance · confidence verified · as of 2026-09-28
115,608Flock ALPR cameras mapped in OpenStreetMap2026-09-26 · source
2,287 of 3,235US counties with at least one mapped camera2026-09-26 · source
What is claimed
Flock's AI-powered camera network watches every car in America.
What was measured, in full, and the caveat
What was measured
This repo's data/flock.json joins three OSM tags (manufacturer, brand and operator = Flock Safety) on ALPR surveillance nodes. It counts 115,608 Flock ALPR cameras mapped worldwide as of 2026-09-26, of which 115,350 are inside a US county. 2,287 of 3,235 US counties have at least one mapped camera; 948 have none mapped.
FTC alleges Rite Aid's AI facial recognition produced 'thousands of false-positive matches'; five-year ban
allegedcharged in a complaint or a lawsuit and not adjudicated; the allegation is the record, not a finding
surveillance · confidence verified · as of 2026-09-28
5years banned from facial-recognition surveillance2023-12-19 · source
2012-2020years deployed, per the FTC complaint2023-12-19 · source
What is claimed
AI surveillance misidentifies people of colour.
What was measured, in full, and the caveat
What was measured
The FTC's complaint says Rite Aid used AI facial recognition in hundreds of stores from 2012 to 2020 and that it generated 'thousands of false-positive matches'. Staff acting on false alerts followed, searched, ejected and called police on customers. The complaint says false positives were more likely in stores in plurality-Black and plurality-Asian communities than in plurality-White ones. The settlement was announced on 2023-12-19 and filed as a stipulated order on 2024-03-08. It bans Rite Aid from facial-recognition surveillance for five years.
Caveat
The harm facts are the FTC's charges, resolved by a stipulated order with no adjudication. Rite Aid said it 'fundamentally disagree[s]' with them and called the system a pilot in a limited number of stores. Only the ban is a court order. 'Thousands' is the FTC's word; it gives no exact count. Rite Aid closed all its remaining stores in 2025 after a second bankruptcy, so the ban now has little practical effect.
17percent growth in data-centre electricity demand, 20252026-04-16 · source
50percent growth in AI-focused data-centre electricity, 20252026-04-16 · source
What is claimed
AI will devour the grid.
What was measured, in full, and the caveat
What was measured
The IEA's Energy and AI report (10 April 2025) estimates data-centre use at about 415 TWh in 2024, about 1.5% of global electricity. It says use grew about 12% a year over the previous five years. The IEA's 16 April 2026 update puts 2025 at 485 TWh, with data-centre demand up 17%. It says electricity use at AI-focused data centres grew even faster, 'surging 50% in 2025'.
Caveat
These are IEA modelled estimates, not metered totals; no public feed meters data-centre use. The same 2026 report says energy per AI task has fallen by at least an order of magnitude a year in recent years. It also says a simple text query now typically uses less electricity than running a television for the same time. The IEA's projections for future years are left out here.
1,171 of 1,877 mapped US datacentre sites sit in counties with D1+ drought (repo's own measurement)
measureda count or a rate from a published dataset or study, with its denominator and its method
energy and water · confidence verified · as of 2026-09-28
1,171 of 1,877mapped US datacentre features in a county with any D1+ drought area2026-09-22 · source
534datacentre features in counties whose worst category is D2 or worse2026-09-22 · source
929datacentre features in counties with at least 50% of area in D1+2026-09-22 · source
What is claimed
Datacentres are draining water from drought-stricken communities.
What was measured, in full, and the caveat
What was measured
This repo's data/datacenters.json (generated 2026-09-28T02:22Z) joins 1,877 OSM-mapped US datacentre features to the US Drought Monitor map of 2026-09-22. The features are 1,769 operating, 95 under construction and 13 proposed. 1,171 are in counties where the worst category covering any part of the county is D1 or worse. 534 are in counties where it is D2 or worse, and 929 are in counties with at least half their area in D1 or worse. All four figures were recomputed from the file and match.
Caveat
This shows where datacentres are, not what they consume; the file itself says 'Nothing here measures a datacentre's power or water draw'. The 1,877 are mapped OSM features, not facilities. At least 284 share a name and county with another feature (for example, 34 'Amazon Web Services' features in one Oregon county), so large campuses count several times. Drought is a share of county area and a weekly expert judgement. OSM coverage is uneven. The site's headline 'in drought' figure (1,350) also counts D0, abnormally dry, which is not drought.
3,462announced job cuts citing AI, August 20262026-09-03 · source
What is claimed
AI is already taking hundreds of thousands of jobs.
What was measured, in full, and the caveat
What was measured
Challenger, Gray & Christmas tallies the reasons employers give when they announce job cuts. AI was cited for 54,836 of 1,206,374 announced cuts in 2025 (about 4.5%), and for 71,825 since the category began in 2023. From January to August 2026, AI was cited for 116,175 of 529,914 (about 22%), making it the most-cited reason so far this year. In August 2026 it fell to 3,462, its lowest monthly total since December 2025.
Caveat
These are announced plans with the employer's stated reason. They are not measured job losses, and they do not prove AI caused the cuts. Total announced cuts in 2026 are down 41% year on year. A separate 'Technological Update (possibly AI)' category counted 20,219 in 2025.
Organization Science study: after ChatGPT, Upwork freelancers in exposed occupations had about 2% fewer monthly jobs and 5.2% lower monthly earnings
measureda count or a rate from a published dataset or study, with its denominator and its method
labour · confidence verified · as of 2026-09-28
2percent fewer monthly jobs for affected freelancers2023-08 · source
5.2percent lower monthly earnings for affected freelancers2023-08 · source
What is claimed
Generative AI is replacing knowledge workers.
What was measured, in full, and the caveat
What was measured
Hui, Reshef and Zhou (Organization Science 35(6):1977–1989, November 2024) use a difference-in-differences design on Upwork. After ChatGPT's release, freelancers in more-affected occupations saw monthly jobs fall 2% and monthly earnings fall 5.2%, relative to less-affected occupations. They find similar effects after image-generation models, and suggestive evidence that top freelancers were hit harder.
Caveat
The exact percentages come from the CESifo working paper (August 2023). The published abstract gives only the direction of the effect, and the journal page returned 403. In a July 2025 Brookings piece summarising the published paper, the authors give about 2% fewer new monthly contracts and roughly 5% lower monthly earnings. The study covers short-run effects on one online freelance platform, not the wider labour market.
Wrongful-death suits over teen suicides: Character.AI and Google settled; the Raine v. OpenAI case continues
allegedcharged in a complaint or a lawsuit and not adjudicated; the allegation is the record, not a finding
safety incidents with deployed chatbots · confidence verified · as of 2026-09-28
2025-05-20date of order granting in part and denying in part the motions to dismiss in Garcia (entered 2025-05-21)2025-05-20 · source
2026-01-07Notice of Resolution filed by Character Technologies; case dismissed without prejudice the same day2026-01-07 · source
100+times ChatGPT directed Adam Raine to crisis resources, per OpenAI's answer2025-11-25 · source
What is claimed
Companion chatbots are killing teenagers.
What was measured, in full, and the caveat
What was measured
Garcia v. Character Technologies (M.D. Fla. 6:24-cv-01903) was filed on 2024-10-22 over the death of 14-year-old Sewell Setzer III. On 2025-05-20 the court granted in part and denied in part the defendants' motions to dismiss. On 2026-01-07 Character Technologies filed a Notice of Resolution, and the court dismissed the case without prejudice pending a stipulated final order. That deadline was extended on 2026-04-09 (by 120 days) and again on 2026-08-04 (by 90 days). CNN reported that the agreement covers Character.AI, its founders and Google, as well as four other cases. Raine v. OpenAI (SF Superior Court CGC-25-628528) concerns 16-year-old Adam Raine. In its 2025-11-25 answer, OpenAI says ChatGPT directed him to crisis resources or trusted people 'more than 100 times' and that his death 'was not caused by ChatGPT'. No resolution of Raine has been reported.
Caveat
No court has found any chatbot maker liable. The settlement terms have not been made public, and a settlement is not a finding of liability. Causation is disputed. The deaths are documented; the role of the chatbots is alleged. The pending status of Raine comes from press and case trackers; the San Francisco Superior Court docket was not read directly.
2,086 court and tribunal decisions dealing with AI-hallucinated material in filings
documenteda specific event recorded in an official finding, a court record or a named organisation's own report
safety incidents with deployed agents or chatbots · confidence verified · as of 2026-09-28
2,086court/tribunal decisions on AI-hallucinated content2026-09-27 · source
What is claimed
AI is corrupting the justice system.
What was measured, in full, and the caveat
What was measured
Damien Charlotin's AI Hallucination Cases database lists 2,086 decisions as of 27 September 2026. Each is a legal decision in which a court or tribunal found, or implied, that a party relied on hallucinated content, usually fake citations. It does not count all fake citations filed, only those a decision addresses.
Caveat
This is one researcher's registry of primary court decisions, not an official statistic. By its own description it includes some decisions where AI use was alleged but not confirmed. Coverage is best in English-language jurisdictions: 1,430 of the decisions are from the US. The filter tags attribute 1,200 to pro se litigants and 828 to lawyers.
UN Panel on Libya documents deployment of the Kargu-2, which it calls a lethal autonomous weapons system (2020)
documenteda specific event recorded in an official finding, a court record or a named organisation's own report
autonomous weapons · confidence verified · as of 2026-09-28
63paragraph of S/2021/229 documenting Kargu-2 use2021-03-08 · source
What is claimed
The first time a killer robot autonomously hunted and killed humans.
What was measured, in full, and the caveat
What was measured
The UN Panel of Experts report S/2021/229 (8 March 2021, para. 63) covers the period after Operation PEACE STORM began on 27 March 2020. It says logistics convoys and retreating HAF forces 'were subsequently hunted down and remotely engaged' by drones 'or' lethal autonomous weapons such as the STM Kargu-2. It says these were 'programmed to attack targets without requiring data connectivity between the operator and the munition'. The Panel's table records a Kargu-2 transfer from Turkey, identified on 27 May 2020. This documents that the system was deployed; the report attributes no casualties to autonomous engagement.
Caveat
The sentence joins drones 'or' Kargu-2, so it does not say which system engaged whom, or whether anyone was killed in autonomous mode. Paragraph 63 ends with a footnote citing a confidential source. An STM official told Shephard (July 2021) that the Kargu-2 keeps a human in the loop during attacks. The IDF's own statements on the Gospel and Lavender targeting systems could not be read today, so they are not included.
Which counters come from public registries?The counters · 6
6 live, 0 dark.
Each is published as its publisher reports it, in its own unit and its own window, and none is set against
any other: the page does not pair a harm counter with a benefit counter, and it adds nothing up.
Window for the windowed ones: 2026-09-15 to 2026-10-15,
the 30 whole UTC days before the anchor day plus the anchor day so far, both ends inclusive; the anchor is generated_at. A counter that failed is dark, with its error, and never a zero or an earlier run’s figure.
Which counters are on the Benefit side?3 counters
live · context only
AlphaFold DB structure predictions, single-chain plus complexes (total)
261,328,474 structure predictions
no window: a running total
Counts
Structure predictions the AlphaFold Protein Structure Database search returns, single chains and complexes together.
Does not count
Experimental structures, validated structures, or anything a prediction was used for. A prediction is not a drug. The total jumps whenever a provider batch lands, so a rise says a batch landed.
ClinicalTrials.gov: AI trials that actually started in the window
13 trials started
2026-09-15 to 2026-10-15
Counts
Registered studies matching "artificial intelligence" or "machine learning" in any of about 40 registry fields, whose start date falls in the window AND is marked ACTUAL by the sponsor. Interventional and observational both.
Does not count
Results, approvals or patient benefit: a trial starting is a commitment, not an outcome. Trials whose sponsor has not yet updated the start to ACTUAL, so the newest days read low. Registered plans whose start is still marked estimated.
filter AREA[StartDate]RANGE[2026-09-15,2026-10-15] AND AREA[StartDateType]ACTUAL
FDA AI-enabled medical device authorisations (cumulative)
1,614 devices authorised, cumulative
no window: a running total
Counts
Every device on the FDA AI-Enabled Medical Device List, one row per authorisation (510(k), De Novo or PMA), from the first in 1995 to the newest the file carries.
Does not count
Use, outcomes or benefit to patients: the list records marketing authorisation. Devices the FDA has not yet added (the list lags by about two months, so the newest months read low). Devices the FDA did not identify as AI-enabled.
newest decision in the file 2026-06-29 · file last modified Fri, 04 Sep 2026 11:54:10 GMT · 0 decisions in the 30-day window, because the list lags
AI Incident Database: incidents first entered in the window
58 incidents entered
2026-09-15 to 2026-10-15
Counts
Incidents whose earliest linked report was submitted to the database inside the window: the editors' intake, in whole UTC days.
Does not count
Harms that happened in the window (the incident date lags entry by months or years). Verified harms: the database is volunteer-curated from press reports and many entries are allegations. Incidents nobody reported to it.
1,736 incidents in the database · incidents 88 and 141 have no entry date and are listed, not guessed · newest entry day 2026-10-10
CISA KEV: vulnerabilities catalogued as exploited in the window (not AI-specific)
35 vulnerabilities catalogued
2026-09-15 to 2026-10-15
Counts
Entries CISA added to its Known Exploited Vulnerabilities catalogue inside the window, for any vendor and any product.
Does not count
Anything specific to AI. When exploitation happened: dateAdded is the cataloguing date. Exploitation the US government did not observe, or did not publish.
15 in calendar month 2026-10 · catalogue version 2026.10.14
OECD AI Incidents and Hazards Monitor: incidents and hazards dated in the window
681 incidents and hazards
2026-09-15 to 2026-10-15
Counts
News-detected AI incidents and hazards dated inside the window, as the OECD monitor classifies them. The classification is made by OpenAI models in the OECD's pipeline, not by a person and not by this site.
Does not count
Verified harm: it counts media-reported events. It does not separate incidents (harm occurred) from hazards (harm plausible). The last three or so days of the window are incomplete.
classified by OECD pipeline: GPT-4o mini filter, GPT-4o incident-or-hazard, o3-mini metadata and clustering · events are processed one to four days after they occur
Context only. Neither index feeds the beam, and the two are never subtracted, compared or combined in this file.
SIREN 4 · ROUTINE
56.6of 100
as of 2026-10-15 05:21:57 UTC
BLISS
awaiting baselineno score
as of 2026-10-15 05:24:59 UTC
BLISS has no score. Its sources answered —
10 of 10 this run —
but no frozen reference distribution exists yet,
so every value is published and none is scored. That is awaiting baseline, not dark, and it is never printed as a number.
BLISS’s 10 sources as read on
2026-10-15 05:24:59 UTC. A value awaiting a baseline is a real measured number with no frozen
history to score it against; a dark source did not answer, and its error is printed instead of a value.
Published in the two files and printed here unedited: the collector’s five lines about the
counts, then the registers’ lines about the rows.
What do the honesty lines say about the counts?
No net score. Nothing in this file adds a harm to a benefit, subtracts one from the other or divides one by the other, because the two sides share no unit.
The beam compares two counts made by one rule over one set of headlines: distinct stories in the newsroom's current window that matched the harm list only, against those that matched the benefit list only. It moves on nothing else.
A headline count is a count of words, not of outcomes. The stories that match neither list are counted and published beside the other two every time, so the two pans never pose as the whole newsroom.
Each counter is published as its publisher reports it, in its own unit and window. A counter that failed is dark, with its error, and never a zero.
SIREN and BLISS are printed as context and never compared here. A BLISS with no frozen reference is printed as awaiting baseline, never as a number.
What do the honesty lines say about the registers?
No net score. The two registers are never summed, subtracted or weighed against each other. A fraud loss in dollars, a lung-function change in millilitres and a count of mapped cameras share no unit, and any exchange rate between them would be a value judgement dressed as arithmetic.
Every figure is taken from a linked source, and every number carries its own source and date. Where sources disagree, the caveat says so. Nothing is estimated.
Curated by hand and dated, not live. Each row was checked against its sources on its as_of date. The live counts are in data/balance.json, which is fetched; this file is not.
Each row records the claim beside the measurement so a reader can see the gap between them. Recording a claim is not endorsing it.
The number of rows on each side is a choice made by the compilers, not a measurement. Equal lengths are not a finding that the two sides are equal, and unequal lengths would not be a finding either.
A status word describes the evidence, not the importance. No status on one side ranks a row against any row on the other.
The verification pass of 2026-09-28 dropped no row outright. dropped[] lists what individual rows left out, and why.
Every precision figure for the word lists is in-sample.
The lists were tuned on the same headlines they were scored on, by one labeller; the first honest test is
the next window. The benefit side’s precision rests on 2 matches in all history.
Where did this come from?
What is the rule on both sides?
Matching. exact word forms after lower-casing, accent-folding and splitting on anything that is not a letter or a digit (wordSequence in collector/news-stories.mjs); phrases of 1 to 4 consecutive words; no stemming.
Not scanned. Items of kind paperlabreleasemodel.
Hedges. a headline ending in "?" (a trailing credit in parentheses allowed after it), or carrying a hedge term, counts on neither side.
Negation. a hit with no, not, never up to 2 words before it is ignored.
Threshold. A side counts on 1 headline hit, or on 2 distinct summary hits.
Symmetry. One rule, applied to both lists in the same pass.
What is on the harm list? 78 terms · lexicon v5
harmharmsharmedharmfulrogueincidentincidentsbreachbreachedbreacheshackedhackingtried to hackattempted to hackattempts to hackattempting to hackattackattacksattackedattackingexploitexploitsexploitedcompromisedunauthorizedunauthorisedleakleaksleakedleakinglawsuitlawsuitssuedsuessuingscamscamsscammerscammersfraudfraudulentdeceptivedeepfakedeepfakesabuseabusedabusivemisinformationdisinformationlethalweaponiseweaponisesweaponisedweaponizeweaponizesweaponizedcatastrophickilleddeathslayoffslaid offjob cutsinfringesinfringementinfringingmalwareransomwarephishingjailbreakjailbreaksjailbrokenstolenmisusemisusedwithout consentwithout permissionfinedfines
The two lists differ in size and reach. Both are printed in full so the difference is
visible, and it is not corrected by any weighting. Any edit to either list makes a new instrument.
Dropped from the previous version after
producing false positives: harm harassedharassmentaddictiontheftendangerendangersendangeringspyingsurveillanceai weaponsautonomous weapons; benefit healhealshealed.
What is the measured precision?
One labeller. An item counts as Y only if it reports something that happened; commentary about what has not happened is not harm, and funding rounds and vendor self-announcements are not benefit. Sample: data/news.json generated_at 2026-09-28T02:46:46.998Z (200 items, 85 non-paper), plus every distinct item in its 483 git revisions (757 items, 324 non-paper).
0 items in the sample matched both lists.
Side and sample
Matched
Precision, strict
Wilson 95%
Known reports found
Harm, current file
11
100%
lower bound 74.1%
11 of 16
Harm, all history
26
88.5%
71 to 96%
23 of 34
Benefit, all history
2
100%
lower bound 34.2%
2 of 4
Benefit, current file
0
—
—
—
What are the registers?
Re-verified by hand against every linked source before as_of_date moves. A figure that changes is re-dated with its new source, never silently replaced. A row that no longer holds moves to retired[] with the date and the reason, and its id is never reused.
Left out of individual rows, and why:
The IDF's own statements on the Gospel and Lavender targeting systems (harm,
noted in un-libya-kargu2). Could not be read on 2026-09-28, so they are not included.
The IEA's projections of data-centre electricity use for future years (harm,
noted in iea-datacentre-electricity). Left out: every row here is a measured or documented figure, and a projection of future years is neither.
No row has been retired.
The file was valid on this run, checksum sha256:009983d7662f7804145c32c9181c073b260a4fc3bb9d8b7c4cad67a0b1e02038.
balance.json v1.0.0 · schema 1 ·
stamped 2026-10-15 05:24:59 UTC, the newest timestamp among the inputs (news, SIREN, BLISS, ledger as_of_date), never the clock ·
built by collector/balance.mjs.
This page does not feed the SIREN index and has no route into it.
Where can I take the data?
/api/balance.jsonThe newsroom counts and every matched and unmatched headline, the beam rule, the six counters with
their windows and errors, SIREN and BLISS as context, and both word lists in full.
/api/ledger.jsonThe two registers: 18 benefit and 18 harm rows, each with
its numbers, sources, status and caveat, key-sorted so the checksum in the repository is the checksum served.
What questions do people actually ask?
How does SIREN count AI harm and benefit side by side?
The utopia case, as far as it has receipts: what AI has measurably done for people, beside what it has measurably done to them. Three layers, each on its own terms — a count of two word lists in the same headlines, six counters read from public registries, and two registers of dated, sourced results — and one refusal. Nothing here is added up, netted or divided into a verdict, because the two sides share no unit.
Does the balance feed the main SIREN number?
The balance · a SIREN view · does not feed the main number. This page does not feed the SIREN index and has no route into it.
Is there a net score of harm versus benefit?
No net score. Nothing in this file adds a harm to a benefit, subtracts one from the other or divides one by the other, because the two sides share no unit.
What do the newsroom word-list counts measure?
24 of 500 stories in the newsroom's current window matched the harm list and 5 matched the benefit list, as of 05:24 UTC on Thu 15 Oct 2026. A count of words, not of outcomes.
What does the beam never tilt on?
the counters the registers in data/ledger.json SIREN against BLISS severity tiers, item scores, recency, corroboration or engagement counts from different windows or source sets any language-model or sentiment judgement
What do the honesty lines say about the counts?
Published in the two files and printed here unedited: the collector’s five lines about the counts.
What do the honesty lines say about the registers?
Published in the two files and printed here unedited: the registers’ lines about the rows.
What is the rule on both sides?
One rule, applied to both lists in the same pass.
What is on the harm list?
Both word lists are printed in full so the difference is visible, and it is not corrected by any weighting.
What is on the benefit list?
Both word lists are printed in full so the difference is visible, and it is not corrected by any weighting. Any edit to either list makes a new instrument.
Which hedge words and negators apply?
A word with “no”, “not” or “never” just before it was ignored and is not marked. The hedge and negator chips are printed under the method section.
What is the measured precision?
Every precision figure for the word lists is in-sample. The lists were tuned on the same headlines they were scored on, by one labeller; the first honest test is the next window.
What are the registers?
Re-verified by hand against every linked source before as_of_date moves. A figure that changes is re-dated with its new source, never silently replaced. A row that no longer holds moves to retired[] with the date and the reason, and its id is never reused.
Which items were set aside by the rule?
These carried a word from one of the lists and still counted on neither side. They are listed so the rule can be checked in the direction that does not flatter it.
Which headlines matched Benefit language?
Headlines in the newsroom window that matched the Benefit word list — a measured count, never a net against the other side.
What is in the Benefit register?
Dated, sourced Benefit register rows (The utopia case), each carrying what was measured beside what is claimed about it.
Which counters are on the Benefit side?
Benefit counters from public registries — each published as its publisher reports it, and never paired with the other side.
Which headlines matched Harm language?
Headlines in the newsroom window that matched the Harm word list — a measured count, never a net against the other side.
What is in the Harm register?
Dated, sourced Harm register rows (The doom case), each carrying what was measured beside what is claimed about it.
Which counters are on the Harm side?
Harm counters from public registries — each published as its publisher reports it, and never paired with the other side.
How do SIREN and BLISS sit beside the beam?
Context only. Neither index feeds the beam. Both are percentile positions in two separate frozen reference distributions of activity tempo, not measures of harm or benefit.
What can this page not say?
Published in the two files and printed here unedited: the collector’s five lines about the counts, then the registers’ lines about the rows.
Where did this come from?
The matching, hedge and negation rules from the published lexicon, both word lists in full, and the dated register rows — with the stamp from collector/balance.mjs.
Where can I take the data?
The newsroom counts and every matched and unmatched headline are at /api/balance.json; the two registers are at /api/ledger.json.
What did the newsroom’s headlines carry?
24 of 500 stories in the newsroom's current window matched the harm list and 5 matched the benefit list, as of 05:24 UTC on Thu 15 Oct 2026. A count of words, not of outcomes.
Which headlines matched a word list?
Newest first. The marked words are the ones the rule counted; a word with “no”, “not” or “never” just before it was ignored and is not marked.
What documented results sit on both sides?
Two registers kept side by side: what AI has measurably delivered or demonstrated, and what harm it has been documented, measured or alleged to cause. Each row separates what was measured from what is claimed about it, with the sources for both.
Which counters come from public registries?
6 live, 0 dark. Each is published as its publisher reports it, in its own unit and its own window, and none is set against any other.
