The State of Play

Where AI Stands in Computer Vision & Sensing

Computer vision is the domain where the model is rarely the problem any more. Object detection, OCR, shelf recognition and organ segmentation have become commodity capabilities: Google Lens handles 20 billion visual queries a month, TotalSegmentator and nnU-Net turn up in clinical research papers as unremarked plumbing, retail practitioners describe the recognition problem at the shelf as largely closed, and LiDAR has become a volume component, with RoboSense shipping 719,200 units in the first half of 2026 and Seyond 450,900. Yet only one practice in the domain, in-store shelf and customer analytics, has reached the point where deployment is the norm rather than the exception, and even there Coresight finds retailers that have scaled store intelligence lose a larger share of sales to execution failures (7.6%) than those still piloting (5.6%). Everywhere else the story has the same shape: a vanguard extracting real value, a long tail that has not started, and a set of constraints that sit outside the model.

What distinguishes this domain from its neighbours is that those constraints are physical and political rather than algorithmic. Sensing degrades with the world: LiDAR loses 59% of detections in fog, production L2+ driver-assistance systems run at 60–70% perception effectiveness in rain and darkness, PPE cameras cannot resolve colour-coded equipment beyond about seven metres or under 50 lux, and the EU's Entry/Exit System lost fingerprint scanners to summer heat at Dover. And where the sensed object is a human face or body, regulators have moved from warnings to shutdowns. Italy's Garante ordered Rome's Olympic Stadium to switch off a facial-recognition access system after four years of live operation; Yoti withdrew its digital ID app from Spain rather than re-architect after a €950,000 fine; Turkey has banned workplace biometric access outright; Boston terminated its Flock Safety pilot within days after the vendor's default nationwide-lookup feature exposed plate data to outside agencies, and Flock lost 90 municipal contracts in August alone. Other AI domains argue about hallucination; this one argues about legal basis.

Momentum concentrates where the sensor points at infrastructure rather than people, and where the buyer is institutional. Wildfire detection has hard numbers: ALERTCalifornia's 1,300 cameras flag roughly half of incidents before a 911 call, FireSat and OroraTech constellations are operational, and Greece runs two national satellite scans a day. Adaptive traffic signals are spreading across continents with quantified delay reductions—Chennai's 105-intersection system claims ₹62 crore a year in time-value savings, Beijing's 62-intersection AI agent cut delay 15–30%, Brisbane has committed $15 million to a 37-intersection trial. Geospatial AI is becoming a business: Satellogic posted its first positive GAAP operating income, Planet grew revenue 58% year on year, and KoBold Metals' Mingomba copper project has moved from AI-driven discovery into $2.3 billion of construction. Clinical imaging has the strongest evidence base of any practice here—MASAI's 105,934-woman RCT, Germany's 500,000-woman PRAIM programme, cardiac 3D modelling now standard of care at 20-plus US children's hospitals, Johns Hopkins narrowing Black-patient screening-adherence gaps from 15.6 to 3.5 percentage points—and the weakest conversion: fewer than 2% of US radiology practices use cleared tools, only 0.2% of 1,357 FDA-cleared AI devices have been tested against patient outcomes, and a prospective crossover study this month found none of four commercial chest X-ray algorithms improved radiologists' accuracy on any finding. Checkout-free retail, the domain's most publicised consumer bet, is now in retreat: Amazon has closed its remaining owned Just Walk Out stores in favour of portable RFID lanes while licensing the technology to 360-plus third-party venues, and AiFi has taken a down-round at roughly $100 million pre-money with a third fewer staff than at its 2023 peak.

What's New, 2026-09-07 to 2026-09-21

Two trajectories moved this fortnight, both downward. Checkout-free retail is now declining rather than merely stalled: alongside AiFi's down-round, CNN documented the Amazon Go closures and the self-checkout reversals at Walmart, Target and Costco; REWE's 1,200 m² Hamburg flagship remains a rarity among roughly 30 German sites, most still in pilot; and Żabka Nano, the category's best European proof point with 50-plus stores, has slowed retail expansion and pivoted to offices, hospitals, universities and its first airport. Perimeter and biometric identification slipped from advancing to stalled after nine Schengen states, including France, Germany and Italy, were granted open-ended flexibility to postpone full Entry/Exit enforcement once the 6 September grace period ended. The same fortnight brought Nigeria's BVAS audit showing biometric authentication failures denying voters at 10% of sampled polling units with no fallback route, Mordor's estimate of a 741% annual rise in injection attacks, and a RAND Europe study warning that human-in-the-loop safeguards do little against automation bias.

The other significant movement is in radiology, where the evidence split sharpened in both directions at once. The Academic Radiology crossover study (1,861 images) found 71% of AI-prompted revisions turned correct reads into incorrect ones; Northwell's shadow-mode Aidoc study lifted aneurysm detection 39% while dropping positive predictive value from 92.7% to 78.2%; and a European Radiology cohort of 879 women found AI-detected breast cancers show no survival advantage once tumour biology is adjusted for. Against that, CMS approved the first radiology AI reimbursement code ($137.53 per Aidoc CT triage case from October), Vara's autonomous mammography triage scaled to 250,000 monthly exams across half of German screening programmes on a CE Class IIb mark, and the FDA funded a $1.29 million LLM-jury framework to evaluate autonomous reports. In law enforcement, the UK Biometrics Commissioner declared live facial recognition fundamentally incompatible with human rights in the same week Kent Police ran its first deployment, the US Interior Department contracted Clearview for missing and murdered Indigenous persons cases, and WIRED revealed Clearview had built and tested InquiryIQ, pairing face matches with web-scraping and LLM profiling to generate personal dossiers. Smaller but telling: Google's own merchant documentation now states virtual try-on does not indicate fit or recommend a size, and Taobao shoppers delete more than half of try-on carts; Helmholtz found satellite remote sensing explains 3–5% of wild-bee diversity against 60% for field surveys; edge vision-language models fabricated 5.9–9.6% of species identifications on camera-trap imagery; a document-automation pilot scoring 94% on DocVQA achieved 41% straight-through processing; Kroger's union filed grievances as Fred Meyer pushed Simbe's Tally toward full-fleet coverage; and confirmation that Gemspring's Trax acquisition came with $170 million of trailing revenue underlined shelf-analytics consolidation.

Key Tensions

• The benchmark-to-field gap. Published accuracy is failing to survive contact with production across the domain. A document pilot scoring 94% on DocVQA delivered 41% straight-through processing; people-counting systems at 97–98% in controlled retail drop to net undercounts of up to 18% on identical hardware in the field; a single unchanged imaging model ranged from AUROC 0.427 to 0.819 across 23 sites; frame-level anomaly-detection benchmarks translate to event-level precision below 10%. IEC 62676-6:2026 exists precisely because vendor figures like "99.995% accuracy" are unfalsifiable.

• Legal basis, not accuracy, decides biometric deployment. The EU AI Act's Article 5 prohibition, France's Conseil d'État precedent requiring a separate statutory basis for algorithmic video analysis, and the Rome stadium shutdown mean four years of successful operation are no defence. The US is bifurcating: ICE fields 1,570 iris scanners and Interior has hired Clearview, while Virginia mandates 98% NIST-validated accuracy, Boston walked away from Flock within days, and Flock is losing cities at a record pace. Vendors are responding architecturally—match-on-card, on-device enclaves—rather than waiting for harmonisation.

• Detection is solved; response is not. Value is stuck downstream of the alert. LungIMPACT cut chest X-ray reporting time from 47 to 34 hours without shortening time to lung-cancer diagnosis (44 versus 46 days); 75% of shelf-monitoring issues go unresolved within 48 hours absent a dedicated response team; Starbucks' NomadGo post-mortem attributed 97% of the failure to governance and change management; 83% of alerts in mature security operations are false and 74% of buyers struggle to realise value within 90 days. The organisations that win, like Denmark's Capital Region with mammography triage, redesigned the workflow around the model rather than bolting the model onto the workflow.

• Physics and geography cap the sensing envelope. The hardest robustness barriers are not software problems. Nanocoatings and 1550nm lasers are being pursued because fog and wet roads defeat current LiDAR; 90% of high-hazard construction zones lack full-time power; wildfire stations cost $50,000-plus a year and constellations $400 million, so high-confidence systems cluster in North America, Europe and Australia while 92.3% of wildfire ML papers lack public code and skip the highest-burn regions. Instituto Humboldt's warning this month about confident false outputs and biases against Indigenous knowledge comes from a megadiverse country the tools were not trained on.

• Regulatory clearance has decoupled from clinical evidence. Radiology accounts for 76% of the 1,500-plus FDA-cleared AI devices, yet only 1.6% cite RCT data, devices lacking clinical evidence carry a 1.39× recall hazard, DermaSensor has logged 11 post-market reports, and a PRISMA review places most healthcare AI at research or early-test stage. Automation bias is now measured rather than hypothesised: incorrect AI advice pushes radiologist false negatives from 2.7% to 33%. The counter-evidence is real—PulmoFoundation's RCT lifted pathologist accuracy from 83.2% to 91.7%—which is why the field's centre of gravity is shifting toward augmentation and post-market monitoring rather than autonomy.

Top 10 Evidence Items

1. CNN: Documented Retreat from Unattended Retail Automation (news-coverage) — The clearest single marker of checkout-free retail's reversal, which the briefing leads with as this fortnight's biggest downward move. https://edition.cnn.com/2026/09/17/business/technology-retail-food-economy
2. AiFi Vendor Analysis: Financial Distress and Market Consolidation (opinion) — Quantifies the financial distress behind the retreat, showing this is a market contraction, not just a few store closures. https://www.teardown.ai/companies/aifi
3. Sports venues face growing scrutiny over use of facial recognition technology (news-coverage) — The Rome stadium shutdown is the briefing's paradigm case that legal basis, not track record, now decides biometric deployment. https://www.pinsentmasons.com/out-law/news/venues-face-growing-scrutiny-facial-recognition-technology
4. Boston Police Department terminated Flock Safety pilot after vendor data-sharing breach exposed licence plate data to outside agencies (news-coverage) — Shows how fast institutional trust collapses once a vendor default exposes data, echoing the domain's shift from warnings to shutdowns. https://www.wbur.org/news/2026/09/14/flock-boston-police-pilot-surveillance-technology-report
5. EU Entry/Exit System operational status: Nine countries granted formal flexibility to postpone full EES enforcement after 6 September grace period (news-coverage) — Documents the EES stall that the briefing cites as perimeter security's demotion from advancing to stalled this fortnight. https://www.biometricupdate.com/202609/eu-lets-9-schengen-countries-delay-full-ees-biometric-checks
6. Aidoc Brain Aneurysm Detection: Sensitivity Gain vs Positive Predictive Value (research-paper) — The Aidoc PPV trade-off is the sharpest illustration of the widening split in radiology's evidence base. https://www.insideprecisionmedicine.com/topics/patient-care/ai-boosts-brain-aneurysm-detection-but-produces-more-false-positives/
7. Long-Term Breast Cancer Survival: AI-Detected vs Radiologist-Detected (research-paper) — Undercuts the value narrative for AI-detected cancers, the uncomfortable counterweight to radiology's positive case studies. https://www.lunit.io/ko/resource/article/does-it-matter-for-long-term-patient-survival-whether-breast-cancer-is-detected-by-ai-or-not/
8. CHOP cardiac 3D modeling in routine clinical practice (case-study) — Grounds the claim that clinical imaging has real, unremarked deployment wins alongside its conversion problem. https://blogs.nvidia.com/blog/childrens-hospital-open-source-ai-cardiac-care/
9. Autonomous diabetic eye disease screening narrows racial testing-adherence gaps: Johns Hopkins real-world study (case-study) — Concrete equity gain that substantiates the briefing's rare unambiguous success story in imaging. https://lanfrica.com/fr/record/autonomous-artificial-intelligence-for-diabetic-eye-disease-increases-access-and-health-equity-in-underserved-populations
10. Indivd Field Accuracy Benchmark — Controlled vs. Production Degradation (adoption-metric) — Directly evidences the benchmark-to-field gap tension with quantified controlled-vs-production degradation. https://www.indivd.com/insights/reports