Public Safety & Emergency ServicesIssue 537 records
Public Safety
Tonight’s public-safety edition is about testing before trusting. A peer-reviewed Penn experiment shows facial recognition errors shift with image quality and fall unevenly across demographic groups. The UK’s Police AI lab proposes supplier testing against held-back data, Chicago plans an evaluated six-month gunshot-detection pilot amid confusion inside its police department, and Berlin starts a four-week trial of behaviour-only AI video analysis. In courts, the England and Wales Civil Justice Council recommends targeted AI declarations rather than blanket disclosure, and an Ohio federal court relies on Rule 11 instead of banning AI for self-represented plaintiffs. Campus police exits from Flock show contract lock-in risk.
- Evidence records
- 7
- Cross-source patterns
- 2
- Evidence classes
- 5 independent reporting1 standards or public-body guidance1 academic research
- Outcomes
- 4 emerging2 cautionary1 mixed
- Source freshness
- 7 new this fortnight
- Research completed
- 2026-10-30
Choose a role to see its takeaway beside every record in the ledger.
Patterns across the evidence
Synthesis · Lighthouse Advisory interpretation
Police AI decisions are being pushed toward testing under real operating conditions before scale-up, but most trials still launch without published metrics
A peer-reviewed Penn experiment shows facial recognition error rates change with blur, resolution and lighting, so a single benchmark figure can mislead investigators. The UK national Police AI lab proposes supplier testing against held-back trusted data before operational use. Chicago is putting gunshot detection through a six-month pilot with a final evaluation report before any award, and Berlin starts its AI video analysis with a four-week trial. The common gap: in Chicago and Berlin the success measures and data plans have not been published, and expansion or a forced contract award may arrive before results.
Operating questionBefore this system goes live or expands, what metrics, comparison data and image or environment conditions will the evaluation use, and who publishes the results?
Supporting evidenceUniversity of Pennsylvania (Department of Criminology; Wharton Statistics and Data Science; School of Engineering)Police AI (UK national policing AI body) / Humberside PoliceCity of Chicago Department of Procurement Services / Chicago Police DepartmentBerlin Police / Berlin Senate Department for the Interior
Courts are relying on existing signature duties and targeted declarations rather than blanket AI bans or disclosure rules
The England and Wales Civil Justice Council recommends declarations only where AI generated or influenced substantive content in witness statements and expert reports, and rejects general AI-use declarations for lawyers. A federal court in Ohio refused to bar self-represented plaintiffs from using AI despite hallucinated citations, holding Rule 11 is the safeguard and leaving any restriction to case management. Both treat access to justice as a reason not to ban tools, and both leave error detection to existing processes, which puts the checking burden on courts and opposing parties.
Operating questionWhich filing types in this court carry enough risk to justify an AI declaration, and who checks citations and evidence when filers rely on AI?
Supporting evidenceCivil Justice Council (England and Wales) working group on AI in court documentsU.S. District Court for the Northern District of Ohio
Evidence ledger
Full record · every source keeps its link and limitations
England and Wales Civil Justice Council final report on AI in court documents: a risk-based Practice Direction, AI declarations for witness statements and expert reports, but no general AI disclosure for lawyers
The Civil Justice Council published its final report on the use of AI to prepare court documents on Oct. 29, 2026. Its working group recommends that the Civil Procedure Rule Committee prepare a Practice Direction taking a functional, risk-based approach focused on whether AI has generated or influenced the substantive content of material submitted to the court. The report makes 15 recommendations: guidance for legal professionals should stay primarily with legal services regulators; the CJC and CPRC should urgently consider guidance for litigants in person; witnesses of fact should declare that AI used in preparing their statements has not generated or influenced their substantive content; and expert reports need a specific declaration, with enough information to scrutinise the evidence, where AI generated or influenced substantive content. It does not recommend general AI-use declarations by legal representatives or a requirement to name the tools used; representatives remain responsible under existing obligations for accuracy.
Why it matters, evidence and limitations
- Why it matters
- A senior court-rules advisory body has chosen targeted, content-based declarations over blanket AI disclosure, a model U.S. state courts weighing certification rules can compare against.
- Evidence and measured results
- Opened the Law Society Gazette report and read it in full. The judiciary.uk CJC announcement page returned an empty HTTP 202 response to curl and was read with the WebFetch tool to confirm the publication; a secondary summary (lawnews.co.uk) was also read. The report PDF itself was not located or opened.
- Limitations and uncertainty
- Recommendations only; the CPRC must still draft and consult on any Practice Direction. Civil courts in England and Wales, not criminal courts or U.S. jurisdictions. Read through press reporting; the full report text was not opened.
Federal court in Ohio refuses to ban self-represented plaintiffs from using generative AI or require AI disclosure, holding Rule 11 is the safeguard despite hallucinated citations, and warns of enforcement
In Boyd v. Beachwood City School Dist. Bd. of Educ., 2026 WL 3181031 (N.D. Ohio Oct. 23, 2026), as summarised and quoted by EDRM, defendants asked the court to prohibit the plaintiffs from using generative AI or require disclosure, pointing to fabricated case law, misrepresented holdings and fictitious quotations, including 12 hallucinations in one motion. The court acknowledged that the plaintiffs’ filings cited hallucinated case law and a hallucinated school-district policy and that these were Rule 11 violations, but declined to ban AI use, finding Rule 11 an appropriate safeguard, and exercised its discretion not to sanction at this time. It noted AI could help people who cannot afford counsel, expressly warned it would enforce Rule 11 going forward, and referred the case to a magistrate judge whose authority includes orders on the plaintiffs’ AI use, up to prohibiting it if necessary.
Why it matters, evidence and limitations
- Why it matters
- Shows a federal court choosing existing signature duties and case management over a categorical AI ban for self-represented parties, the same choice many state courts face as AI-drafted pro se filings rise.
- Evidence and measured results
- Opened the EDRM post and read it in full; it quotes the memorandum opinion with pinpoint citations. The opinion itself (Westlaw citation) was not opened.
- Limitations and uncertainty
- Single trial-court order read through a practitioner blog’s quotations; not binding precedent; the opinion was not opened; whether plaintiffs are self-represented is inferred from the court’s quoted discussion of pro se parties.
Peer-reviewed University of Pennsylvania experiment finds facial recognition error rates shift sharply with blur, low resolution and lighting in police-style images, with persistent race and gender disparities
Cuellar, To and Mehrotra generated 50,000 synthetic faces with StyleGAN3, labelled them with FairFace, and built race-and-gender-balanced samples (a 167-face control database and 167 treated probes, 83 target-present and 84 target-absent). They degraded probes by contrast, brightness, motion blur, pose and resolution and ran 1:n identification with DeepFace using ArcFace. Low resolution, high blur and contrast, and low brightness substantially increased the false negative identification rate (FNIR), with reduced resolution the largest effect. At the selected operating point the false positive identification rate (FPIR) was substantially higher than FNIR and highest near baseline image quality, generally falling as degradation increased. Disparities persisted: FPIR was often higher for Black individuals, especially Black women, while the group with the highest FNIR varied by degradation type. The authors argue FRT should not be summarised with one accuracy figure and should be evaluated on the low-quality images used in investigations.
Why it matters, evidence and limitations
- Why it matters
- Gives police agencies, prosecutors and courts a peer-reviewed basis for asking for image-quality-specific error rates before relying on a facial recognition lead, especially as courts begin to require disclosure of FRT use.
- Evidence and measured results
- Opened the PLOS One article and read the methods, results, discussion and limitations sections in full. Peer-reviewed journal article; data are deposited at Harvard Dataverse (not opened).
- Limitations and uncertainty
- Synthetic faces and controlled single-factor degradations, which may not capture real surveillance imagery; one open-source pipeline (DeepFace/ArcFace), not commercial police systems; small 167-face gallery; FairFace labels and simplified race categories; FFHQ-derived faces skewed white before balancing. U.S. academic study.
UK national Police AI lab proposes supplier benchmark testing against held-back trusted data before operational use, as Humberside describes knowledge-base-limited chatbot and staged CCTV analysis
At a panel reported by Emergency Services Times on Oct. 29, 2026, Police AI’s Lewis Lincoln Gordon said the body’s nascent AI lab is working on standards and testing so systems can be assessed in the context where they will be used; one proposal is an evaluation process in which suppliers test products against benchmarks, with trusted data held back to check performance, before operational use. Humberside Police’s Scott Dunn argued forces should invest in common foundations rather than "spending money 43 different ways." Near-term uses cited: interview transcription and summarisation, and staged retrospective CCTV analysis; real-time analytics would wait on reliable testing and safeguards. Humberside’s Bobbi chatbot is restricted to a pre-defined knowledge base and hands unanswered queries to a person. Speakers said deepfake detection tools are imperfect and data quality and cultural barriers to sharing remain major obstacles.
Why it matters, evidence and limitations
- Why it matters
- A national police body proposing held-back test data for supplier claims is a concrete assurance model that U.S. state police standards boards or consortia could adapt for multi-agency procurement.
- Evidence and measured results
- Opened via the WebFetch tool after curl returned HTTP 403, and read in full; trade-press write-up of a conference panel moderated by the author; no Police AI testing documentation was opened.
- Limitations and uncertainty
- Conference panel commentary, not a published standard or evaluation; the testing process is described as a proposal; no performance data; UK national context; the author moderated the panel.
Chicago plans a six-month gunshot-detection and violence-technology pilot in six South and West Side police districts with a final evaluation report, while police leaders say they know little about it
In an Oct. 20, 2026 letter, Chicago chief procurement officer Sharla Roberts told aldermen the city would launch pilots of "violence interruption and public safety technologies, including gunshot detection systems" in police districts 2, 5, 7, 9, 10 and 11, operational within about two weeks, saying procurement decisions should be informed by evidence. A city statement said the pilot will run six months, is separate from pending procurements, and that a final evaluation report and pilot data will be reviewed with user departments to decide next steps. At an Oct. 28 budget hearing, aldermen said there had been no CPD involvement, and acting Superintendent Fred Waller said he wished he knew more. The city ended ShotSpotter in September 2024, issued an RFP in early 2025 that drew nine bids, and has not awarded a contract; a council committee advanced an ordinance forcing an award within 30 days, and a non-binding ballot question goes to voters.
Why it matters, evidence and limitations
- Why it matters
- A large city is trying to put gunshot detection through a time-boxed, evaluated pilot before procurement, under political pressure to buy now, which tests how evidence-led police technology decisions hold up.
- Evidence and measured results
- Opened the Block Club Chicago article and read it in full; the procurement letter and city statement are quoted as reported; the pilot vendors, evaluation design and metrics were not disclosed and no city document was opened.
- Limitations and uncertainty
- Local news report of a letter and statement; pilot scope, vendors, costs and evaluation criteria undisclosed; no results yet; prior studies of ShotSpotter are referenced but not re-examined here.
Berlin police start a four-week trial of AI video analysis at Kottbusser Tor in November: 30 cameras flag fight-like movement for officer review, no facial recognition, as opponents challenge the law in court
Berlin police will begin AI-assisted video surveillance at Kottbusser Tor in early November 2026, with 30 cameras at 15 locations, starting with a four-week trial phase after installation and testing. According to the police project manager, the software reacts only to movement patterns typical of fights and other physical altercations; when it detects one, officers at the local station get an acoustic signal and live view, assess the situation themselves and dispatch colleagues. The report says biometric facial recognition is not possible, the system does not register skin colour, gender or clothing, and recordings are deleted after 30 days unless needed as evidence. Police recorded more than 900 offences in the area in 12 months, over 300 of them physical assaults. The capability comes from a reformed police law that the Left Party and Greens have challenged at Berlin’s constitutional court; the interior senator plans expansion to Warschauer Brücke, Alexanderplatz and Görlitzer Park.
Why it matters, evidence and limitations
- Why it matters
- A behaviour-only video analytics design with human assessment, short retention and no biometrics is a concrete reference for U.S. cities weighing AI-enabled camera networks and real-time crime centres.
- Evidence and measured results
- Opened the Berlin Spectator article and read it in full; it summarises reporting by Tagesspiegel and RBB; no police or Senate documents, evaluation plan or vendor details were opened.
- Limitations and uncertainty
- English-language local outlet summarising German media; no false-alert rates, evaluation design or results; non-U.S. legal context (German police law and data protection); expansion plans may change with coalition talks.
Colleges begin to "de-Flock": Emory removed campus plate readers and Cornell and Penn State are ending contracts, as a study counts at least 150 institutions using Flock and one college could not cancel
Inside Higher Ed reports that Emory University removed about a dozen Flock automated license plate readers in September and did not renew, Cornell did not renew, and Penn State plans to end its contract, amid student campaigns. A recent study identified at least 150 colleges and universities using Flock as of May, including large publics and two-year colleges; Flock declined to give a current count. Contract costs cited include $33,974 a year at Old Dominion University and $15,000 a year for five cameras at East Tennessee State. Elgin Community College tried to cancel after a student was detained by ICE, but Flock refused; staff covered the cameras with trash bags, the bags were removed, and Flock uninstalled the cameras when the contract expired in April. The University of Delaware says its campus police have strong policies limiting ALPR access, use, retention and sharing; students there want no replacement with other vendors such as Axon or Motorola Solutions. Flock has changed some default sharing settings.
Why it matters, evidence and limitations
- Why it matters
- Campus police are a large, little-noticed part of the ALPR market, and their exit problems (contract lock-in, data sharing defaults) mirror those facing cities.
- Evidence and measured results
- Opened the Inside Higher Ed article and read it in full; the cited study of institutional Flock use was not opened; contract figures are as reported.
- Limitations and uncertainty
- News feature centred on student activism; institution-level reasons for non-renewal partly as reported; underlying study not opened; no data on ALPR effectiveness on campus.
How to read this edition
Source findings, measured results and limitations come from the cited publications. Patterns, operating questions, role takeaways and implementation considerations are Lighthouse Advisory interpretation, stated as questions to validate locally rather than guaranteed outcomes. Vendor and operator claims are labeled as claims. Full research method.
- Independent reporting
- Independent reporting with attributable sources but without a formal evaluation design.
- Standards or public-body guidance
- Normative or advisory guidance from a standards body or public institution.
- Academic research
- Research produced through an academic institution or peer-reviewed venue.