The Longcat Gazette · Special Report
Criminal Defense Desk · September 2026 · Fully cited · 39 sources

The Algorithm on Trial: How AI, Geofence Warrants, and Forensic Reform Are Reshaping Criminal Defense

**A system built on hair microscopy, secret DNA algorithms, and digital dragnets is being challenged from courthouses to statehouses — and defense attorneys are beginning to fight back with the same tools prosecutors used to wield.** [3][4]

1. The Geofence Era Ends at the Supreme Court

For nearly a decade, prosecutors used "geofence warrants" — bulk-data requests that compelled Google to hand over location data for every cellphone near a crime scene — as a routine investigative tool. [3][25] In June 2026, the Supreme Court drew a line. In *Chatrie v. United States*, the justices ruled 6-3 that using a geofence warrant to obtain evidence against a Virginia man accused of a 2019 bank robbery constituted a "search" under the Fourth Amendment. [22][26] Justice Elena Kagan, writing for the majority, held that individuals retain a reasonable expectation of privacy in records of their cellphone's location, and that police intrude on that interest even when the data comes from a third-party tech company and covers only a limited time. [22] The decision did not ban geofence warrants outright; it sent the case back to the lower courts to determine whether the search was "reasonable." [24] But the ruling reframed the legal landscape: location data is no longer a free-for-all for law enforcement. [22][26]

The *Chatrie* decision landed alongside another structural change. In December 2023, Google announced it would shift Location History storage from its servers to users' devices — a move that effectively ended the company's role as the primary target of geofence warrants. [41][42] As TechCrunch reported at the time, the use of geofence warrants had "exploded" in the years since the practice was first revealed in 2019. [41] Google's policy change, combined with the Court's constitutional ruling, has forced prosecutors to seek more targeted investigative tools. [41][42][43]

2. AI Comes to Public Defense — Ethically

The same generative AI that prosecutors have used to analyze evidence and draft motions is now being deployed on the other side of the aisle — but with guardrails. In July 2026, the National Association of Criminal Defense Lawyers released *Parity in Practice: The Defender's Duty to Ethically Use AI*, a white paper offering defense attorneys a roadmap for putting generative AI to work "thoughtfully, ethically, and to the benefit of their clients." [40][21] The paper argues that AI parity is not optional: if prosecutors use AI to build cases, defenders must use AI to test them, or risk violating their ethical obligation to provide competent representation. [39][21]

The NACDL's framework arrives as defenders across the country experiment with the technology. In Miami-Dade County, the Public Defender's office has integrated AI for legal research and case preparation. [1] In California, Berkeley Law's Criminal Law and Justice Center has launched an AI-for-Public-Defenders initiative. [2] A 2025 arXiv study by Inyoung Cheong, Patty Liu, Dominik Stammbach, and Peter Henderson — based on interviews with 17 public defense professionals — mapped the daily workflow of defenders into five pillars (evidence investigation, legal research, document review, client communication, and courtroom advocacy) to identify where AI can genuinely help and where it risks harm. [60]

3. Bodycam Transcription: AI as Force Multiplier

One of the most concrete AI-in-defense success stories involves body-worn camera footage. Public defenders routinely face hundreds of hours of video per case; reviewing it all is mathematically impossible given existing staffing. [15] JusticeText, a legal tech company, rolled out AI-powered bodycam transcription for the Kentucky Department of Public Advocacy in May 2026. [62] The Kentucky DPA, established in 1972, handles more than 140,000 cases annually with a team of nearly 400 attorneys. [62] The partnership allows defenders to search, clip, and timestamp footage rather than watch it linearly — a task-level intervention that the Cheong et al. study identified as ripe for AI assistance. [60][62]

4. When AI Meets the Courtroom: The Puloka Precedent

Not all AI evidence survives contact with a jury. In *State of Washington v. Puloka* (2024), a Washington state superior court judge rejected the admission of video exhibits "enhanced by artificial intelligence" for use in a jury trial — an apparent issue of first impression. [49] The defendant, Joshua Puloka, was charged with three counts of murder from a 2021 shooting. The unaltered 10-second source video had already been entered into evidence. The defense sought to admit an AI-enhanced version, arguing that the original was low-resolution, motion-blurred, and "blocky." [49] The judge's refusal to admit the enhanced video signaled judicial skepticism toward AI-modified evidence — a precedent prosecutors and defenders alike are watching closely as generative AI tools become cheaper and more accessible. [49]

5. The Liar's Dividend: Deepfakes and the Burden of Proof

The flip side of AI-generated evidence is AI-denial. The "liar's dividend" — a term coined by Robert Chesney and Danielle Citron — describes the phenomenon in which bad actors exploit growing awareness of deepfakes by claiming authentic evidence was AI-generated. [61] As one scholar put it, "as education about the risks of AI-generated content grows, it becomes easier for bad actors to leverage that skepticism and claim that authentic evidence was created by AI." [61] For defense attorneys, the liar's dividend cuts both ways: it can be used to challenge genuine prosecution evidence, but it also makes it harder to introduce legitimate digital evidence that might be dismissed as synthetic. [61]

Congress has moved to address AI-generated child sexual abuse material through the TAKE IT DOWN Act, which criminalizes the non-consensual publication of intimate images — including AI-generated depictions. [30][31] But the liar's dividend problem extends well beyond CSAM into the heart of criminal adjudication: when any video, audio, or photograph can be plausibly denied as a deepfake, the epistemological foundations of evidence themselves are shaken. [61]

6. Confrontation and the Forensic Analyst: *Smith v. Arizona*

The Supreme Court's 2024 decision in *Smith v. Arizona* clarified a critical Sixth Amendment question: when an expert witness conveys an absent lab analyst's statements, and those statements support the expert's opinion only if they are true, the statements come into evidence for their truth and thus implicate the Confrontation Clause. [36][37] The 9-0 ruling, authored by Justice Kagan, means that prosecutors cannot circumvent a defendant's right to confront witnesses by having one analyst vouch for another's lab work without producing the original analyst for cross-examination. [36] The decision reverberates through forensic disciplines — from DNA to drug chemistry to fingerprint analysis — where surrogate testimony has been common. [36][37]

7. Probabilistic Genotyping: The Secret Algorithm Problem

Beneath the surface of modern DNA evidence lies a controversy over proprietary software. Probabilistic genotyping (PG) programs — such as TrueAllele and STRmix — use complex statistical models to analyze DNA mixtures that would otherwise be uninterpretable. [6] The problem: the source code is often shielded as a trade secret, preventing defense experts from scrutinizing the algorithms that helped secure a conviction. [6][7] A Criminal Legal News investigation in 2025 documented how PG software "promises to unravel the chaos of complex DNA mixtures" but operates in secrecy, raising the question: can defendants challenge evidence derived from algorithms they are not allowed to see? [6] The National Institute of Justice has funded research into the uncertainty inherent in probabilistic genotyping, but admissibility standards vary wildly by jurisdiction. [7][6]

8. The FBI Hair-Testing Scandal: A Cautionary Tale

The limits of forensic science are not theoretical. In a stunning admission, the FBI acknowledged that its microscopic hair analysis testimony contained errors in at least 90 percent of cases under ongoing review. [52] A National Registry of Exonerations report documented how hair comparison evidence — once presented as near-certain identification — contributed to wrongful convictions across the country. [53] The hair-analysis scandal predates the AI era but offers a template for understanding how forensic "science" gains institutional authority before its error rates are properly understood. [10][11]

9. Exonerations in 2024: 147 and Counting

The National Registry of Exonerations documented 147 exonerations in 2024, a figure that "underscores both the resilience of those wrongfully convicted and the systemic failures that led to their imprisonment." [5] The cases reveal patterns of official misconduct, racial disparities, and the critical role of specialized organizations. [5][27] Meanwhile, the Innocence Project reports that its clients have collectively spent 4,102 years wrongfully incarcerated; 205 clients have been exonerated by DNA evidence. [29] Conviction Integrity Units — specialized prosecutor-office divisions that review past convictions — have expanded, with the NRE tracking CIUs across the country. [54]

10. Risk Assessment Under Scrutiny

The First Step Act's PATTERN risk-assessment tool, used to classify federal prisoners for recidivism risk, underwent a mandated 2023 revalidation by the National Institute of Justice. [8][9] Critics argue that risk-assessment instruments encode racial bias. A University of Michigan Ford School study found that pretrial risk-assessment tools are "subjective and biased," potentially reproducing the disparities they claim to eliminate. [38] The Department of Justice continues to refine PATTERN, but the debate over whether algorithmic neutrality is achievable — or even coherent — remains unresolved. [8][9][38]

11. California's Racial Justice Act: Promise and Frustration

In 2020, California enacted the Racial Justice Act (RJA), allowing defendants to challenge racial bias at multiple stages of prosecution. Four years later, the law has produced only a handful of successful motions. [46] CalMatters reported in November 2024 that a "Garrison Project and CalMatters analysis of the Racial Justice Act found about a dozen successful cases in which judges took action on racial bias in the legal process." [46] Stanford Law School researchers have since developed a statistical toolkit to help defenders bring RJA claims grounded in empirical disparities rather than anecdotal evidence. [47][48]

What to Watch

The next frontier is integration. As AI tools become standard in both prosecution and defense, the defense bar's challenge is not access to technology but literacy, ethics, and quality control. [21][39] The NACDL's *Parity in Practice* framework is a start, but implementation will require funding, training, and a willingness to treat AI as an instrument of zealous advocacy rather than a cost-cutting substitute. [40][39] Geofence warrants are constrained but not dead. [22][41] Forensic reform is piecemeal. [10][11] And the liar's dividend is only beginning to reshape how juries evaluate evidence. [61] The algorithm is on trial — and the verdict is still out.

Sources

  1. Florida Bar: Miami-Dade PD using AIhttps://www.floridabar.org/the-florida-bar-news/miami-dade-public-defender-is-using-artificial-intelligence-for-research-and-for-case-preparation
  2. Berkeley Law AI in Criminal Defense portalhttps://www.law.berkeley.edu/research/criminal-law-and-justice-center/our-work/ai-for-public-defenders
  3. ABA: Are Geofence Warrants Headed for Extinction?https://www.americanbar.org/groups/criminal_justice/resources/magazine/2024-summer/geofence-warrants-headed-extinction
  4. Criminal Legal News: Chatrie reverse-location surveillancehttps://www.criminallegalnews.org/news/2026/may/1/beyond-geofence-warrants-chatrie-reverse-location-surveillance-and-the-fourth-amendment-fight-over-digital-dragnets
  5. Criminal Legal News: 150 exonerations 2024https://www.criminallegalnews.org/news/2025/apr/15/nearly-150-exonerations-2024-highlight-persistent-flaws-us-criminal-justice-system
  6. Criminal Legal News: Probabilistic genotyping on trialhttps://www.criminallegalnews.org/news/2025/aug/1/probabilistic-genotyping-trial-can-we-trust-secret-algorithms-deciding-guilt
  7. Thompson 2023: Uncertainty in probabilistic genotyping (JFS)https://onlinelibrary.wiley.com/doi/10.1111/1556-4029.15225
  8. NIJ 2023 PATTERN revalidationhttps://nij.ojp.gov/library/publications/2023-review-and-revalidation-first-step-act-risk-assessment-tool
  9. NACDL on PATTERNhttps://www.nacdl.org/Media/PATTERN-First-Step-Act-Risk-Assessment-Tool
  10. ProPublica: Is It Forensics or Junk Science?https://www.propublica.org/article/understanding-junk-science-forensics-criminal-justice
  11. Criminal Legal News: bite marks Louisiana junk sciencehttps://www.criminallegalnews.org/news/2025/mar/15/bite-marks-and-broken-justice-louisiana-mans-life-and-death-struggle-against-junk-science
  12. RCFP: bodycam video access litigationhttps://www.rcfp.org/bodycam-videos-local-legal
  13. NACDL AI Task Forcehttps://www.nacdl.org/Landing/ai
  14. Chatrie v. United States (SCOTUS 2026)https://www.supremecourt.gov/opinions/25pdf/25-112_0am4.pdf
  15. State of Surveillance: Chatrie SCOTUS rulinghttps://stateofsurveillance.org/news/chatrie-supreme-court-ruling-geofence-warrant-fourth-amendment-search-2026
  16. CLN: geofence warrants eliminated by creatorhttps://www.criminallegalnews.org/news/2024/feb/15/geofence-warrants-mass-location-surveillance-and-privacy-threat-created-google-may-be-eliminated-its-creator
  17. SCOTUSblog: Chatrie geofence search rulinghttps://www.scotusblog.com/2026/06/court-rules-that-law-enforcements-use-of-geofence-warrant-was-a-search
  18. NRE 2024 Annual Reporthttps://exonerationregistry.org/sites/exonerationregistry.org/files/documents/2024_Annual_Report.pdf
  19. Innocence Project: exonerations datahttps://innocenceproject.org/exonerations-data
  20. Wikipedia: TAKE IT DOWN Acthttps://en.wikipedia.org/wiki/TAKE_IT_DOWN_Act
  21. CRS: TAKE IT DOWN Act summaryhttps://www.congress.gov/crs-product/LSB11314
  22. Smith v. Arizona (SCOTUS 2024)https://www.supremecourt.gov/opinions/23pdf/22-899_97be.pdf
  23. SCOTUSblog: Smith v. Arizonahttps://www.scotusblog.com/cases/smith-v-arizona
  24. U Michigan Ford School: pretrial risk tools biasedhttps://fordschool.umich.edu/news/2023/pretrial-risk-assessment-tools-found-be-subjective-and-biased
  25. NACDL Parity in Practice white paper (2026)https://www.nacdl.org/getattachment/e0a68ac8-31cf-4a86-9fb9-917570634f60/parity-in-practice-the-defenders-duty-to-ethically-use-ai.pdf
  26. NACDL news release: Parity in Practicehttps://www.nacdl.org/newsrelease/News-Release-~-Parity-in-Practice
  27. TechCrunch: Google moves to end geofence warrantshttps://techcrunch.com/2023/12/16/google-geofence-warrants-law-enforcement-privacy
  28. EFF: Is This the End of Geofence Warrants?https://www.eff.org/deeplinks/2023/12/end-geofence-warrants
  29. Bloomberg Law: Google location data move reshapes geofence warrantshttps://news.bloomberglaw.com/privacy-and-data-security/googles-location-data-move-will-reshape-geofence-warrant-use
  30. CalMatters: California Racial Justice Act few successeshttps://calmatters.org/justice/2024/11/california-racial-justice-act
  31. Stanford: RJA data pathwayshttps://law.stanford.edu/2025/04/15/data-disparities-and-discrimination-how-californias-racial-justice-act-creates-new-pathways-to-challenge-and-evaluate-racial-bias
  32. Stanford RJA statistical toolkithttps://law.stanford.edu/wp-content/uploads/2025/04/Statistical-Claims-Under-the-California-Racial-Justice-Act-Toolkit.pdf
  33. Greenberg Traurig: State v. Puloka AI video rejectedhttps://www.gtlaw.com/en/insights/2024/5/washington-court-rejects-novel-use-of-ai-enhanced-video-in-trial
  34. FBI: hair testimony errors in at least 90% of caseshttps://www.fbi.gov/news/press-releases/press-releases/fbi-testimony-on-microscopic-hair-analysis-contained-errors-in-at-least-90-percent-of-cases-in-ongoing-review
  35. NRE: Microscopic Hair Analysis Comparison Reporthttps://exonerationregistry.org/sites/exonerationregistry.org/files/documents/NREReportMHCAv1.3.pdf
  36. NRE: Conviction Integrity Units trackerhttps://exonerationregistry.org/conviction-integrity-units
  37. How Can AI Augment Access to Justice? Public Defenders (arXiv)https://arxiv.org/abs/2510.22933
  38. Deepfakes and the Liar's Dividend - UST Law Journalhttps://ustlawjournal.blog/2025/10/08/ais-lurking-danger-deepfakes-and-the-liars-dividend
  39. JusticeText: AI bodycam transcription for Kentucky DPAhttps://www.justicetext.com/blog/kentucky-dpa