This chapter is a synthesis. Chapters 25 through 31 have analyzed each algorithmic system in depth — the legal frameworks that govern it, the constitutional arguments it generates, and the specific deficiencies that make its outputs contestable. This chapter assembles those tools into a practical sequence for the lawyer who walks into court knowing that algorithmic evidence is in the case but needing to know where to start, what to ask for, and how to frame the challenge.
The sequence is deliberately organized by procedural stage. Algorithmic evidence is not challenged most effectively at trial. It is challenged earliest — at intake, at arraignment, at bail, in discovery — and the trial argument is built from the record created at those earlier stages. A lawyer who first confronts the algorithmic question during closing argument has already lost the best opportunities.
I. Stage one: intake audit — identify every algorithmic system in the case
Brady v. Maryland, 373 U.S. 83 (1963) Johnson v. State, 2025 WL 2237582 (Appellate Court of Maryland, August 6, 2025) (unreported)
The first task, at intake, is identification. Defense counsel should review the charging documents, police reports, and any available discovery not for the names of witnesses but for the presence of algorithmic systems — any reference to facial recognition, risk scoring, predictive analysis, acoustic detection, data fusion, watchlisting, or automated flagging. These systems are often described in police reports in language that obscures their nature: “the database indicated,” “the system returned a match,” “the assessment identified.” Each of those phrases is a marker of an algorithmic output that requires investigation.
The systems to look for, drawn from the preceding chapters, include: facial recognition tools such as Clearview AI or any agency-run biometric matching system (Chapter 27); gunshot detection alerts from ShotSpotter/SoundThinking (Chapter 28); risk assessment scores such as COMPAS at sentencing or the PSA at bail (Chapters 25 and 26); predictive policing outputs from any place-based or person-based analytical tool (Chapter 28); surveillance-derived investigative leads from FISA Section 702 collection or related data fusion systems (Chapters 17 and 30); and text analytics flags from systems such as Pangaea Text or ATLAS in immigration-adjacent criminal proceedings (Chapter 20).
Once the systems are identified, the Brady disclosure demand should go out immediately — before arraignment if possible. As established in Chapter 31 through Johnson v. State, 2025 WL 2237582, the Brady obligation in the algorithmic context requires disclosure of: the specific software or system used, including the version; the complete output, not just the top result; the confidence or similarity score associated with the match; the quality and provenance of the probe input; and documentation of the human review steps taken before the output was acted upon. The Johnson court reversed a conviction specifically because the State could not identify the facial recognition software it had used and produced only a partial output. That level of disclosure failure is itself a Brady violation — make the demand specific enough that any gap in compliance is documentable.
II. Stage two: the records blitz — attack the inputs before the score
Privacy Act of 1974, 5 U.S.C. § 552a(d) Law Enforcement Directive, Directive (EU) 2016/680, Articles 16 and 17 Mathews v. Eldridge, 424 U.S. 319 (1976)
Every algorithmic system examined in this module produces outputs that are only as reliable as their inputs. As Chapter 22 established in the immigration context and Chapter 31 applied to criminal proceedings, rectification of inaccurate inputs is often the most effective and least costly challenge available — more effective than a constitutional argument about the algorithm’s architecture, because it defeats the specific output in this case rather than attacking the system in general terms.
The records request should identify every database that may have contributed to the algorithmic output: criminal history records from state and federal repositories; federal agency records under Privacy Act, 5 U.S.C. § 552a, including FBI Integrated Automated Fingerprint Identification System records, DHS biometric databases, and any relevant System of Records Notice for the agency involved; and, in the EU context, records held by competent authorities subject to the Law Enforcement Directive, Directive (EU) 2016/680.
Common errors to look for, drawn from the case patterns in preceding chapters: a dismissed arrest still coded as a conviction in the criminal history that feeds the PSA’s prior-conviction factor (Chapter 26); an incorrectly coded violent offense flag that drives the NVCA component of the PSA or the COMPAS score (Chapter 25); an identity database error attributing another person’s record to the client, of the type that produced the wrongful arrests documented in Chapter 27; and a prior failure to appear attributed to the client based on a court-system data entry error rather than actual non-appearance.
Each confirmed error should be formally challenged in writing, creating a record of the error and the agency’s response. Under the Mathews v. Eldridge framework established in Chapter 12, the client’s interest in accurate inputs at the pretrial stage is substantial — the consequence of an inaccurate PSA input is pretrial detention; the probability of error in administrative criminal history records is documented and non-trivial; and the corrective procedure — checking the inputs against source records — is straightforward. That analysis supports a procedural due process argument for disclosure of the scoring worksheet and an opportunity to contest inputs before the bail hearing concludes.
III. Stage three: the motion to compel — source code, training data, and validation studies
U.S. Constitution, Amendment VI — Confrontation Clause Ake v. Oklahoma, 470 U.S. 68 (1985) Daubert v. Merrell Dow Pharmaceuticals, Inc., 509 U.S. 579 (1993)
Where the algorithmic output is proprietary — the system’s architecture, training data, or weighting methodology protected as a trade secret — the motion to compel is the vehicle for testing whether the prosecution can simultaneously rely on the system and withhold the information necessary to challenge it.
As established in Chapter 29, the Sixth Amendment confrontation argument and the Fifth Amendment due process argument from Ake v. Oklahoma, 470 U.S. 68 (1985), both support disclosure. The confrontation argument is that meaningful challenge to the evidence requires access to enough information about the methodology to test whether it was reliably applied. The Ake argument is that where the state’s case rests on evidence requiring expert interpretation, the defendant has a due process right to access expert assistance — and where the expert cannot examine the methodology, the assistance is not meaningful.
The motion should request specifically: the name, version, and developer of the system; model documentation describing its intended use, known limitations, and demographic performance differentials; training data documentation, including the demographic composition of the training dataset and the geographic scope of the validation study; the full ranked output with confidence or similarity scores for all candidates generated; audit logs showing how the system operated at the time the evidence was generated; and any internal error or performance reports generated by the agency or the vendor after deployment.
Where trade secret protection is asserted, the defense should propose a protective order permitting defense-retained expert examination under confidentiality restrictions — the arrangement courts have increasingly accepted in proprietary forensic tool litigation. If the vendor refuses even limited disclosure under protective order, press the exclusion argument: the prosecution cannot use a proprietary tool to generate evidence and deny the defense the information needed to test it. State v. Puloka, King County Superior Court, Washington, 2024, provides a recent example of a trial court that applied simultaneously Frye, Federal Rule of Evidence 702, Rule 401, and Rule 403 to AI-enhanced video evidence — the judicial willingness to impose multiple, cumulative reliability burdens on algorithmic and AI-processed evidence is increasing, not decreasing. That trend benefits the defense.
IV. Stage four: the expert — finding and deploying technical testimony
Federal Rules of Evidence, Rule 702 Proposed Federal Rule of Evidence 707 (public comment closed February 16, 2026)
Algorithmic evidence involves methodological complexity that cannot be communicated to a jury without expert assistance. As established in Chapter 29 through the Advisory Committee’s reasoning in developing Proposed Rule 707, the reliability standards that govern algorithmic evidence require specialized knowledge to apply. Defense counsel should engage a technical expert with specific qualifications relevant to the system at issue — not a general AI commentator but a researcher or practitioner with documented knowledge of the specific type of system: facial recognition accuracy under real-world conditions, gunshot detection false-positive rates, risk assessment tool validation methodology, or machine learning bias analysis.
The expert’s mandate should cover three areas. First, an assessment of whether the system satisfies the Daubert reliability factors — testability, peer review, known error rate, and general acceptance within the relevant scientific community — in light of the specific conditions of its use in this case. Second, an analysis of the demographic performance data for the specific system, with emphasis on whether the NIST FRVT findings documented in Chapter 27 or the equivalent data for other systems are consistent with, or worse than, the claimed accuracy rates. Third, a documentation of any specific deficiency in the input data, search parameters, confidence threshold, or human review process that would make the particular output in this case unreliable independent of the system’s general validity.
That third category — specific deficiency — is often the most powerful argument. A technically sophisticated court may be persuaded that facial recognition is generally reliable and still be persuaded that this facial recognition match, from this probe image quality, at this confidence threshold, with this human review record, is not reliable enough to go to the jury. The expert provides the technical foundation; the lawyer provides the legal argument for exclusion under Daubert or Rule 403.
V. Stage five: the courtroom — framing the argument for judge and jury
The automation bias problem documented in Chapter 26 — where 79 percent of judges reported that the PSA “always” or “often” informs their decisions — operates in the courtroom as well as the bail hearing. Jurors who have no technical background and no reason to distrust computer-generated results may treat algorithmic output as objective truth. The defense lawyer’s task is to reframe that perception without appearing to oppose technology in general, which courts and juries will resist.
The most effective framing treats the algorithm as a witness the prosecution has chosen not to call and cannot call. The system generated an output. The output was produced by a methodology the defense cannot examine. The defense has asked to examine it and been denied or given incomplete information. The jury is being asked to rely on a conclusion without being permitted to evaluate how that conclusion was reached. That framing resonates with the confrontation principle in terms a lay jury can understand — it is not a technical argument, it is a fairness argument.
For cross-examination of the witness who used or interpreted the algorithmic output — whether an investigating officer, a forensic analyst, or a risk assessment supervisor — the key questions are not technical. They are procedural. Did you know the confidence score associated with the match? Did you review the full ranked list of candidates, or only the top result? Did you review the probe image quality assessment? Did you review the system’s documented error rate for images of this quality and this demographic? Did you verify the identification through any independent method before taking action? If the answers are negative — and in most documented cases they are — the cross-examination establishes that the officer treated a probabilistic hypothesis as a confirmed identification, which is the core process failure in every wrongful arrest case examined in Chapter 27.
The closing argument synthesizes both layers: the procedural failure by the officer who acted on the output without adequate verification, and the methodological unreliability of the output itself. Together, they support reasonable doubt not through abstract constitutional argument but through a concrete demonstration that the identification process the prosecution is asking the jury to trust failed at each step where human judgment was available to catch the machine’s error.
VI. Preserve the record for appeal — every argument, every stage
Nothing in this chapter is worth building if the arguments are not preserved. The constitutional issues — Fourth Amendment, Fifth Amendment, Sixth Amendment, Fourteenth Amendment — must appear in writing, in timely filed motions, with specific factual and legal grounding, before trial. Arguments raised for the first time in closing argument or on appeal are waived or subject to plain error review, which is a substantially higher standard. The record must show that the defense identified the algorithmic system, demanded disclosure, challenged the inputs, moved to compel additional disclosure where warranted, sought expert assistance, and raised each constitutional argument before the court that had the power to remedy it.
That record is also the foundation for Schlup v. Delo, 513 U.S. 298 (1995), if the wrongful conviction argument must be made post-conviction. As established in Chapter 31, the Schlup gateway requires that the new evidence, considered in light of all the evidence, make it more likely than not that no reasonable juror would have convicted. The pre-trial record — the disclosure failures, the input errors, the confidence score deficiencies, the absence of independent verification — is part of that evidence. It is built before trial and preserved through appeal.
Next: Module 7 — The transatlantic comparison: two legal systems, one algorithmic challenge.
Leave a Reply