Chapter 29 — Key Takeaways

A scannable one-page card. For the full argument, the figures, and the cold-case beat, see index.md.

The core claims

  • Hold the future to the same yardstick as the past. Every emerging method gets the NAS 2009 / PCAST 2016 question — what is the measured error rate, and how do we know? Newness, sophistication, and the words "artificial intelligence" are not evidence that a method works. They are reasons for more caution.
  • Rapid DNA = validated STR chemistry, automated. Swab-in/profile-out in ~90–120 minutes, no analyst. Excellent for clean single-source reference swabs (CODIS booking-station use, enabled by the Rapid DNA Act of 2017); dangerous the moment it is pushed onto degraded crime-scene mixtures, because automation removed the analyst who would have said "this is a mixture — stop." Validity depends on the use, not the box.
  • Microbial forensics / the necrobiome = promising, not court-ready. A microbial cousin of insect succession (Ch.13) for postmortem-interval estimation, plus trace association and bioterrorism attribution. The science (microbial sequencing) is real; the forensic validation (reproducible, condition-spanning studies; measured error rates) is, for most uses, incomplete. Interested, hopeful, unwilling to convict anyone on it yet.
  • Forensic isotope analysis = a region, not a name. Stable-isotope ratios (H/O for water, C/N for diet, strontium for geology) written into tissues that record different life stages (enamel = childhood; hair = recent months; bone = years). Rigorous chemistry, coarse geographic inference — an investigative lead to narrow a missing-persons search, never an identification.
  • AI/ML = real power, new perils. Already inside AFIS (Ch.14), probabilistic genotyping (Ch.9), digital triage (Ch.25). Three perils: the black box (unauditable reasoning), training-data bias (inherits the injustices of its data), automation bias (humans over-trust the machine). Plus a pervasive validation gap. Treat an algorithm's output as a lead to confirm, demand the measured subgroup error rate, and distrust any number whose basis you cannot audit. A vendor's "99% accurate" is marketing, not a validation study.
  • Familial searching ≠ investigative genetic genealogy. Both reach a suspect through relatives, but: familial searching = CODIS / STR markers / close kin (partial match → a parent/child/sibling already in the criminal database); IGG = consumer database / SNP markers / distant cousins (SNP profile → family-tree reconstruction). Both end identically: an investigative lead that a conventional, direct STR match (Ch.7) confirms. The new step points; the gold standard identifies.
  • §29.6 is the chapter's purpose: an "evaluate before you trust" gate for the future. The four themes, forged on past failures, must hold when pointed at methods that do not yet exist.

The method-validity verdict (NAS 2009 / PCAST 2016)

Method Core claim Validity verdict Honest verb
Rapid DNA (clean reference swab) This reference sample's STR profile Strong — inherits validated STR typing, automated; high inside its envelope "is the profile of this reference sample"
Rapid DNA (degraded crime-scene mixture) Identifies the source of a complex sample Off the spectrum — misapplied outside its validated envelope; analyst removed (should not be offered)
STR confirmation after an IGG/familial lead This person is the source Strong — gold-standard DNA does the identifying "matches, with RMP/LR" (Ch.7, 9)
Familial searching / IGG (the lead step) Points to a relative → a candidate Investigative, not a courtroom-ID method; judged by reliability of pointing + privacy cost "an investigative lead, to be confirmed"
Forensic isotope analysis Region of origin / life history Sound chemistry, coarse forensic inference; reference-data-limited "consistent with a region; a lead"
Microbial forensics (PMI / trace) Time since death / association Emerging — real science, incomplete forensic validation, no established error rate "consistent with (research-stage)"
AI/ML forensic tool Match / identification / score Depends entirely on independent validation, measured subgroup error rates, and auditable reasoning; an opaque unvalidated tool offered as proof of identity sits at the bottom "a lead to confirm — not an ID"

Where they sit: validated methods (STR confirmation; rapid DNA in-envelope) at the top; sound-instrument / modest-inference methods (isotopes) in the middle, strong only for what they honestly claim; emerging methods (microbial PMI) below, pending validation; and unvalidated black-box AI at the bottom with the discredited methods, no matter how advanced its architecture — because the spectrum cares about measured error, not cleverness.

What you can honestly say on the stand

  • Rapid DNA: "This instrument generated an STR profile from a single-source reference swab; that is its validated use. It is not validated for, and I would not rely on it for, a degraded or mixed crime-scene sample, which requires laboratory interpretation."
  • Isotopes: "The isotope ratios are consistent with an individual who spent childhood in this broad region and later years in another — an investigative lead to narrow the search. They do not identify the person or prove a specific town or route."
  • IGG / familial searching: "Genetic genealogy [or a familial database search] generated an investigative lead; the identification before this court is the conventional STR comparison between the crime-scene profile and the defendant's known sample."
  • AI output: "The system returned this candidate as a similarity-ranked lead. That score is not a probability of correctness, and absent independent validation with measured error rates I cannot offer it as an identification — only as a lead investigated by other means."
  • What you must NOT say: any rapid-DNA result on a complex sample as an identification; that isotopes "prove" a town or itinerary; that a genealogy "names" or "identifies" the suspect (it generates a lead); that an AI score is the probability the match is right; or that any method is reliable because it is new or sophisticated.

The cold-case line

Rapid DNA and IGG establish that the minor contributor to the gas-can mixture is not an unknown stranger — the "random intruder" theory is excluded. An exclusion, not an identification: it closes a door and narrows the field, but names no one. Status: "stranger theory excluded." The book's first theme in its purest form. One door closed is not a walk through it.

The themes this chapter advanced

  • Exclusion over proof (PRIMARY) — the cold case delivers an exclusion (stranger theory excluded), not a name; isotopes narrow to a region; IGG/familial searching generate leads. The future, like the past, excludes more reliably than it proves.
  • The validity spectrum (PRIMARY) — used prospectively: the same NAS/PCAST question sorts every emerging method; §29.6 turns the audit into a forward-looking gate; rapid DNA shows validity is a property of method-plus-sample-plus-question.
  • (Also advanced: cognitive bias — automation bias and bias laundering in AI, §29.4; the genealogy confirmation-bias cascade. And the CSI effect — the headline-driven hype of "rapid," "AI," and "breakthrough," the mirror image of the field's old overconfidence.)