Chapter 35 — Key Takeaways

The big idea

Artificial intelligence wears two faces in the forensic lab — a force multiplier that spends your scarcest resource, human attention, where it matters, and a synthesis engine that can fabricate the very evidence you are handed — and the same discipline governs both: AI output is a lead; the human examiner is the accountable author of every finding. A model emits a probability, not a fact — a number that correlates with truth on data resembling its training set and degrades, often silently, on data that does not. The defining problem of modern casework is volume, and AI solves it not by deciding but by prioritizing: it sorts the haystack so a person can judge the needles. Let a score flow straight into a report and you have laundered machine error into testimony. Technology can now write its own evidence, and still the method holds — image first, work the copy, corroborate across sources, document everything, report every finding with its limits intact.

The pipeline has a non-negotiable shape

The "how" of AI-assisted work is fixed regardless of which model you run:

• Read-only and copy-based. Models are hungry and stochastic and write temp files; they run on hashed working copies of a verified image, never the original. The original is sacred does not pause for convenience.
• Auditable. Every stage records the model, version, threshold, and inputs, so opposing counsel's expert can replay it. A result you cannot reproduce is a result you cannot defend.
• A single human gate. Data reduction, enrichment, and ranking produce suggestions; the examiner's review is the only place a finding is born. Skip that gate and the work stops being forensic.

The hashing spectrum — know which rung you are on

Robust hashing finds known material only — PhotoDNA-style matching never classifies an unknown image, and novel AI-generated content will not match a hash list at all.

Triage is half a dozen tools, each with its own failure mode

Known-file filtering (drop the haystack first), CNN classification and object detection, embeddings for clustering and near-duplicate collapse, OCR/ASR/translation, named-entity recognition with link analysis, and unsupervised anomaly detection on logs and timelines. Each is a force multiplier and each fails differently — OCR mis-reads digits, ASR invents words, translation flattens meaning, NER mislabels, co-occurrence is not collusion, an anomaly is a deviation not a crime. You use them to decide where to read the real evidence, then prove the point with corroborated artifacts.

Detection is losing an arms race — so prove the real

There is no button that prints "fake," only a battery of independent tests and a discipline of convergence: blend boundaries, lighting and eye-reflection inconsistencies, blink/pupil/teeth tells, temporal flicker, frequency-domain GAN fingerprints, the missing rPPG heartbeat; in audio, spectral artifacts, flat prosody, absent breathing and room acoustics; in biometrics, head-pose and identity drift. Every tell is a lead, and the headline reliability fact is generalization: detectors collapse on unseen generators (the Deepfake Detection Challenge fell from the low-80s to ~65%). Because generators improve faster than detectors, the field's center of gravity shifts from catch the fake to prove the real — C2PA Content Credentials, watermarking (SynthID), and known-synthetic hashing. Read provenance like EXIF: a valid credential corroborates history, a failed one signals tampering, and an absence is silent — it neither indicts nor exonerates.

In court: the liar's dividend and the Daubert problem

The liar's dividend requires no fake at all — the mere existence of deepfakes lets the guilty wave away genuine evidence. The answer is not a better detector but authentication under FRE 901: chain of custody, hash verification, container/encoder metadata, provenance, corroborating timeline, and a detection battery whose convergence and stated limits you can defend. Meanwhile every AI tool is a method the court may scrutinize under Daubert/FRE 702, where AI strains all five factors — unknown out-of-distribution error rate, black-box explainability (SHAP/LIME/Grad-CAM only approximate), documented bias (NISTIR 8280 found order-of-magnitude demographic differences), poor reproducibility, and trade-secret inaccessibility. Pin the model, version, threshold, validation set, and measured error rate — reproducibility is admissibility.

You can now…

• ☐ Build an auditable, read-only, copy-based AI triage pipeline — known-file filtering and the hashing spectrum, classification, clustering, OCR/ASR/translation, entity extraction, anomaly detection — that prioritizes human attention and never replaces the human gate.
• ☐ Distinguish cryptographic, fuzzy, and perceptual hashing, and handle robust-hash matching of known illegal material clinically — its procedure, mandatory-reporting duties, and limits (it never classifies the unknown).
• ☐ Run a deepfake detection battery across visual, audio, and biometric artifacts and report each result as a corroborated lead, stating plainly that detectors generalize poorly to unseen generators.
• ☐ Read provenance correctly — C2PA Content Credentials, watermarking, known-synthetic hashing — and authenticate contested media under FRE 901 against the liar's dividend.
• ☐ Articulate the Daubert/FRE 702 reliability, bias, explainability, and reproducibility problems of AI tools and pin a model-and-version manifest that makes machine-assisted work admissible.

Looking ahead

Chapter 36 — The Forensic Toolkit. Step back from any single technique to the instruments themselves — Autopsy and The Sleuth Kit, FTK and EnCase, Cellebrite, Volatility, Wireshark — how to choose, validate, and combine open-source and commercial tools, weighing the AI features now embedded in every major suite against exactly the reliability standard this chapter set. Technology changes; principles don't.

One sentence to carry forward: A machine can sort a terabyte and forge a face in the same afternoon — but only a human who names the tool, states its limits, and points to the artifact turns a score into a finding; the score is never the finding.