Fargo, West Fargo & Cass CountyCriminal defense informationCall 701-929-4886
Fargo and North Dakota criminal defense guide

AI-Generated Evidence in Criminal Cases

Artificial intelligence can search, classify, transcribe, translate, summarize, predict, and generate. Those functions are not interchangeable. An AI…

ElementsPenaltiesEvidenceNext steps

Police AI summaries, automated transcripts, translations, image and object labels, facial-recognition leads, report drafting, generative reconstructions, chatbot records, risk scores, prompts, model versions, source grounding, confabulation, human verification, discovery, experts, authentication, hearsay, confrontation, demonstratives, and reasonable doubt

AI-Generated Evidence in Criminal Cases

Artificial intelligence can search, classify, transcribe, translate, summarize, predict, and generate. Those functions are not interchangeable. An AI output may be an investigative lead, a human witness’s work aid, an expert’s disclosed tool, a demonstrative, a provider business record, or an exhibit offered for its truth. Before a court relies on it, counsel should identify the source data, exact model and version, prompt and settings, retrieval material, complete output, human edits, validation, known error, and the person who accepts responsibility for the final assertion.

An AI output is not a witness and confidence is not personal knowledge. Preserve the original evidence before AI processing, the complete input and output, system and model version, prompt, settings, account and audit logs, retrieved sources, intermediate steps available to the user, edits, rejected outputs, human verification, and every derivative exhibit.

AI can assist an investigation without becoming admissible proof

Investigative leads and trial evidence are different

Police may use a system to flag an image, rank a face, identify an object, cluster records, translate a message, or suggest a link. Officers can investigate that lead through admissible evidence. If the State offers only the independently obtained evidence, the AI may still matter to discovery, warrant candor, or bias, but it is not automatically a trial exhibit. If the State offers the output or an opinion based on it, the system’s reliability and foundation move to the foreground.

Automated transcription converts signal into proposed words

Speech-to-text systems can create body-camera, interview, jail-call, 911, and surveillance transcripts. Performance varies with noise, overlap, dialect, names, code words, language, compression, and speaker distance. A polished sentence may conceal a low-confidence word or an inaudible segment. The native audio controls. Preserve timestamps, speaker segmentation, confidence fields, model, language setting, custom vocabulary, human corrections, and earlier transcript versions.

Machine translation adds a second inference layer

An automated transcript of non-English speech can be wrong before translation begins; the translation then introduces ambiguity, idiom, grammar, and context choices. A bilingual reviewer should work from the audio when possible, not merely translate the machine transcript. Identify dialect, speaker, disputed terms, alternative translations, and reviewer qualifications. A model’s fluent English can make uncertainty disappear without making the source clearer.

Generative summaries can omit, merge, or invent

A tool may summarize reports, calls, messages, medical records, or thousands of documents. It can compress repetitive material usefully, but it may combine speakers, misstate dates, overlook qualifications, convert allegations into facts, or supply content not in the sources. NIST calls confidently generated false content “confabulation” in its Generative AI Profile, NIST AI 600-1. Every important assertion should be traced to the underlying record.

Computer vision labels are not observations by a person

Systems may label “gun,” “person,” “fight,” “license plate,” “blood,” or “vehicle,” track objects, estimate pose, or select key frames. Output depends on training categories, threshold, image quality, angle, occlusion, lighting, compression, and deployment population. A box around an object does not prove the legal identity of the item. The original frames, complete video, alternate classes, scores, threshold, and human review are necessary.

Facial recognition returns candidates, not identification

A face-search system compares a probe image with a gallery and returns ranked candidates. Gallery composition, image quality, algorithm, threshold, demographic performance, and human review affect the result. A candidate lead should be corroborated independently. If an officer later identifies the defendant, discovery should show whether the officer saw the algorithm’s candidate first because that exposure can influence the supposedly independent identification.

AI can draft reports that appear to be human memory

An officer may use AI to convert notes or audio into a narrative. The final report can contain model language, edited facts, or omissions not evident from the document. Identify what the officer personally perceived, what source material the tool received, what it generated, and what the officer changed and verified. A signed report does not transform a model’s unsupported addition into firsthand knowledge.

Generated reconstructions are opinions or aids, not recordings

A system can animate a crash, fill missing frames, create an age-progressed face, reconstruct a scene, generate a suspect image, or simulate a sequence. The output depends on selected inputs and assumptions. It should be labeled, validated, and offered under the rules governing expert opinion or illustrative aids. Photorealism can imply that disputed events were observed when they were modeled.

Chatbot conversations can be evidence of human use

A defendant’s prompts and a service’s outputs may show research, planning, knowledge, fantasy, testing, or innocent work. Provider exports, account history, device artifacts, and session data may authenticate the exchange. The model’s response is generated content, not proof that its statements are true. Attribution requires account control, device use, timing, and context, especially when accounts are shared or compromised.

AI reliability depends on task, data, version, and deployment

Fluency can conceal factual error

Generative systems are optimized to produce responsive output, not sworn testimony. A coherent answer can contain nonexistent facts, fabricated citations, altered chronology, or confident uncertainty. Repeated prompting may produce different answers. Human reviewers can become anchored by the first fluent result and miss source conflicts. Verification means comparing each material statement with preserved evidence, not asking the model whether it is correct.

Model versions can change silently

Cloud services update models, safety rules, system instructions, retrieval indexes, and output formatting. A brand name and date may not identify the version used. Preserve model identifier, endpoint, release, account tier, system instructions available to the user, temperature or randomness settings, seed if supported, tools, retrieval collection, and timestamp. A later rerun can differ and does not necessarily reproduce the original analysis.

Inputs can be incomplete or contaminated

A summary is no better than the material supplied. Investigators may omit exculpatory files, use OCR with errors, combine duplicate records, or include an analyst’s assumptions as source facts. Prompt injection hidden in a document can direct some systems to ignore content or change output. Inventory every input with hashes and show which files failed, were truncated, excluded, duplicated, or inaccessible.

Retrieval can supply the wrong source

A retrieval-augmented system searches a document collection and provides selected passages to a model. Chunking, embeddings, query phrasing, ranking, permissions, and index freshness determine what is retrieved. A citation can point to a real document that does not support the generated proposition. Preserve the query, returned passages, scores, collection version, permissions, and source pages, and verify every citation directly.

Benchmarks may not represent the case

Published accuracy can reflect clean test data, broad categories, or an earlier version. Criminal evidence may involve dark video, rural accents, code-switching, damaged audio, rare objects, unusual faces, or adversarial inputs. Ask whether validation included comparable conditions and whether the selected threshold was fixed before examining this case. Overall accuracy can hide severe performance differences among categories.

Automation bias affects human reviewers

A person told that software found a match may interpret ambiguous evidence in that direction. Confirmation can occur in face comparison, transcription, object recognition, document review, and link analysis. Blind or sequential review, independent verification, alternate candidates, and documentation of pre-AI conclusions can reduce bias. The human’s agreement is not independent corroboration if it followed exposure to the same output.

Vendor secrecy does not eliminate the need for testing

A proprietary model may limit access to training data or code. Courts can consider protective orders, validation reports, test sets, version records, logs, deposition, or independent case-specific testing. The proponent still bears the evidentiary foundation. “Trade secret” cannot itself establish reliability, and lack of access can affect the defense’s ability to confront the opinion and the court’s ability to evaluate it.

Security filters can alter what investigators receive

Safety rules, moderation classifiers, privacy filters, and account permissions may block an input, redact a field, refuse an answer, or substitute a generalized response. Those controls can change between users and dates. Preserve refusal messages, filtered material, account role, and vendor notices. An analyst should not describe the model as having reviewed the complete evidence when a filter prevented part of the collection from reaching it.

AI output can embed human hearsay

A model may summarize witness statements, databases, reports, web pages, and analyst notes. The output adds a transformation layer but does not erase the underlying declarants. Identify each source, whether offered for truth, and any exception or nonhearsay purpose. A business-record certificate for the software log does not make every source statement or generated conclusion a business record.

A reproducible AI record includes far more than the final answer

Preserve the complete prompt hierarchy

Relevant instructions can include a system prompt, vendor rules, agency template, user prompt, prior chat, uploaded documents, tool calls, and hidden retrieval context. Producing only the officer’s last question omits instructions that shaped the answer. Preserve exact text, order, roles, attachments, parameters, and session. If the vendor will not disclose a layer, document that limitation rather than claiming complete reproducibility.

Preserve every output, not only the preferred one

Investigators may regenerate, branch, revise, or ask leading follow-ups until an answer fits their theory. Every result can bear on reliability and bias. Save timestamps, output IDs, edits, thumbs-up or feedback, deleted branches, retries, and human selections. A report that quotes only the final answer should identify how many alternatives preceded it and why that version was chosen.

Link source evidence with hashes and stable identifiers

Create an inventory of every uploaded or retrieved item with file name, source, custodian, hash, size, format, and relevant page or time. Preserve extraction and OCR versions separately from originals. A model citation should map to an exact record location. If an input was a screenshot or analyst summary, identify that derivative rather than implying the system reviewed the original evidence.

Document human review at assertion level

“Reviewed by an officer” is too broad when a 100-page generated report contains hundreds of facts. Identify reviewer, source checked, date, corrections, unresolved items, and final approval for each material section. Preserve tracked changes. A witness should testify from knowledge and disclosed sources, not adopt a model’s prose wholesale and claim memory after the fact.

Keep investigative and evidentiary versions separate

An internal lead list may use low thresholds and tolerate false positives. A trial exhibit requires reliable foundation and fair presentation. Preserve the original lead output, subsequent verification, final report, and rejected candidates. Do not relabel an exploratory score as an identification or an investigative summary as a fact chronology simply because later evidence supported part of it.

Maintain audit and access logs

Account logs can show users, sessions, uploads, prompts, tool calls, exports, sharing, deletions, and administrative changes. Provider retention may be short. Prompt preservation should address both local agency records and vendor data. Logs also reveal whether confidential evidence left the approved environment, whether an outside model trained on it, and whether other users accessed or modified the session.

Ordinary evidence and constitutional rules apply to AI-assisted proof

Authentication must match the claimed item

Under N.D.R.Ev. 901, the proponent must support a finding that the exhibit is what it claims. A provider export may authenticate the prompt-output record; it does not prove generated factual content. A human witness can authenticate an edited report as the report submitted; that does not establish every sentence from personal knowledge. System evidence under Rule 901(b)(9) should explain accurate operation for the asserted task.

Electronic certifications have limited reach

N.D.R.Ev. 902(13) and (14) can authenticate a qualifying electronic-process record or copied data through certification and notice. A hash can show copy integrity. Neither provision establishes that a model’s classification, translation, summary, or generated reconstruction is accurate. Distinguish file identity from substantive reliability.

Rule 702 governs specialized AI opinions

If a witness relies on AI to identify, classify, predict, translate, compare, or reconstruct beyond ordinary understanding, N.D.R.Ev. 702 requires sufficient data, reliable principles and methods, and reliable application. Daubert v. Merrell Dow Pharmaceuticals, Inc., 509 U.S. 579, 589–95 (1993), and Kumho Tire Co. v. Carmichael, 526 U.S. 137, 147–52 (1999), require case-specific gatekeeping.

Experts must disclose their real basis

N.D.R.Ev. 703 and 705 govern facts, data, and disclosure supporting expert opinions. An expert who used an AI assistant should identify the task, input, output, verification, and whether the model influenced the conclusion. The witness cannot avoid scrutiny by presenting generated analysis as unrecorded mental work. The ultimate opinion must be the expert’s, supported by methods the expert can explain and defend.

Original evidence remains more probative than a generated summary

N.D.R.Ev. 1001–1003 govern electronic originals and duplicates. When the State offers a summary to prove contents of messages, audio, video, or records, the underlying admissible material and applicable summary rule matter. An AI answer is not an “original” of its sources. A transcript or summary can assist, but disputed content should be tested against the native record.

Rule 16 and constitutional disclosure reach the AI workflow

N.D.R.Crim.P. 16 covers relevant data, reports, examinations, and expert disclosures within its terms. Favorable material under Brady v. Maryland, 373 U.S. 83, 87 (1963), cannot be hidden because a model labeled it irrelevant. Requests should identify inputs, outputs, alternate candidates, rejected results, errors, validation, prompts, logs, human edits, and vendor communications in the prosecution team’s possession or control.

Confrontation applies to human testimonial statements

AI has no conscience to cross-examine, but its output may repeat testimonial statements from absent people or transmit a human analyst’s conclusions. Identify the actual declarants and purpose. Machine-generated values can present foundation and reliability questions even when not hearsay. The State should not use a model to restate an unavailable witness’s accusation and argue that the transformation removes the Confrontation Clause.

Warrant affidavits must distinguish leads from verified facts

An affidavit relying on an AI match should state the system’s role, source, score or rank, threshold, limitations, and corroboration. It should not call a candidate a positive identification. Material falsehoods or omissions concerning error, alternate candidates, bad input, or failed verification can support a request under Franks v. Delaware, 438 U.S. 154, 155–56, 171–72 (1978).

AI-assisted trial presentation needs transparent labels and human accountability

Generated demonstratives are not camera evidence

A photorealistic reconstruction, interpolated frame, translated voice, synthetic narration, or generated scene should be labeled as created material. Under N.D.R.Ev. 107, an illustrative aid used only to help understand evidence is not itself evidence unless admitted under another rule. The jury should see the supporting data and assumptions and receive an instruction preventing confusion.

AI transcripts should remain tied to the source

Play the audio or video and identify disputed words, inaudible portions, human corrections, and speaker labels. Do not let a fluent transcript replace difficult source media. A transcript displayed as an aid should not go to deliberations as evidence unless independently admitted. Competing or neutral transcripts may be fairer when key language is contested.

Confidence graphics can mislead

Percentages, rankings, heatmaps, boxes, and color scales may look scientific while encoding a vendor-specific score. The witness should explain the denominator, candidate population, threshold, calibration, and error. N.D.R.Ev. 403 permits exclusion when misleading effect or unfair prejudice substantially outweighs probative value. Use accurate terminology such as candidate, label, or model output.

The human witness must own every opinion

A witness may use tools, but must identify personal observations, assumptions, sources, method, verification, and conclusion. “The AI said so” is not a foundation. Counsel should ask whether the witness could reach and explain the opinion without the output, whether contrary results existed, and what would falsify the conclusion. Responsibility cannot be delegated to an unavailable vendor.

Jury instructions can prevent categorical thinking

An instruction may distinguish an investigative lead from evidence, a generated aid from a recording, and a model score from certainty. It can tell jurors to evaluate the underlying evidence, witness credibility, method, limits, and corroboration. The instruction should not imply that all AI is unreliable or authoritative; it should match the precise admitted use.

Defense counsel should reconstruct the complete AI workflow

Ask early where AI was used

Request identification of AI or automated systems used for leads, transcription, translation, summarization, review, redaction, face or object search, enhancement, reconstruction, report drafting, risk scoring, or expert work. Broad marketing names are insufficient. Identify vendor, product, model, version, deployment, date, users, task, and whether output influenced a warrant, interview, identification, charging decision, or exhibit.

Preserve native evidence before challenging the output

Obtain the original audio, video, images, messages, documents, databases, and device records supplied to the system. Hash them and compare the actual inputs. If the source was truncated, converted, OCR-processed, or filtered, preserve every stage. A model dispute cannot be evaluated when the original evidence was overwritten by the derivative report.

Demand prompts, versions, settings, and every result

Seek system and user prompts, prior turns, retrieval passages, tools, model ID, parameters, session, logs, account records, all outputs, retries, branches, deleted results, feedback, edits, and final approval. Ask for validation, benchmark details, case-specific tests, known limitations, release notes, and vendor communications. Use protective orders where legitimate confidentiality concerns can be managed without preventing a defense.

Trace every generated assertion to a source

Create a table listing the output claim, cited source, exact page or timestamp, source meaning, accuracy, human reviewer, and correction. Mark claims with no support, conflicting sources, or omitted qualifications. This often converts a complex AI dispute into ordinary record comparison that a court can evaluate directly.

Investigate automation bias and derivative conduct

Determine what officers believed before the output, which candidate they saw, how later witnesses were exposed, what alternative results existed, and whether verification was blind. Map the output to later searches, interviews, photo arrays, reports, and warrants. An AI error can shape human evidence even when the output itself is never offered at trial.

Use experts for a defined technical question

A machine-learning specialist can analyze model evaluation; a forensic examiner can preserve inputs and logs; a linguist can review transcription or translation; a vision specialist can address image classification; a domain expert can test whether the output fits substantive evidence. Define the task narrowly. No single expert automatically covers software, statistics, digital forensics, language, and the underlying forensic discipline.

Seek focused remedies

Potential relief includes preservation, native production, vendor process, protective access, continuance, independent testing, Rule 702 hearing, suppression of a warrant’s fruits, exclusion of unreliable opinions, correction of reports, alternate transcripts, limitation of demonstratives, Rule 403 exclusion, and jury instructions. Identify the exact output and prejudice rather than asking the court to decide whether AI is generally trustworthy.

Keep the burden of proof on admissible evidence

A model can organize an investigation, but it cannot supply missing identity, intent, possession, causation, or credibility. The State must prove every element beyond a reasonable doubt through admissible evidence. When the model’s work cannot be reproduced, its sources are incomplete, or human verification is circular, the defense should explain the resulting evidentiary gap in ordinary language.

Frequently asked questions about AI-generated evidence

Is an AI output automatically admissible?

No. Its admissibility depends on purpose, authentication, hearsay, expert reliability, completeness, original evidence, discovery, relevance, prejudice, and other applicable rules.

Can police use AI to generate investigative leads?

They can use tools to assist investigation, but a lead is not proof. Any search, arrest, identification, or trial claim still requires lawful authority and reliable evidence.

Is an automated transcript the evidence?

The original audio or video is the source. A transcript may assist, but errors, speaker labels, human edits, model version, and disputed words must be disclosed.

Can AI translate evidence?

It can assist, but translation should be checked by a qualified person against the source audio or text, with dialect, ambiguity, and alternatives preserved.

What is AI confabulation?

NIST uses the term for confidently presented erroneous or false generative output. A fluent answer must be verified against the preserved record.

Does a facial-recognition match identify the defendant?

It usually returns a ranked candidate from a gallery. Independent verification, algorithm and gallery details, threshold, image quality, and human-review procedures matter.

Can AI write a police report?

A tool can assist drafting, but the officer must verify every factual representation and distinguish personal knowledge, source material, and generated language.

Can a generated reconstruction be shown to a jury?

Potentially as a properly founded expert opinion or illustrative aid, with assumptions and generated nature disclosed. It should not be presented as an actual recording.

Why does the model version matter?

Cloud models and settings change. Different versions can return different outputs, and a later rerun may not reproduce the original work.

Are prompts discoverable?

Prompts, inputs, outputs, logs, and settings can be material to evaluating a report, examination, expert opinion, or warrant. The applicable discovery rule and case facts control.

Must the State disclose rejected AI outputs?

Contrary candidates, failed runs, alternate outputs, and errors may be material to reliability, impeachment, or the defense and should be specifically requested and preserved.

Does a high confidence score mean the result is correct?

Not necessarily. It may be an internal model score. Meaning, calibration, threshold, candidate population, and false-positive and false-negative performance require explanation.

Can proprietary software be challenged?

Yes. Protective procedures may address legitimate secrets, but the proponent still must establish reliability and provide a fair opportunity to test the opinion.

Can AI output contain hearsay?

Yes. It may restate human assertions from reports, witnesses, databases, or web sources. Generation does not eliminate the underlying declarants or hearsay layers.

Does AI create a Confrontation Clause issue?

The model is not a human declarant, but output may relay testimonial statements or a human analyst’s conclusions. The actual declarants and purpose must be identified.

When is an AI expert hearing appropriate?

When specialized output will support an opinion and reliability is genuinely disputed, the court can examine data, method, testing, errors, fit, version, and application under Rule 702.

What should defense counsel request?

Native sources, exact inputs, prompts, retrieval passages, model and version, settings, logs, all outputs, rejected results, human edits, validation, error information, vendor materials, and derivative uses.

Can AI fill a missing element of a criminal charge?

No. The State must prove each element through admissible evidence. A model’s unsupported assertion, candidate, or prediction is not a substitute for proof beyond a reasonable doubt.

Official authorities cited

Legal information, not legal advice. AI-generated-evidence issues depend on the task, source data, model and version, prompts, settings, retrieval, complete outputs, validation, error, human review, discovery, expert method, authentication, hearsay, confrontation, original evidence, demonstratives, attribution, and current law.


Call nowChoose a time