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

Deepfakes and Altered Digital Evidence in Criminal Cases

A realistic image, voice, or video no longer proves its own origin. At the same time, calling inconvenient evidence a “deepfake” does not make it false.…

ElementsPenaltiesEvidenceNext steps

AI-generated images, face swaps, cloned voices, synthetic video, altered photographs, edited recordings, repurposed media, provenance, source devices, native files, metadata, hashes, compression, enhancement, content credentials, detector limits, experts, authentication, discovery, exhibits, jury instructions, and reasonable doubt

Deepfakes and Altered Digital Evidence in Criminal Cases

A realistic image, voice, or video no longer proves its own origin. At the same time, calling inconvenient evidence a “deepfake” does not make it false. Criminal courts need disciplined proof: what device captured or generated the media, who controlled it, where the first-generation file went, what software touched it, which changes were routine, whether any content was inserted or removed, and whether a disclosed method can reliably answer the question presented. Provenance and corroboration matter more than visual confidence.

The right question is not simply “real or fake?” Determine whether the exhibit is captured, copied, compressed, enhanced, edited, repurposed, partially synthetic, or fully generated; identify every lifecycle stage; preserve the native source; and test whether any change affects the fact the State offers the exhibit to prove.

Digital media can change without becoming a deceptive deepfake

A copy is not automatically false

Sending a photograph, downloading a video, exporting a security clip, or copying audio can change file name, container, metadata, resolution, compression, and timestamps while preserving the relevant content. NIST’s official Examining Digital Media guide cautions that a copy or copy of a copy is not inherently fake or inauthentic. The issue is whether the exhibit accurately represents the claimed source and event and whether any lifecycle change matters to the disputed fact.

Ordinary editing can be material or immaterial

Cropping, rotating, color correction, stabilization, noise reduction, volume adjustment, trimming, captions, slow motion, and format conversion are common. A crop can remove another person; noise reduction can suppress a syllable; stabilization can distort edges; a caption can tell the viewer what to perceive. An edit need not be malicious to create a misleading exhibit. The proponent should disclose what changed and preserve the untouched source.

Repurposed media can be technically genuine but factually false

A real photograph or video from another date, place, person, or event can be reposted with a false description. No pixel synthesis is necessary. Reverse chronology, account history, source posts, weather, landmarks, shadows, clothing, device metadata, and independent witnesses can reveal repurposing. A detector designed to find generated imagery may correctly call the file nonsynthetic while missing the false attribution that matters in court.

Partially synthetic media changes selected content

A face can be replaced, lips retimed, a person inserted or removed, a background regenerated, an object erased, or words changed while most of the original remains. Audio can clone a voice for a new phrase or alter only one word. The alteration may be brief and survive inside an otherwise authentic recording. Analysis should localize the disputed region and not average confidence across the entire file.

Fully synthetic media begins with generation rather than capture

Text-to-image, text-to-video, and voice-generation systems can create content without a camera or microphone recording the depicted event. A generator may leave metadata, watermarks, content credentials, characteristic structure, or account records, but those indicators can be removed. The absence of a label does not prove capture, and the presence of an editing application does not prove generation. Source history is stronger than one artifact viewed in isolation.

Enhancement can create a new demonstrative interpretation

Super-resolution, frame interpolation, colorization, deblurring, voice isolation, and generative fill can produce information that the source did not directly record. Some tools estimate plausible detail rather than recover unique ground truth. A generated intermediate frame may look like captured evidence. Courts should distinguish a faithful enlargement or conventional filter from a model-created visualization and decide whether the output is evidence, expert opinion, or an illustrative aid.

Compression artifacts can resemble manipulation

Block boundaries, ringing, motion smearing, missing frames, variable frame rates, rolling shutter, dropped audio, resampling, and platform transcoding can produce strange faces, hands, edges, timing, or lip movement. Multiple social-media uploads make diagnosis harder. An analyst who treats every anomaly as synthesis risks a false accusation. Compare the suspected artifact with codec behavior, neighboring frames, known platform processing, and earlier versions.

“AI-generated” does not answer intent or legal relevance

Synthetic content may be satire, an illustration, an investigative reenactment, a fraudulent impersonation, an alibi fabrication, or contraband. The same generation method can serve lawful and unlawful ends. The State must prove the applicable offense, authorship, knowledge, purpose, and effect rather than rely on fear of artificial intelligence. The defense likewise needs evidence connecting suspected generation to the exhibit and disputed issue.

Provenance reconstructs the media’s lifecycle

Begin with the claimed capture or generation event

Identify device, owner, operator, application, account, date, time, location, settings, sensor, microphone, lens, network state, and surrounding circumstances. If generated, identify service, model, account, prompt, seed, input files, revision history, output, and billing or activity logs. A screenshot of a social-media post starts near the end of the lifecycle. Work backward toward the earliest available source.

Preserve the source device before routine use changes it

Phones and cameras synchronize, optimize, delete, transcode, create thumbnails, update applications, and overwrite caches. Cloud services can expire originals. A forensic or documented native export may preserve file-system metadata, application databases, edits, project files, account tokens, thumbnails, deleted artifacts, and share history. Merely emailing the visible media can discard the evidence most useful for authenticating or challenging it.

Collect every known version

Obtain the camera original, edited project, exported file, cloud copy, message attachment, platform download, repost, law-enforcement copy, forensic extraction, prosecutor exhibit, and enhanced output. Hash each version and record its source. Differences can reveal when content changed. Identical hashes establish identical bytes; different hashes can result from harmless conversion or material alteration and require comparison.

Metadata helps but can be absent or forged

EXIF, file-system, container, codec, edit-history, application, cloud, account, and provider metadata can identify devices, times, software, coordinates, or transformations. Platforms often strip fields. Editing tools can rewrite them, and anyone with access can falsify some values. Treat metadata as evidence to corroborate against device databases, provider logs, neighboring files, known events, and physical facts, not as a self-proving certificate.

Hashes prove sameness after a defined moment

A cryptographic hash can show that a file presented at trial matches the file hashed during acquisition. It does not prove the hashed file was authentic before acquisition, complete, or captured by the claimed device. Document who selected the file, how it was acquired, the algorithm, tool, time, and storage. A hash without a trustworthy acquisition point can preserve a fabrication perfectly.

Content credentials and signatures are useful but not conclusive

Some devices and applications attach signed provenance assertions describing capture and edits. Verify the signature, certificate chain, claim, ingredients, time, device or service, and whether the final file still contains the manifest. Missing credentials do not prove fakery because many legitimate workflows do not create or preserve them. Valid credentials support only what was signed and do not necessarily establish the truth of a scene or identity of a human operator.

Account and project history can be more informative than the final file

Generation services, editing applications, cloud galleries, and collaboration tools may retain prompts, source uploads, masks, layers, timelines, revisions, exports, login sessions, payment records, and sharing. Seek prompt preservation before retention expires. A flattened image can conceal hundreds of edits that remain visible in a project file. Conversely, a camera roll with neighboring images and consistent sensor output can support ordinary capture.

Independent corroboration protects against both false acceptance and false denial

Other cameras, witnesses, location records, messages, transactions, weather, physical evidence, device possession, and surrounding files can confirm or contradict the depicted event. A person should not defeat reliable evidence by merely asserting that modern tools could create it. Nor should a realistic appearance override missing provenance and contrary facts. Corroboration tests the event, not just the pixels.

Media-forensic methods must fit the file and the claimed alteration

Visual inspection is triage, not a final method

Odd blinking, teeth, hands, reflections, shadows, lip sync, or boundary motion may suggest review. New generators improve quickly, while compression creates similar anomalies. Human observers can become biased after being told a clip is fake. Record observations before running tools, use the original at native resolution and rate, compare surrounding frames, and avoid conclusions based on a viral checklist.

File and codec structure can reveal transformations

An examiner can inspect containers, streams, timestamps, group-of-pictures structure, frame types, quantization, audio layout, variable frame timing, encoder tags, and discontinuities. These features may show re-encoding, splicing, export, or screen capture. Re-encoding is not synonymous with deception. The method should explain which structural finding supports which claimed lifecycle event and what common innocent process could produce it.

Pixel and signal analysis can localize inconsistency

Techniques may examine sensor noise, demosaicing, compression grids, lighting, reflections, geometry, frequency patterns, motion, optical flow, facial landmarks, audio phase, room acoustics, electrical hum, or background continuity. Validity depends on source quality and assumptions. Social-media compression, stabilization, filters, screenshots, and repeated exports can destroy cues. Report inconclusive results honestly rather than convert missing signal into proof.

Detector performance depends on the tested population

A model trained on known generators may perform poorly on newer models, unseen post-processing, different demographics, low resolution, or adversarial changes. NIST’s Guardians of Forensic Evidence program identifies generalization and robustness to post-processing and anti-forensics as central challenges. A vendor’s overall accuracy number does not establish performance on this file, generator, compression level, or decision threshold.

A probability score needs a defined meaning

A detector may output model confidence, class probability, anomaly score, or rank—not the probability that the exhibit is fake. Ask what classes were trained, base rates, threshold, calibration, false-positive and false-negative rates, test set independence, confidence interval, and effect of preprocessing. Running many detectors and selecting the most accusatory result is not validation. Disclose every run and output.

Audio cloning requires both provenance and acoustic analysis

Voice conversion and text-to-speech can imitate identity while changing content. Examine source file structure, codec history, speech timing, prosody, background, reverberation, breaths, room response, phase, electrical noise, and any generation metadata. Speaker recognition and synthetic-speech detection answer different questions. A voice can be the correct speaker in an edited sentence, or a clone can sound similar while a detector remains uncertain.

Face recognition and deepfake detection are different tasks

Face comparison asks whether visible features support identity; deepfake analysis asks whether media was generated or manipulated. A real video can show the wrong person, and a face swap can use features from the correct person. Quality, angle, lighting, occlusion, demographic performance, and reference-image selection affect comparison. Experts should not extend training in one task into unsupported opinions on the other.

Reproducibility requires the exact inputs and environment

Preserve the analyzed file hash, tool, version, model weights, settings, threshold, hardware where relevant, preprocessing, regions selected, scripts, logs, and outputs. A cloud detector can change without notice. If the model is proprietary, the defense needs enough information to test reliability and challenge the opinion, subject to protective conditions where justified. A screenshot of a result is not a reproducible examination.

Existing evidence rules address authenticity without presuming or denying AI alteration

Rule 901 asks what the proponent claims

A photograph offered only to show what a witness received requires a different foundation from one offered as an accurate depiction of an event. A generated illustration offered as an expert demonstrative is not authenticated as camera footage. Define the claim precisely: source, date, speaker, depicted person, completeness, or process. Rule 901’s threshold is satisfied by evidence allowing a reasonable juror to find authenticity; the issue of weight then remains for trial.

North Dakota recognizes multiple authentication methods

State v. Burton, 2025 ND 83, ¶¶ 20–25, explains that Rule 901 places no limit on the number of authentication methods and requires proof sufficient for a reasonable juror to find the item is what it purports to be. Its 911-recording holding used circumstantial evidence and testimony. The same flexible approach permits source, witness, system, metadata, and expert evidence in a deepfake dispute without making any one method mandatory.

Rule 902 certifications establish limited propositions

N.D.R.Ev. 902(13) and (14) permit qualifying certified electronic-process records and copied electronic data with notice and inspection. A hash-based certification can show that data was copied accurately from a specified device or file. It does not prove the copied media depicted a real event, identify the person who created it, or rule out alteration before acquisition. State the certification’s exact scope.

Expert reliability must be method-specific

Under Daubert v. Merrell Dow Pharmaceuticals, Inc., 509 U.S. 579, 589–95 (1993), courts consider whether specialized reasoning is reliable and fits the case. Kumho Tire Co. v. Carmichael, 526 U.S. 137, 147–52 (1999), extends gatekeeping to technical expertise. For AI detection, examine testing, peer review, error, standards, acceptance, data, model drift, validation on comparable media, and reliable application—not job title or software output alone.

Original-recording rules focus attention on the native source

N.D.R.Ev. 1001 treats an output accurately reflecting electronically stored information as an original. Rule 1003 generally permits duplicates unless a genuine authenticity question or unfairness makes that inappropriate. When the dispute concerns alteration, generation, metadata, or missing context, the native file, source device, project, and earliest version become important. A social-media screen recording may fairly show the post but be inadequate to prove how the underlying media was made.

Discovery should include examination material and exculpatory evidence

Under N.D.R.Crim.P. 16, the defense can seek relevant documents, data, tangible objects, reports, examinations, and expert disclosures within the rule. Constitutional disclosure duties extend to favorable material evidence. Request native files, source devices, every version, acquisition logs, tools, model versions, settings, outputs, failed tests, validation, notes, communications, prompts, project history, and contrary findings. An analyst should not disclose only the successful detector run.

Unsupported speculation does not create a genuine authenticity dispute

The possibility that software could fabricate media is relevant context, not proof that it happened. A challenge should identify missing provenance, inconsistent metadata, undisclosed edits, structural anomalies, witness conflicts, account evidence, or a qualified analysis. Courts can require a focused proffer before extensive expert proceedings. This prevents “deepfake” from becoming a universal objection while preserving meaningful scrutiny where facts support it.

Trial presentation should reveal transformations and uncertainty

Show the jury the source before an enhanced derivative

If an enlarged, stabilized, brightened, filtered, slowed, transcribed, or AI-assisted version will be used, identify it clearly and preserve the unprocessed file. Explain each transformation and whether it adds estimated content. Allow side-by-side comparison at correct scale and speed. A demonstrative should not carry the same exhibit label as the native evidence if the distinction matters.

Rule 107 separates illustrative aids from evidence

N.D.R.Ev. 107 governs material used only to help the factfinder understand evidence. A generated reconstruction, colorized frame, model heatmap, or detector visualization may illustrate an expert’s opinion without independently proving the event. The court can control use, instruct the jury, and prevent deliberation-room access when the aid is not admitted evidence. Labels should state that the image is illustrative.

Heatmaps and boxes can imply more certainty than the model supplies

A colored region may show where an algorithm attended, not where manipulation occurred. Bounding boxes can be chosen by the analyst. Smooth confidence bars hide thresholds and error. Require the witness to explain what every visual encodes, which output came directly from the model, and which annotations were added later. Rule 403 permits exclusion or correction of displays likely to mislead.

The jury may need a precise limiting instruction

An instruction can distinguish authentication from ultimate truth, native evidence from an aid, source media from an enhanced version, and detector opinion from certainty. It can explain that copying or compression does not itself prove falsity and that an absence of detectable artifacts does not guarantee authenticity. Instructions should match the admitted evidence rather than teach general fear about artificial intelligence.

Hearsay and confrontation remain independent

An authentic recording can contain inadmissible human statements. A synthetic voice may be offered to prove that a defendant created a fraudulent file rather than for the truth of its words. Identify the proponent’s purpose and every declarant. The Confrontation Clause, hearsay, privilege, relevance, completeness, and unfair prejudice do not disappear because authenticity is disputed or because an expert calls media manipulated.

Authorship and criminal intent need human evidence

An account or device may contain a generated file without proving who created it, entered a prompt, knew it was synthetic, or used it for an unlawful purpose. Examine logins, sessions, payments, prompts, source images, messages, project files, possession, sharing, admissions, and motive. Shared computers and cloud accounts complicate attribution. Generation evidence can prove media origin while leaving the charged actor and mental state unresolved.

A disciplined defense investigation starts with preservation, not a detector website

Send immediate source-specific preservation demands

Identify the device, account, post, message, original upload, platform, cloud gallery, editing application, generator, project, prompt, input media, native output, metadata, audit logs, account sessions, billing, content credentials, thumbnails, caches, and every police copy. Ask that automatic deletion and transcoding be suspended. Preserve reference recordings and images before appearance, voice, or account data changes.

Acquire evidence through documented forensic methods

Use an examiner who can preserve device state, obtain native files, image relevant storage where authorized, export cloud records, compute hashes, and record tool versions. Do not test the only copy in an application that rewrites metadata. Work from verified duplicates and retain a pristine source. If the State controls the device, seek a protocol, joint observation, protective order, or independent image as appropriate.

Create a media-lifecycle table

For every version, list source, custodian, date, device, application, file name, size, duration or dimensions, codec, metadata, hash, edits, export, upload, download, and relationship to the next version. Mark unknown stages. This compact record often reveals more than a detector because it shows where alteration was possible and whether the trial exhibit can be traced to the alleged event.

Define the disputed proposition before testing

Ask whether the issue is capture versus generation, changed words, swapped face, removed object, wrong date, repurposed event, incomplete context, speaker identity, or misleading enhancement. Select methods that answer that question. A generic deepfake classifier may be irrelevant to a real video posted with a false date. Narrow scope improves reliability and prevents costly analysis of undisputed content.

Audit the State’s complete examination

Request hashes, source files, software, model, version, weights, settings, thresholds, preprocessing, regions, logs, notes, screenshots, reports, validation, training and test descriptions, error information, all outputs, and unsuccessful or contrary results. Determine whether the tool changed after analysis or requires a cloud service. Reproduce the test where possible and evaluate performance on comparable compressed media.

Use independent corroboration aggressively

Find other cameras, witnesses, original posts, backups, device-location records, account sessions, messages, source images, weather, lighting, landmarks, transactions, and physical evidence. A second angle or earlier file can resolve what a complex detector cannot. Corroboration also protects the defense from an unsupported fabrication claim that distracts from stronger identity, context, or intent arguments.

File focused motions

Potential relief includes native production, device inspection, preservation, additional expert disclosure, continuance, Rule 702 hearing, exclusion of unreliable detector opinion, authentication hearing, limiting instruction, production of the complete recording, correction of a misleading demonstrative, and exclusion under Rule 403. Identify the exact claim and exhibit. A motion asserting only that deepfakes exist is unlikely to frame a testable legal issue.

Present uncertainty without overstating it

An expert may conclude authentic, manipulated, consistent with a workflow, unsupported, or inconclusive. Inconclusive does not mean fifty-fifty and does not prove fabrication. Cross-examination should expose assumptions, missing sources, detector limits, transformations, and alternative explanations in plain language. Reasonable doubt can arise from broken provenance and weak attribution without promising that a tool can name the generator.

Frequently asked questions about deepfakes and altered digital evidence

What is a deepfake?

It is visual, audio, or multimodal content artificially generated or modified, often to make a person appear to say or do something. The term can include fully or partially synthetic media.

Is every edited image or video a deepfake?

No. Cropping, compression, color correction, trimming, and stabilization are ordinary edits. They can still matter if they remove context or create a misleading impression.

Does a copied file become inauthentic?

No. A copy may remain accurate even if metadata or compression changes. The proponent should trace the copy to its source and disclose material transformations.

Can real media be misleading without alteration?

Yes. Genuine media can be repurposed with a false date, place, identity, or description. Event attribution requires provenance and corroboration, not only synthetic-content detection.

How is digital media authenticated in North Dakota?

Rule 901 requires evidence sufficient to support a finding that the item is what the proponent claims. Witnesses, source devices, circumstances, systems, metadata, hashes, and experts may contribute.

Does a hash prove a file is real?

No. It proves byte-for-byte sameness between hashed copies after a defined acquisition. It does not prove the file was genuine before acquisition or depicts a real event.

Do missing metadata prove a deepfake?

No. Messaging and social platforms commonly strip metadata. Missing fields reduce available provenance but do not by themselves establish generation or alteration.

Can metadata be falsified?

Yes. Some metadata can be edited or regenerated. It should be corroborated with source-device, application, provider, account, neighboring-file, and physical evidence.

Are deepfake detectors reliable?

Performance varies by generator, media type, compression, post-processing, demographics, and unseen methods. Case-specific validation, error measures, version, threshold, and reproducibility are essential.

What does a detector confidence score mean?

It may be a model score rather than the probability the media is fake. The developer’s definitions, calibration, threshold, classes, and test results must be examined.

Can compression look like manipulation?

Yes. Blocking, smearing, dropped frames, resampling, and repeated transcoding can create anomalies. Analysts must compare them with normal codec and platform behavior.

Can enhanced video add details?

Some conventional processing reveals existing signal; generative super-resolution or frame interpolation can create plausible estimated content. The original and exact method must be disclosed.

Can AI clone a person’s voice?

Yes. Voice cloning can synthesize new speech or convert another speaker. Provenance, acoustic analysis, source files, account records, and corroboration are needed.

Are content credentials conclusive?

No. Valid signed provenance can support a documented lifecycle, but many legitimate files lack credentials, credentials can be lost, and a signature proves only the signed claim.

When is an expert needed?

An expert may be necessary for file structure, signal analysis, detector interpretation, voice cloning, provenance systems, enhancement, or complex alterations beyond ordinary observation.

What should defense counsel request?

Source devices, native files, all versions, project and prompt history, metadata, hashes, account and provider records, acquisition logs, detector inputs and outputs, tools, versions, validation, notes, and enhancements.

Can a defendant simply claim a video is AI-generated?

A bare possibility is usually weak. A meaningful challenge identifies provenance gaps, conflicting sources, anomalies, undisclosed edits, account evidence, or qualified forensic findings.

Does proof that media is synthetic prove the defendant created it?

No. The State still must prove authorship, knowledge, purpose, possession or use, and every element of the charged offense with admissible evidence.

Official authorities cited

Legal information, not legal advice. Deepfake and altered-media issues depend on the precise authenticity claim, source device, native file, complete lifecycle, edits, compression, enhancement, metadata, hashes, credentials, accounts, prompts, projects, corroboration, expert method, detector validation, disclosure, trial presentation, attribution, intent, and current law.


Call nowChoose a time