On September 30, the Arizona Court of Appeals upheld a manslaughter conviction but ordered a new sentencing hearing because the trial judge had considered an AI-generated video of the deceased victim.

According to the ABA Journal’s report, the victim’s family created a video depicting Christopher Pelkey speaking about forgiveness and addressing the defendant directly. Pelkey, an Army veteran, was 37 when he was killed.

The three-judge panel found that the trial judge’s consideration of the video so prejudiced the defendant that the sentencing procedure was fundamentally unfair. The conviction remained in place, but the 10.5-year sentence did not.

This was not another case about a lawyer filing fictional citations. It concerned synthetic media that put new words and a simulated performance into the likeness of a person who could not approve either one.

Synthetic media creates a provenance problem

Courts and lawyers already deal with edited photographs, surveillance recordings, reconstructions, demonstrative exhibits, and victim-impact material. AI adds a different problem: it can generate a persuasive performance that never occurred.

A viewer may understand intellectually that a video is synthetic and still react to the apparent voice, facial expression, timing, and direct address. Those choices are created by living people and software, not the person shown on screen.

That does not mean every AI-assisted image or video is automatically inadmissible or improper. The Arizona decision turned on that case and that sentencing. It does show why “this was made with AI” is not enough information for a meaningful review.

The reviewer needs to know where the source material came from, who wrote the words, how the voice was produced, what the tool generated, what a person edited, and how the final product will be described.

Start asking questions at intake

Synthetic evidence will not always arrive with a label. A client may bring a recording from social media, an opposing party may produce an image, or an employee may create a demonstrative without realizing that the software changed more than the background.

Firms need a consistent intake process for audio, video, and images whose origin matters. That process should apply whether the material appears helpful or harmful.

Useful questions include:

Who created or obtained the file?
Is this the original file or a downloaded, compressed, or edited copy?
What device, application, or AI tool produced it?
Were a face, voice, background, object, or spoken words generated or altered?
What source recordings, photographs, prompts, scripts, and edit history still exist?
Who can authenticate the underlying material?
Has the original file and its metadata been preserved?
How will the synthetic or edited portions be disclosed?

Those questions are not a substitute for an evidentiary analysis. They help preserve the facts the attorneys and experts will need to perform one.

Do not let the polished version replace the source material

Generative tools make it easy to produce a cleaner, shorter, or more persuasive version of messy source material. That convenience can also erase the record of how the final exhibit was made.

Keep the original files. Preserve the prompt history, script, tool name, account used, settings, output versions, and human edits. If a vendor or expert creates the material, the engagement should address retention and access to those records.

The firm should also decide who reviews the finished product. Technical review can identify generated or altered elements. Legal review addresses how the material may be used and described. The attorney responsible for the matter needs enough information to explain the complete chain without guessing.

An unlabeled export sitting in the matter folder is not enough. Six months later, the employee who created it may not remember which parts came from a recording and which parts came from the model.

Detection tools cannot carry the whole process

Deepfake detectors may provide another data point, but they do not solve the provenance problem. A detection result can be uncertain, tools can disagree, and ordinary editing or compression can affect the analysis.

The stronger control is a documented chain from source to final output. That chain helps the firm evaluate material it creates and challenge material it receives.

For incoming evidence, preserve the file before converting or uploading it to another service. Record where it came from and when it was collected. If authenticity is disputed, the firm may need qualified forensic assistance rather than an online detector.

For material the firm creates, use approved tools and retain enough information for another person to reproduce or explain the process. A lawyer should not learn at a hearing that a staff member used an AI feature hidden inside an editing application.

Iowa lawyers are already being asked to consider this

The Iowa State Bar Association’s current six-part AI Training Series includes a session on AI-generated evidence and deepfakes. The series runs from September 30 through November 19 and also addresses automated workflows, ethics, malpractice risk, engagement letters, and trial preparation.

That does not tell us how an Iowa court would decide the Arizona issue. It does show that synthetic evidence has moved from a hypothetical technology question into practical legal training.

Firms do not need to wait for a disputed exhibit to set basic rules. Decide which tools may create or alter evidentiary material, what records must be retained, who reviews the output, and how the use of AI will be disclosed.

Test the process with one file

Choose a sample video, audio recording, or demonstrative and ask the litigation team to trace it.

Can they find the original? Can they identify every edit? Do they know whether an AI feature changed the voice or image? Is the script preserved? Can the attorney explain who approved the final version and how it differs from the source?

The Arizona court did not erase the conviction. It did require the sentencing to be done again. That is a costly way to discover that the process for synthetic material was not ready.

If AI-generated material is entering your matters without a documented review trail, Artech can help build the tool, access, retention, and approval controls around the workflow.