AI Forensics, System Reconstruction and Digital Evidence

AI Systems Reconstructed From Code, Prompts, Logs and Training Data

When the AI system itself is in dispute, the questions are how it was built, what it was trained on, what it was told and who did what along the way. Global Digital Forensics preserves the code, data, prompts and logs before a provider purge or a version update erases them, reconstructs the history of the system from those records and documents every step so another examiner can repeat it.

Get a Free Consultation Call (800) 868-8189

Audio waveform and video frames under forensic review.

Media Authentication, System Reconstruction and Security Testing

Media authentication asks whether a recording or image is what it claims to be. System reconstruction asks how a model or application was built and how it behaved, using the retained prompts, outputs, retrieval context, configuration, logs and versions. Security testing asks whether a defined attack path works under authorized rules of engagement. Each needs a different method, and the first thing a senior examiner does on your call is name which one your dispute turns on, so the budget goes to the question the case actually asks.

Global Digital Forensics has operated since 1992. Our founder, Joseph Caruso, has worked in technology since 1980 across software development, mainframe databases, real time transaction systems and cybersecurity incident response. His consulting expert work includes Motorola Mobility v. Microsoft and Apple v. Samsung, and his work in an American Express consumer class action involved airline reservation architecture, source code and system configuration. Reading source code, configuration and transaction records to establish what a system did is the same discipline an AI dispute requires, and we start every engagement by preserving the original requests, outputs and configuration before any experiment touches them.

Our vocabulary for attacks on models and their mitigations follows NIST AI 100-2 E2025, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations.

If the dispute involves a model, a prompt log or a generated recording, get a free consultation and a senior examiner will tell you which examination the evidence supports before anyone quotes a scope.

Get a Free Consultation

Technical methods and scope

Matching the Evidence to the AI Question

AI forensics covers synthetic or altered media, disputed model activity, prompts and outputs, retrieval sources, agent actions and intellectual property questions involving models and training data. Media authentication starts with original files and their recording history. Application reconstruction starts with versions, configuration, prompts, logs and controlled tests. We pin down the disputed claim first, so every hour of examination goes where the case turns.

Findings rest on independent observations that agree with one another, with the confidence and the alternatives written down. Compression, transcoding, screen recording, ordinary editing and platform processing can look like generation or manipulation, and recompression can strip the traces a generator leaves behind. A detector score is therefore weighed against the history of the file, and the report states what the evidence shows rather than what a tool printed.

Synthetic and Altered Media

We go after the earliest available file, because it carries more context than a social media download or a messaging preview. The examination records container and codec structure, metadata, any edit history, frame timing, compression, audio continuity and the relationship between picture and sound. Reference recordings, device records, cloud originals and publication history show whether an isolated visual anomaly matters before anyone calls the file a deepfake.

Video work covers consistency between frames, lighting, geometry, face and mouth movement, synchronization of sound with picture and changes in encoding. Audio work covers waveform continuity, spectral behavior, room sound, edits and speaker comparison where adequate exemplars exist. Still images are examined for metadata, resampling, compositing boundaries, sensor patterns and provenance. The conclusion comes from the combined record, and our deepfake detection and media authentication practice carries the specialized media work.

Describe the devices and the deadline and the first call sets the plan.

Model, Application and Agent Records

An AI application leaves records well beyond the model: system and developer instructions, prompts, outputs, model and application versions, retrieval results, tool calls, identity activity, content filters, rate limits and the actions taken downstream. We establish which records the provider holds, which the customer holds and which were never logged, so the reconstruction rests on the coverage that actually exists. When the dispute is about who built what, how it was trained and what it was told, these are the records that answer it.

Controlled testing runs on the preserved configuration or the closest documented replica, and repeated runs separate deterministic application behavior from sampling variation. We record inputs, parameters, environment, time, outputs and side effects, which gives the report a basis anyone can rerun. Technical observations are reported on their own terms, cleanly separated from questions of intent, authorship and policy that the parties will argue.

Training Data, Provenance and Intellectual Property

With the model, the data and authorized access in hand, an agreed empirical test can measure specific behavior, such as whether a model reproduces particular material under documented conditions. We write down the test assumptions, the evidence that was available and any specialist requirements, so the scope of the result is explicit. Where source code and training pipelines are in dispute, the source code analysis this practice has performed for decades applies directly.

Tell us what happened and when, and a senior examiner will say which sources to secure first.

A Report Built on Observations You Can Check

The final record can include a source inventory, a preservation log, a media or system timeline, reproducible test cases, comparison exhibits and a technical report. We explain why each method fits the source and the question, including its error risks, so another qualified examiner can test the work and the decision maker can rely on what the system actually did.

Methodology

Four Steps From Disputed Output to Tested Finding

Define the disputed AI activity

The claim is translated into specific model, application, data and tool behaviors that can be tested, which keeps the opinion answerable and the cost in proportion.
Preserve media and system records

Original media, model and application versions, prompts, outputs, logs and configuration are secured before any test changes the environment or a retention purge erases them.
Test the relevant behaviors

Controlled comparisons and reproducible cases across the authorized model, retrieval, identity, tool and data boundaries test each competing explanation under documented conditions.
Report the evidence and conclusions

Every conclusion is tied to source records and repeatable observations, with confidence and alternatives stated, so counsel knows where the other side will push.

Preservation is the step with a clock on it. If provider logs or conversation histories are still live, call (800) 868-8189 and we will tell you what to hold before the retention window closes.

Evidence commonly examined

AI Evidence We Examine

Original media, prompts and outputs fix the questioned material and the recorded interaction. Model, application and policy configuration documents the conditions at the time of the event, and retrieval sources and data lineage show which information entered the workflow. Identity, tool and application logs connect accounts to recorded actions, version records and test cases let the examiner compare behavior under documented conditions, and device and account artifacts supply the independent records that confirm or challenge the AI activity.

What you can expect

The Record You Receive

An evidence and architecture map connects the sources, systems and boundaries that matter. A reproducible test record documents the conditions, inputs and observed results. A source inventory lists the evidence examined, and the provenance or security report sets out each finding and the observations behind it in language counsel can put in a motion or an expert disclosure.

Scope, Access and Deliverables Agreed Before Work Starts

Tell us about the system, the dispute, the operating constraints and who needs to read the report. A senior examiner agrees the authorized methods and the deliverables with you before anything is touched, and the first consultation is free.

Preview the report contents

Every AI forensic report is organized the same way.

The source inventory lists the prompts, logs, versions and media preserved. The timeline places each recorded interaction and system change in order. The findings tie each conclusion to the source records and reproducible tests behind it, and the report is written to support a declaration, a deposition or trial testimony.

Get a Free Consultation

Frequently asked

Scoping an AI Forensics Engagement

What should we define before this work begins?

The AI system, the disputed output, the available prompts and logs, the model versions and the retention period that applies to each. Retention is the urgent item, because provider logs and conversation histories expire on their own schedule, not the court's.

Can you examine an AI system we no longer control?

Often, yes. Retained prompts, outputs, exports, application logs, account records and device artifacts can establish much of what the system was told and what it did after the provider has moved on. Where a model version or provider record is gone, the report names the gap and builds the account from the records that remain.

Can you tell whether confidential information was shared with an AI tool?

Yes, where the records exist. We examine prompts, uploads, browser and application artifacts, account activity and network records to establish what was entered, by whom and when, and we place those findings on the same timeline as the device and account evidence in the matter.

What does the AI forensics report deliver?

A source inventory, a preservation log, a timeline of the AI activity, reproducible test cases where testing was authorized and a technical report that connects each conclusion to a record and states its confidence and the alternatives considered.

Free consultation

Describe the AI System in Dispute

On the first call a senior examiner hears what the system was supposed to do and what it did, identifies the prompts, logs and versions to preserve before they expire, and outlines the examination, the tests and the report.

A 24/7 hotline and examiners accepted by federal and state courts.

GDF does not accept pro se litigation cases, and GDF may decline matrimonial or family law cases at our discretion.

Reviewed by Joseph Caruso, Founder, CEO and CTO, CISSP. Last reviewed November 3, 2026.