When retained by a party, GDF gives New York counsel an independent technical evaluation of the AI evidence and reports material results whether they support, narrow, or conflict with the proposed theory. The work reconstructs the deployed system and event record across models, instructions, prompts, retrieval, tools, application logic, outputs, and human action. GDF does not provide legal advice, and the court or tribunal determines admissibility and ultimate legal issues.
AI expert witness analysis begins with the deployed system
The relevant object may be a model, generative service, retrieval application, recommendation engine, classifier, or workflow that uses AI for one step. GDF identifies the provider, version, application code, instructions, policy settings, tools, knowledge sources, input path, output handling, and human approvals. Terms such as hallucination, bias, autonomous, trained on, and AI-generated are converted into an observable proposition about retrieval, reproducibility, controls, scoring, editing, or approval.
Questions counsel can convert into technical tests
An AI expert assignment can be framed around questions such as:
- Which model, application, instructions, retrieval sources, tools, and human steps contributed to the disputed result?
- Do preserved records support that a particular prompt, source, output, edit, or approval occurred?
- Can the claimed behavior be reproduced under disclosed conditions, and how variable are repeated results?
- Does an evaluation measure performance on the relevant task and population rather than a convenient benchmark?
- What do false-positive, false-negative, threshold, provenance, and uncertainty records show?
- Which conclusions are prevented by provider opacity, version drift, incomplete logs, or missing human-workflow records?
AI evidence records to preserve
Useful evidence can include model and application identifiers, prompts, system instructions, conversation context, retrieved passages, index versions, tool calls, parameters, safety settings, sessions, timestamps, input files, raw and displayed outputs, moderation results, API responses, user edits, approvals, and downstream actions. Collection records the export or API method, query, permissions, paging, time zone, hashes, and omissions. Provider notes add context but do not establish what occurred in a specific request.
- Model, application, provider, version, and deployment context
- System instructions, prompts, retrieval sources, tools, and parameters
- Raw and displayed outputs with timestamps and downstream handling
- Human review, edits, approvals, and business actions
- Collection method, hashes, gaps, and provider-controlled limits
Reproduction is a controlled test, not a time machine
Probabilistic systems can return different results from the same visible prompt because of sampling, hidden instructions, routing, updates, retrieval, tools, and context. GDF records known parameters, preserves all trials, and states whether the historical environment can be recreated. A matching output shows possibility under recorded conditions, not historical provenance; a nonmatch does not prove an artifact false. Contemporaneous logs, stored requests, hashes, and edit history carry different weight from a later reenactment.
Prompt injection and tool-path analysis
Prompt injection can originate in direct user text, retrieved documents, web content, files, tool output, or instructions passed between agents and services. GDF reconstructs the instruction hierarchy, trust boundaries, retrieval path, tool permissions, filtering, logging, and resulting actions. Controlled tests vary the suspect input while preserving model, application, system prompt, corpus, tools, parameters, and account context where those elements are available. The analysis distinguishes an exploitable path from a single unexpected answer and records whether safeguards blocked, transformed, or permitted the requested action.
Evaluating reliability around the claimed use
Accuracy must be defined for the task. A classifier may require labeled examples, thresholds, subgroup review, and false-positive and false-negative rates. Retrieval testing can address source coverage, ranking, citations, and answer support. Generative testing can address factuality, human review, refusal behavior, and prompt sensitivity. GDF also reviews evaluation-set provenance, leakage risk, sample size, scoring, and deployment changes. Public benchmarks and vendor claims do not establish performance on the disputed population. The NIST AI Risk Management Framework is voluntary and does not itself establish compliance.
Model and training-data provenance
A claim about an LLM or its training data requires a defined model, version, provider, training stage, dataset, and asserted relationship to the disputed material. GDF can examine model cards, dataset records, manifests, licenses supplied for technical context, filtering and deduplication records, training or fine-tuning jobs, checkpoints, deployment identifiers, API records, and reproducible output tests where available. Similar output does not by itself establish that a work appeared in training data, and a provider's general statement does not establish the contents of a particular model version. Missing provider-controlled weights or corpora are stated as limits rather than filled with inference.
Provenance, synthetic content, and human contribution
For authorship or provenance, GDF traces inputs, outputs, and later edits through metadata, application history, session logs, content credentials, cloud records, local artifacts, and versions. Detection scores alone do not prove AI generation; the detector, population, threshold, and error rates must be known. Deepfake authentication requires examination of the original available file, encoding and edit history, container and stream metadata, content credentials, source-device or platform records, frame or signal discontinuities, and comparison material. Compression, transcoding, enhancement, and reposting can erase or imitate artifacts, so no single visual or detector cue controls the conclusion. File-level media acquisition and authentication are routed to GDF's media-authentication service, while the AI expert scope addresses the model, workflow, provenance claim, and technical limits. GDF does not decide copyrightability or ownership.
AI-assisted invention and software development
An AI-assisted development record may include prompts, generated code, commits, notebooks, experiments, model outputs, human revisions, tickets, and builds. GDF organizes that sequence, ties material to a person or system where supported, and tests whether a feature appears in available versions. USPTO guidance states that only natural persons can be inventors and the same inventorship standard applies to AI-assisted inventions. Technical records can document workflow and contributions, but counsel addresses inventorship, patentability, ownership, and infringement.
Five stages for an AI expert opinion
The process is designed around the deployed application and the claim being tested, not a generic model demonstration:
- Frame: confirm conflicts, party-retained role, disputed proposition, relevant system and period, legal assumptions from counsel, decision context, and expected expert use.
- Reconstruct: map the model, application, instructions, retrieval, tools, controls, data path, output handling, and human actions that could affect the result.
- Preserve: collect available prompts, context, sources, calls, parameters, outputs, logs, edits, approvals, versions, hashes, and provider notices with documented gaps.
- Evaluate: design task-specific tests, retain every material trial, measure relevant error and variability, test alternatives, and distinguish current behavior from historical evidence.
- Report and testify: connect opinions to preserved records and disclosed tests, prepare rebuttal and demonstratives, and explain uncertainty at deposition or hearing.
AI findings for report, deposition, hearing, and rebuttal
Deliverables may include a system and data-flow map, evidence inventory, event chronology, prompt and retrieval table, evaluation protocol, error analysis, provenance findings, technical report, declaration support, deposition exhibits, or demonstratives. The working record identifies scripts, queries, datasets, versions, parameters, exclusions, failures, and all material trial results. Confidential prompts, personal data, model artifacts, credentials, and case evidence are transferred through an approved controlled channel, never a public form.
Rebuttal analysis checks whether the opposing opinion identifies the actual system, preserves the complete context, uses a relevant evaluation population, discloses failed trials, distinguishes current from historical behavior, and accounts for human edits or downstream processing. GDF also reviews whether a detector's claimed error rates apply to the material tested. Testimony separates observed records, experimental results, technical inference, provider assertions, and legal assumptions supplied by counsel.
- AI system, data-flow, and evidence maps
- Prompt, retrieval, tool, output, and human-action chronologies
- Task-specific evaluation protocols and complete result sets
- Provenance analysis with detector and metadata limits
- Expert report, rebuttal analysis, deposition and hearing preparation, demonstratives, exhibits, and testimony
Limits belong in the opinion
Hosted systems may hide model weights, training corpora, service-side prompts, safety layers, routing, and historical versions. Logs may be incomplete, retrieval can change, and human review may occur elsewhere. These conditions can prevent reconstruction or narrow attribution. A current test cannot stand in for a prior version. Guidance, model cards, and governance frameworks can assess process but do not decide the law or facts. GDF does not present a probability or detector score as certainty.
AI intake: model, event, evaluation, and human review
Identify the parties, forum, deadlines, AI-enabled product, provider, relevant dates, disputed result, known version, available logs, evaluations, human-review process, and expected expert role. State the technical proposition to test and identify any source-code, synthetic-media, cloud, privacy, or regulated-system interface. Begin with conflict screening and scope. Do not send confidential prompts, datasets, outputs, credentials, personal data, code, or evidence through the public form. Counsel remains responsible for legal theories, privilege, discovery, and the governing standard.
Expert witness frequently asked questions
Can an AI detector prove that content was machine-generated?
No detector score should be treated as proof by itself. The relevant error rates, threshold, test population, content type, model coverage, preprocessing, and known edits must be examined. Metadata, application history, prompt records, stored outputs, cloud events, and version history can provide stronger provenance evidence when available.
Can GDF reproduce a disputed AI output?
GDF can test whether an output or behavior occurs under documented conditions. Exact reproduction may be prevented by nondeterministic sampling, hidden instructions, service routing, model updates, retrieval changes, tool results, or missing context. A later match shows possibility, not that the same process produced the historical artifact.
Can an AI expert evaluate bias or reliability?
A technical evaluation can measure defined performance, error rates, subgroup results, threshold effects, retrieval support, variability, or human-review outcomes on a relevant dataset. It cannot replace the legal definition of discrimination, liability, or compliance. Counsel supplies the governing legal questions and intended use of the results.
Is access to model weights or training data always required?
No. Some questions can be addressed through application records, input and output history, configuration, evaluations, repository evidence, provider documentation, and controlled testing. Other questions cannot be answered reliably without provider-controlled internals or training records. GDF states which conclusions the available access supports and which remain open.
Does alignment with the NIST AI Risk Management Framework prove compliance?
No. The framework is voluntary and supplies risk-management concepts rather than a legal conclusion. GDF can compare documented technical and operational practices with identified framework outcomes. Counsel determines whether a statute, regulation, contract, policy, or other authority applies and what compliance requires.
When should counsel retain an AI expert witness?
Early retention is useful before prompts, logs, retrieved content, model versions, evaluations, and human-review records change or expire. It also allows the parties to define a testable proposition, preserve the deployed-system context, agree on test materials, and identify separate source-code, cloud, media, privacy, or statistical expertise if needed.
How do jurisdiction and cost affect an AI expert assignment?
Forum requirements, provider-controlled access, model availability, dataset size, evaluation design, repeated trials, specialist media work, and testimony dates affect technical scope and cost. GDF can phase preservation, reconstruction, evaluation, reporting, and testimony around counsel's priorities. Counsel determines jurisdiction, discovery obligations, procedure, and legal strategy; GDF does not provide legal advice or promise a fixed cost before the deployed system and evidence are defined.