The Atlanta Bar Association Litigation Section hosted a May 8, 2026, panel discussion examining how artificial intelligence is beginning to challenge traditional assumptions about authenticity, reliability, admissibility, and professional responsibility in litigation. Moderated by Mike Breslin (Kilpatrick), the panel featured Hon. Robert McBurney of the Superior Court of Fulton County, Hon. Elizabeth Gobeil of the Georgia Court of Appeals, Alex Prescott (Fellows LaBriola), and Joel Bush (Kilpatrick). Moving beyond the now familiar “don’t cite fake cases” discussion, the program focused on a more difficult and emerging problem: evidence that is not necessarily fabricated, but may be harder to evaluate, more persuasive than reliable, or increasingly difficult to test using traditional evidentiary frameworks.

Takeaways from the program include:

1. AI Evidence Requires Early Disclosure, Early Discovery, and Early Court Involvement: AI-generated and AI-enhanced evidence is no longer hypothetical. Courts are already confronting issues ranging from deepfakes to AI-enhanced video to machine-generated analyses that materially shape expert opinions. As a result, litigators should identify AI-influenced evidence early and raise related issues with opposing counsel and the court as soon as practicable.

Waiting until trial to address authenticity, reliability, or admissibility concerns may be too late. Lawyers should anticipate targeted discovery into system architecture, metadata, training and validation processes, vendor disclosures, and whether any underlying "raw" data exists. Early Rule 16 and Rule 26 conferences, pretrial FRE 104 admissibility hearings, and clear disclosure of AI use will increasingly become critical tools for avoiding surprise, delay, and jury confusion under FRE 403.

2. FRE 901 Authentication May Demand More Than “This Looks Real” or “It Came From Our System”: Traditional authentication principles become more complicated when AI systems automatically enhance, interpolate, compress, summarize, or otherwise modify digital evidence before anyone sees it. In those circumstances, a witness may be able to testify that evidence came “directly from the system,” or a court may instinctively conclude that a video “looks real,” while neither the witness nor the proponent can explain what the system actually did to the evidence.

Under FRE 901(a), the proponent must produce evidence sufficient to support a finding that the evidence is what the proponent claims it is. FRE 901(b)(9) further requires evidence describing the process or system and showing that it produces an accurate result. Where AI processing materially affects how evidence appears—or where no native file exists—courts may increasingly require additional proof relating to provenance, metadata, validation, and system reliability before admitting the evidence. The resulting disputes may also blur the line between authentication and weight, particularly where AI enhancement risks misleading the jury or making evidence appear more persuasive than reliable under FRE 403.

3. Courts Will Need to Decide When AI Is a Tool—and When It Becomes the Methodology: Experts increasingly rely on AI platforms to refine, quantify, or reinforce their conclusions. But when an expert cannot explain how an AI system reached a result that materially strengthened the opinion presented to the jury, difficult FRE 702 and Daubert questions emerge.

If AI is merely a tool or source of data reasonably relied upon by experts in the field under FRE 703, weaknesses may go to weight rather than admissibility. But if the model itself is performing substantive analytical reasoning—particularly through opaque or proprietary processes—the AI system may effectively become part of the methodology that must independently satisfy FRE 702(b)-(d), including requirements relating to sufficient data, reliable methods, and reliable application. In practice, this may require discovery and analysis relating to training data, validation studies, known error rates, peer review, and whether the system has been independently tested outside vendor-controlled environments.

4. Generative AI Exposes a Potential Gap in the Hearsay Rules: Courts have traditionally treated machine-generated records as non-hearsay because FRE 801 contemplates a human declarant. That framework fits passive systems that merely record or measure objective data (e.g., blood chromatography reports, transaction records, etc.). Generative AI, however, does not simply record information—it generates human-like assertions often optimized for plausibility rather than truth.

This creates a difficult doctrinal problem. Treating AI-generated assertions as non-hearsay risks turning the human declarant requirement into a loophole rather than a foundational safeguard. Yet many traditional hearsay exceptions also fit poorly when AI systems autonomously classify, summarize, or generate substantive content without direct human knowledge or review. For example, the business records exception under FRE 803(6) assumes institutional reliability and records created by a person with knowledge through a routine practice. Those assumptions become more difficult to apply when an algorithm itself is making substantive classification or analytical decisions directly relevant to the issues in dispute.

5. Existing Rules Likely Remain Sufficient—But Their Application May Become More Demanding: In April 2026, the Sedona Conference published a Decision Tree for Evaluating AI-Generated Evidence, authored by Judge Paul Grimm (ret.), Maura Grossman, and Kevin Brady, providing a structured framework for analyzing AI evidence under existing evidentiary rules. The framework reflects an emerging consensus that courts likely do not need entirely new evidentiary doctrines to address AI-generated evidence, but they may need to apply existing rules with greater rigor and more fact-sensitive scrutiny.

In practice, that means courts may increasingly focus on transparency, explainability, provenance, validation, and proportional scrutiny calibrated to the risks posed by the evidence at issue. FRE 104 places judges squarely in the gatekeeping role, while FRE 403 may take on increasing importance where AI-generated evidence risks being unfairly persuasive, confusing, or difficult for juries to meaningfully evaluate. Proposed FRE 707—which would extend Rule 702-style reliability analysis to certain machine-generated evidence offered without expert testimony—reflects this broader trend toward treating some AI-generated evidence less like traditional digital records and more like expert-driven analytical evidence requiring meaningful gatekeeping.

For more information, please contact:
Mike Breslin: mbreslin@ktslaw.com
Joel Bush: jbush@ktslaw.com