海角黑料

1) Start by Classifying the AI Evidence

Most 鈥淎I evidence鈥� in litigation falls into one of four buckets. The bucket often determines which evidentiary rules do the heavy lifting:

• AI-generated content (e.g., the chatbot鈥檚 output offered for its truth).
• AI-assisted human content (e.g., a draft contract or memo created with AI assistance, later adopted/edited by a human).
• AI as part of a business system (e.g., an internal 鈥淪afetyBot鈥� that creates tickets, logs, and audit trails as part of routine operations).
• AI-altered or 鈥渄eepfake鈥� media (e.g., audio/video/image altered by AI, which intensifies authentication and prejudice concerns).

If you treat all four categories the same, you鈥檒l miss the best arguments鈥攅specially on authentication and Rule 702 reliability.

2) A Federal Rule Checklist for Chatbot Chats

In federal court, AI chat evidence almost always triggers a predictable checklist:

• Relevance: Rules 401 & 402 (and then Rule 403 balancing).
• Authentication: Rule 901 (and, increasingly, Rule 902(13)鈥�(14) certifications for electronic evidence).
• Hearsay: Rules 801鈥�803 (including 鈥減arty-opponent鈥� admissions, business records, etc.).
• Expert gatekeeping: Rule 702 if the chatbot output functions like technical or specialized opinion.
• Completeness / context: Rule 106 to prevent cherry-picked snippets.
• Best Evidence: Rules 1001鈥�1003 (especially where a screenshot is offered instead of the native export).
• Special exclusions when relevant: e.g., Rule 407 (subsequent remedial measures) in negligence cases.

3) Authentication: The 鈥淚s This What You Say It Is?鈥� Foundation

Authentication is usually the first real battleground with AI chats. Under Rule 901, the proponent must offer enough evidence for a reasonable factfinder to conclude
the exhibit is what the proponent claims it is (e.g., a true and accurate export of a particular user鈥檚 chat, from a specific account, on a particular date).

Common ways to authenticate a chat transcript

• Witness-with-knowledge testimony: a user, custodian, or investigator explains how the chat was created and preserved.
• Distinctive characteristics: account identifiers, timestamps, conversation continuity, internal references, and corroborating surrounding evidence.
• Process/system evidence: a qualified witness explains the export process and how the system reliably produces accurate logs.
• Rule 902(13)鈥�(14) certifications: for certain electronic records, a certification can substitute for live testimony (with proper notice).

Courts have long been skeptical of 鈥渋nternet printouts鈥� offered without meaningful foundation鈥攅specially when authorship or control is disputed. AI chat logs can face the same
vulnerability if the proponent can鈥檛 tie the chat to the right account and user with reliable metadata and testimony.

4) Hearsay: Whose 鈥淪tatement鈥� Is an AI Output?

Here is the core hearsay puzzle with chatbots: the rules define hearsay around a person鈥檚 assertion and a human declarant.
That means the human user鈥檚 prompt is usually easy to categorize (it鈥檚 a person鈥檚 statement), while the chatbot鈥檚 output is harder鈥攂ecause it may be treated as machine output,
or as an 鈥渆xpert-like鈥� opinion, or as a statement adopted by a party.

Practical framing options

• Non-hearsay use (often strongest): offer the AI output to show notice, knowledge, state of mind, or effect on the listener
  (e.g., 鈥渢he manager was warned and then decided to repair鈥�), rather than to prove the output was true.
• Party admissions: if a party鈥檚 employee typed the prompt, forwarded the output, or adopted it as accurate, portions of the exchange may qualify as admissions (or adoptive admissions).
• Business records: if the organization routinely keeps AI-chat logs as part of its operations (e.g., ticketing, compliance, incident reporting), the recordkeeping layer may fit a business-records theory鈥�
  but you still must confront reliability and purpose for which offered.
• Rule 702 (expert-like evidence): if you鈥檙e offering the chatbot鈥檚 output as technical truth (鈥渢he stairs were structurally unsafe鈥�), a judge may treat it as expert territory and require Rule 702 reliability.

Translation: the more the proponent uses AI output like an authoritative conclusion, the more likely the court is to demand robust foundation, context, and鈥攕ometimes鈥攅xpert testimony.

5) Rule 702: When AI Output 鈥淎cts Like鈥� Expert Testimony

If the AI output is being offered as specialized knowledge (medicine, engineering, finance, safety standards), it can collide with Rule 702.
That rule requires the proponent to show the opinion is helpful, rests on sufficient facts/data, and is reliably applied.

This is especially important because generative AI can produce fluent answers that are incomplete, wrong, or untraceable (鈥渉allucinations鈥�).
As a result, some judges and rulemakers have been actively debating whether to create a dedicated rule for AI-generated evidence that functions like expert testimony.

6) Best Evidence & Completeness: Screenshots Invite Rule 1002 / 106 Arguments

If the content of a chat is central, be ready for best-evidence and completeness disputes:

• Rule 106 (completeness): if one side offers a snippet, the other may demand surrounding context to avoid misleading impressions.
• Rules 1001鈥�1003 (best evidence / duplicates): a screenshot might be a 鈥渄uplicate鈥� or 鈥渟econdary鈥� representation鈥攆ine in many cases, but risky if authenticity is challenged.

In practice, the 鈥渂est鈥� version of an AI chat exhibit is usually: a native export + metadata + custodian declaration + a clear explanation of how the system logs were generated and preserved.

7) Worked Example: Introducing a 鈥淐hat Box鈥� Transcript at Trial

Hypothetical: Paul sues Dell鈥檚 Department Store in a slip-and-fall case, alleging a broken stair step caused his injury. Immediately after the incident, Dell鈥檚 manager Mark uses an internal AI assistant (鈥淪afetyBot鈥�) to create an incident ticket and get recommendations.

The SafetyBot chat (proposed exhibit)

Mark: 鈥淐ustomer fell near stairwell. Step 3 looked cracked/loose. Could that cause a fall?鈥�

SafetyBot (AI): 鈥淵es. A cracked/loose step is a common hazard that can cause falls. Recommend immediate repair and documenting condition. If there are repeated reports, consider closing the stairwell until fixed.鈥�

At trial, while Mark is on the stand, the following occurs:

1. Q: 鈥淩ight after Paul fell, you typed into SafetyBot that Step 3 looked cracked and loose鈥攃orrect?鈥�
   A: 鈥渊别蝉.鈥�
2. Q: 鈥淎nd SafetyBot responded that a cracked step can cause falls and recommended immediate repair鈥攃orrect?鈥�
   A: 鈥渊别蝉.鈥�
3. Q: 鈥淵ou forwarded that SafetyBot response to maintenance the same day鈥攔ight?鈥�
   A: 鈥渊别蝉.鈥�
4. Q: 鈥淎nd the stair was repaired the next day鈥攔ight?鈥�
   A: 鈥渊别蝉.鈥�

Plaintiff moves to admit: (A) a printed screenshot of the chat and (B) the native export plus audit logs.

A) Likely Federal (FRE) objections & rulings

1) Authentication (Rule 901 / 902)

Objection: 鈥淟ack of foundation / not authenticated.鈥�
Likely ruling: If plaintiff offers only a screenshot with no metadata and no witness who can explain where it came from and whether it was altered, the objection has traction.
If plaintiff offers a native export + audit logs + a custodian or Mark鈥檚 testimony tying it to his account and the system, authentication is much more likely to be satisfied.

2) Hearsay (Rules 801鈥�803)

Mark鈥檚 prompt: If Mark is an employee speaking within the scope of his duties, the prompt can be offered against Dell as a party-opponent statement (commonly via the agent/employee path).
SafetyBot鈥檚 output: Defense will argue it鈥檚 being offered for its truth (鈥渂roken step causes falls鈥�), and that the jury can鈥檛 cross-examine 鈥淪afetyBot.鈥�
Plaintiff鈥檚 cleanest response is often: 鈥淣ot for truth鈥攐ffered to show notice/knowledge and why Dell acted,鈥� which is a non-hearsay purpose.
If plaintiff insists it is offered for truth (i.e., as substantive proof of defect/causation), the judge may scrutinize whether this is really expert-like opinion requiring Rule 702 reliability.

3) Rule 702 (expert reliability) if offered as technical truth

Objection: 鈥淚mproper expert opinion / unreliable methodology.鈥�
Likely ruling: If the AI output is being used as a substitute for an engineer or safety expert, the court may exclude it absent a proper expert foundation.
If it鈥檚 used only to show Mark was advised and reacted, the Rule 702 pressure drops dramatically.

4) Rule 403 (unfair prejudice / misleading the jury)

Objection: 鈥淓ven if relevant, the AI output will mislead the jury into treating it as authoritative.鈥�
Likely ruling: Courts may consider limiting instructions, redactions, or allowing the evidence only for the non-hearsay purpose (notice) if the 鈥淎I authority effect鈥� risks confusion.

5) Subsequent remedial measures (Rule 407) as to the repair

Objection: 鈥淩epair the next day is a subsequent remedial measure.鈥�
Likely ruling: If plaintiff offers the repair to prove negligence or culpable conduct, Rule 407 typically bars it. If offered for another permitted purpose (e.g., ownership/control, feasibility if disputed, or impeachment), it may be allowed with care.

B) California (CEC) version: key objections and 鈥渆xemptions/exceptions鈥�

In California Superior Court, you鈥檙e still fighting the same war鈥攋ust with CEC section numbers:

1) Relevance & balancing (CEC 210 / 352)

The SafetyBot exchange is likely relevant if it tends to show notice, condition, or causation. Even relevant evidence can be excluded if its probative value is substantially outweighed by undue prejudice, confusion, or misleading the jury.

2) Authentication (CEC 1400 / 1401)

Plaintiff must authenticate the writing (the chat record) before it is received. A native export plus audit logs and testimony from Mark or a custodian generally beats a screenshot standing alone.

3) Hearsay rule (CEC 1200) and key 鈥渆xemptions/exceptions鈥�

• Party admissions (CEC 1220) and related admissions:
  Mark鈥檚 typed prompt may come in if it qualifies as an admission offered against the party (and related doctrines may apply depending on agency/authorization).
• Adoptive admissions (CEC 1221):
  If Dell (through Mark or another authorized actor) adopted or agreed with the AI output鈥攅.g., forwarding it with 鈥渢his is accurate鈥濃�攊t strengthens an admissions theory.
• Business records (CEC 1271):
  If Dell regularly keeps SafetyBot tickets/logs as part of routine operations and the foundational requirements are met, the recordkeeping layer may fit a business-records exception.
• Non-hearsay purpose:
  As in federal court, the cleanest move is often to offer the AI output to show notice/knowledge/effect on the listener rather than to prove the output鈥檚 truth.

4) Subsequent remedial measures (CEC 1151)

Evidence of the next-day repair is generally inadmissible to prove negligence or culpable conduct, though it may be permitted for other limited purposes in the right posture (ownership/control, impeachment, etc.).

5) 鈥淏est evidence鈥� concept in California: Secondary Evidence Rule (CEC 1521)

California鈥檚 approach is often framed through the Secondary Evidence Rule: the content of a writing may be proved by otherwise admissible secondary evidence,
but the court may exclude it in certain circumstances鈥攅specially where fairness or authenticity concerns are substantial.
Practically, a screenshot is more vulnerable than a native export with reliable metadata.

8) A Field Checklist: How to Make AI Chats Trial-Ready

• Preserve early: treat AI chats like ESI (litigation holds, retention policies, export procedures).
• Prefer native exports: capture full threads, timestamps, account IDs, and system audit trails.
• Lock down integrity: hashes/digital signatures where available; document chain of custody.
• Capture the 鈥減rompt + output鈥� together: context matters for meaning and for completeness objections.
• Decide your purpose: 鈥渘otice/effect on listener鈥� is often an easier admissibility path than 鈥渢ruth of the AI鈥檚 conclusion.鈥�
• Anticipate Rule 702 fights: if the AI output sounds like an expert, expect expert standards.
• Prepare limiting instructions: if admitted, frame the permitted use tightly (especially to avoid jurors treating AI as authoritative).

9) Bottom Line

Courts already have tools to handle AI chats: authentication rules, hearsay doctrine, expert gatekeeping, and prejudice balancing.
The winning strategy usually comes from choosing the cleanest theory for why the chat matters鈥攖hen building a record that proves authenticity and prevents the jury from over-weighting AI output.

Disclaimer: This article is for general informational purposes and is not legal advice.

Sources (for footnoting or 鈥淔urther Reading鈥�)