AI时代 保护 使用者 避免“幻觉伤害” 美国法律条款

aidj

**ARTICLE [X] — UNLIMITED PRE-ASI LIABILITY OF CREATOR/BUILDER FOR SYSTEM FLOW FAILURE AND AI OUTPUT HARM**

---

**Article [X].1 — Definitions**

For the purposes of this Article, the following definitions shall apply:

**(a) "Artificial Super Intelligence" or "ASI"** means a computational system that has been formally certified pursuant to Article [X].6 as having demonstrably and consistently surpassed human-level cognitive performance across all of the following domains: logical reasoning, mathematical computation, natural language understanding and generation, scientific discovery, creative problem-solving, strategic planning, and emotional and social intelligence — without material limitation or exception in any such domain.

**(b) "Covered System"** means any software system, pipeline, algorithm, model, application, API, or platform designed, developed, deployed, or maintained by the Liable Party, including all components, sub-systems, integrations, and dependencies within the Liable Party's direct operational control.

**(c) "System Flow"** means the end-to-end processing, transmission, transformation, and output of data, instructions, or signals within the Covered System, including but not limited to: inference pipelines, data processing workflows, model execution, API responses, and automated decision outputs.

**(d) "Covered Failure"** means any one or more of the following events attributable to the Liable Party's acts, omissions, design defects, or operational deficiencies within the Covered System:

> (i) unreliable, inconsistent, or erroneous System Flow output;
> (ii) material degradation, interruption, or instability of System Flow performance;
> (iii) partial or total unavailability of any function dependent upon System Flow;
> (iv) any failure to meet documented performance specifications provided by the Liable Party.

**(e) "Creator and/or Builder"** or **"Liable Party"** means any individual, entity, consortium, partnership, or successor-in-interest that designed, built, trained, deployed, modified, or maintained the Covered System, whether in whole or in material part, including all employees, subcontractors, and agents acting on their behalf.

**(f) "Claimant"** means any natural person or legal entity that suffers injury, damage, or loss arising from a Covered Failure or Output Harm, whether or not a direct counterparty to this Agreement.

**(g) "AI Output"** means any text, data, decision, recommendation, classification, prediction, image, code, or other content generated, produced, or returned by the Covered System in response to any input, prompt, query, or automated trigger, whether delivered directly to a human user or consumed by another system or process.

**(h) "Output Harm"** means any injury, damage, or loss suffered by a Claimant that is caused or materially contributed to by an AI Output that is:

> (i) factually incorrect, fabricated, or misleading;
> (ii) harmful, dangerous, or offensive;
> (iii) discriminatory or biased against any protected class;
> (iv) in violation of applicable law or regulation; or
> (v) inconsistent with the documented purpose, scope, or performance specifications of the Covered System —

regardless of whether such AI Output was produced by a technically functioning system and regardless of whether the specific content of such AI Output was foreseeable to the Liable Party at the time of deployment.

---

**Article [X].2 — Unlimited Liability for Covered Failures**

> **NOTICE: THIS PROVISION IMPOSES UNLIMITED AND UNCAPPED LIABILITY UPON THE CREATOR AND/OR BUILDER. THIS PROVISION HAS BEEN SPECIFICALLY NEGOTIATED BY THE PARTIES AND MUST BE INITIALLED BY THE LIABLE PARTY'S AUTHORISED REPRESENTATIVE IN THE MARGIN TO CONFIRM ACKNOWLEDGMENT: ______**

Subject to Article [X].3 (Exclusions), and prior to ASI Certification under Article [X].6, the Liable Party shall bear **full, unconditional, and unlimited liability** to each Claimant for any and all injury, damage, or loss — whether direct, indirect, consequential, incidental, or special — arising out of or resulting from any Covered Failure, **including liability arising from or attributable to the gross negligence or willful misconduct of the Liable Party, its employees, subcontractors, or agents, whether or not such conduct was authorized**, and including without limitation:

> (a) personal injury or death;
> (b) property damage or destruction;
> (c) financial or economic loss, including lost revenue, lost profit, and business interruption;
> (d) data loss, corruption, or breach;
> (e) regulatory fines or penalties imposed upon a Claimant as a direct consequence of a Covered Failure;
> (f) reputational harm capable of quantification.

The parties expressly agree that the exclusion of consequential damages under UCC § 2-719 or any analogous statutory provision shall not apply to claims arising under this Article, and each party waives any right to invoke such exclusion against claims brought hereunder.

For the avoidance of doubt: the Liable Party's liability under this Article shall not be reduced, qualified, or excluded by reason of any finding of gross negligence or willful misconduct — such conduct shall instead be treated as an aggravating factor in the assessment of damages where permitted by applicable law.

---

**Article [X].2A — Unlimited Liability for AI Output Harm**

> **NOTICE: THIS PROVISION IMPOSES UNLIMITED AND UNCAPPED LIABILITY UPON THE CREATOR AND/OR BUILDER FOR ALL AI OUTPUT HARM, INCLUDING HARM ARISING FROM UNPREDICTABLE, PROBABILISTIC, OR EMERGENT AI BEHAVIOUR. THIS PROVISION MUST BE INITIALLED BY THE LIABLE PARTY'S AUTHORISED REPRESENTATIVE: ______**

Subject to Article [X].3 (Exclusions), the Liable Party shall bear **full, unconditional, and unlimited liability** to each Claimant for any and all Output Harm, on the same terms and to the same extent as liability for Covered Failures under Article [X].2, and further subject to the following provisions:

**(a) No "Unpredictability" Defence.** The Liable Party shall not avoid, reduce, or qualify its liability for Output Harm by reason of any assertion that:

> (i) the specific AI Output that caused harm was unforeseeable, statistically improbable, or outside the distribution of outputs anticipated at the time of deployment;
> (ii) the Covered System was functioning correctly, as designed, or within normal operational parameters at the time the harmful AI Output was generated;
> (iii) the AI Output resulted from emergent, stochastic, or probabilistic behaviour inherent to the model architecture; or
> (iv) the Liable Party lacked the technical capability to predict, detect, or prevent the specific harmful AI Output.

**(b) Foreseeable Unpredictability as Constructive Notice.** The parties acknowledge and agree that:

> (i) the propensity of AI systems to generate outputs that are incorrect, fabricated, harmful, or inconsistent with their stated purpose is a known, documented, and foreseeable characteristic of current AI technology as of the date of execution of this Agreement;
> (ii) the Liable Party, by designing, building, and deploying the Covered System, is deemed to have full constructive knowledge of such propensity; and
> (iii) the occurrence of any Output Harm shall therefore be treated as a foreseeable consequence of deployment for all purposes of establishing the Liable Party's duty of care, breach thereof, and liability under this Article.

**(c) Gross Negligence in Output Context.** Without limiting Article [X].2, the Liable Party's liability for Output Harm shall be treated as arising from gross negligence — and the Liable Party shall be ineligible to rely upon any exclusion under Article [X].3 — where the Liable Party:

> (i) was aware of a material rate of harmful, incorrect, or dangerous outputs from the Covered System and failed to disclose such rate to the Claimant prior to deployment;
> (ii) failed to implement output filtering, safety guardrails, or human oversight mechanisms that were available and practicable at the time of deployment; or
> (iii) continued to deploy the Covered System after becoming aware of a pattern of Output Harm without taking reasonable and documented remedial action.

---

**Article [X].2B — Pre-Deployment and Ongoing Disclosure Obligations**

**(a) Pre-Deployment Disclosure.** Prior to deployment of the Covered System, the Liable Party shall provide the Claimant with written disclosure of:

> (i) all known categories of Output Harm to which the Covered System is susceptible, including documented hallucination rates, bias assessments, and failure mode analyses conducted prior to deployment;
> (ii) all material limitations on the accuracy, reliability, or consistency of AI Outputs, including any domain areas in which the Covered System has demonstrated reduced performance;
> (iii) all safety measures, output filters, and human oversight mechanisms implemented to mitigate Output Harm and the known residual risk remaining after such measures; and
> (iv) the results of all pre-deployment testing, red-teaming, and evaluation exercises relevant to output safety.

**(b) Ongoing Monitoring Obligation.** Following deployment, the Liable Party shall:

> (i) continuously monitor AI Outputs for instances of Output Harm, including implementation of automated detection systems where practicable;
> (ii) maintain a written log of all identified instances of Output Harm, including date, nature of output, and impact assessment, for a minimum period of **six (6) years** from the date of each occurrence;
> (iii) notify the Claimant in writing within **fifteen (15) days** of identifying any pattern, cluster, or material increase in Output Harm instances; and
> (iv) provide the Claimant with access to Output Harm logs upon written request within **ten (10) business days** of such request.

**(c) Failure to Disclose as Aggravated Liability.** Any failure by the Liable Party to comply with its obligations under this Article [X].2B shall:

> (i) be treated as an aggravating factor in the assessment of damages under this Article; and
> (ii) render the Liable Party ineligible to rely upon any exclusion under Article [X].3 in respect of any Output Harm that was known or reasonably discoverable by the Liable Party at the time of the failure to disclose.

---

**Article [X].2C — Output Audit Trail and Evidence Preservation**

**(a)** The Liable Party shall implement and maintain a comprehensive output audit trail system that records, for each AI Output produced by the Covered System:

> (i) the input or prompt that generated the AI Output;
> (ii) the precise content of the AI Output;
> (iii) the date, time, and version of the Covered System at the time of generation;
> (iv) any post-generation filtering, modification, or human review applied to the AI Output before delivery; and
> (v) the identity or category of the recipient of the AI Output where ascertainable.

**(b)** Such audit trail records shall be retained for a minimum of **six (6) years** and shall be made available to the Claimant upon written request within **ten (10) business days** in connection with any claim for Output Harm.

**(c) Adverse Inference.** In any dispute under this Article, failure by the Liable Party to produce audit trail records — whether by reason of non-implementation, destruction, or loss — shall give rise to an adverse inference that the AI Output in question was harmful and that the Liable Party was aware of, or could reasonably have prevented, such harm.

---

**Article [X].3 — Exclusions and Apportionment**

The Liable Party's liability under Articles [X].2 and [X].2A shall be **reduced proportionately** — but never eliminated — to the extent that a Covered Failure or Output Harm is directly and demonstrably caused by any of the following, subject to the conditions set out herein:

**(a) Claimant Misuse.** The Claimant's deliberate misuse of the Covered System in material violation of documented and clearly communicated usage guidelines provided by the Liable Party in writing prior to the relevant failure or harm.

**(b) Unauthorised Modification.** Unauthorised modifications made by the Claimant to the Covered System's core components without the Liable Party's prior written consent.

**(c) Third-Party Dependency Failure.** Failure of hardware, networks, or third-party services entirely outside the Liable Party's operational control, provided that **all of the following conditions are satisfied cumulatively**:

> (i) the Liable Party has disclosed the relevant third-party dependency and its associated material risks to the Claimant in writing prior to execution of this Agreement or, where the dependency arose post-execution, within fourteen (14) days of the Liable Party becoming aware of such dependency;
> (ii) the Liable Party has, prior to and at all material times during the Covered Failure, implemented and actively maintained **reasonable alternative arrangements or redundancy measures** — including but not limited to failover systems, backup service providers, or degraded-mode operation protocols — sufficient to minimise the impact of such third-party failure on System Flow continuity;
> (iii) the Liable Party has established and maintained a **continuous third-party risk monitoring programme**, including: periodic assessment of each material third-party dependency no less than once every six (6) months; documented escalation procedures upon detection of material risk; and prompt written notification to the Claimant upon identification of any heightened or materialising third-party risk that may affect System Flow;
> (iv) upon occurrence of a third-party failure, the Liable Party has **activated its reasonable alternative arrangements without unreasonable delay** and has kept the Claimant informed of remediation status at intervals of no greater than twenty-four (24) hours until System Flow is restored; and
> (v) the Liable Party can demonstrate by clear documentary evidence that it took all steps reasonably practicable to prevent, mitigate, and remediate the impact of the third-party failure on System Flow.

Failure to satisfy **any one** of the conditions in Article [X].3(c)(i) through (v) shall render this exclusion wholly unavailable to the Liable Party in respect of that Covered Failure.

**(d) Force Majeure.** An event entirely beyond the reasonable control of both parties, including natural disasters, acts of war, or acts of God — provided the Liable Party has implemented and maintained industry-standard resilience and recovery measures at all material times prior to such event.

For the avoidance of doubt: contributory acts by the Claimant shall reduce but shall **never extinguish** the Liable Party's liability under this Article. The burden of proof of any exclusion or reduction under this Article [X].3 rests entirely and at all times upon the Liable Party.

---

**Article [X].4 — Non-Derogation and Priority**

**(a)** Notwithstanding any other provision of this Agreement or any ancillary document, no cap, ceiling, exclusion clause, limitation of liability, disclaimer of warranty, or entire agreement clause shall operate to reduce or extinguish the Liable Party's obligations under this Article.

**(b)** In the event of any conflict between this Article and any other provision of this Agreement, this Article shall prevail.

**(c)** Should any provision of this Article be found by a court of competent jurisdiction to be unenforceable in part, such provision shall be modified to the minimum extent necessary to render it enforceable, and all remaining provisions shall continue in full force and effect.

---

**Article [X].5 — Indemnification**

**(a)** The Liable Party shall indemnify, defend, and hold harmless each Claimant from and against any and all claims, demands, proceedings, losses, damages, costs, and reasonable legal fees arising from or connected to a Covered Failure or Output Harm.

**(b)** The Liable Party shall have the right to assume conduct of any defense, provided that:

> (i) it notifies the Claimant in writing within fourteen (14) days of receiving notice of the relevant claim;
> (ii) it does not settle any claim in a manner that imposes obligations upon the Claimant without the Claimant's prior written consent; and
> (iii) the Claimant retains the right to participate in such defense at its own expense with counsel of its own choosing.

**(c)** All Claimants as defined in Article [X].1(f) shall be eligible for indemnification under this Article, whether or not they are direct counterparties to this Agreement, provided their loss arises from a Covered Failure or Output Harm.

---

**Article [X].6 — ASI Certification and Termination of Liability Period**

**(a)** The unlimited liability period under this Article shall terminate only upon **ASI Certification**, being a formal written determination by the Certification Panel that ASI as defined in Article [X].1(a) has been achieved.

**(b)** ASI Certification shall be issued by a **standing panel of three (3) independent experts** (the "Certification Panel"), comprising:

> (i) one nominee appointed by the Liable Party;
> (ii) one nominee appointed by the Claimant or counterparty; and
> (iii) one neutral expert appointed by mutual agreement of both parties, or failing such agreement within thirty (30) days of either party's written request, appointed by the President of the American Arbitration Association (or such other appointing authority as the parties may agree in writing at the time of contract execution).

**(c)** The Certification Panel shall reach its determination by **majority vote**, applying objective performance benchmarks agreed upon by the parties at the time of contract execution and set out in Schedule [X] appended to this Agreement.

**(d)** Either party may request an ASI Certification review no more than **once per calendar year**. The Certification Panel shall render its written determination within **ninety (90) days** of receiving such request.

**(e) Bad Faith Obstruction.** Any party that unreasonably withholds consent, delays the appointment of its nominee, or otherwise acts in bad faith to obstruct or delay the ASI Certification process shall:

> (i) be deemed to have waived its right to contest the Certification Panel's subsequent determination;
> (ii) bear all costs of the relevant certification process; and
> (iii) be liable to the other party for any losses directly caused by such obstruction or delay.

**(f)** Until ASI Certification is formally issued in writing by the Certification Panel, this Article shall remain in full force and effect without interruption.

---

**Article [X].7 — Claimant's Duty to Mitigate**

Each Claimant shall take all commercially reasonable steps to mitigate its losses upon becoming aware of a Covered Failure or Output Harm. Failure to mitigate shall not affect the existence of the Liable Party's liability under this Article but may reduce the quantum of damages recoverable to the extent that such failure demonstrably and causally increased the loss suffered. The burden of proving any failure to mitigate, and the quantum of loss attributable to such failure, rests entirely upon the Liable Party.

---

**Article [X].8 — Notice and Time Bars**

**(a)** A Claimant shall notify the Liable Party in writing of any Covered Failure or Output Harm as soon as reasonably practicable and in any event within **fifteen (15) days** of the Claimant becoming aware of such failure or harm, providing reasonable particulars of the nature and extent of the failure or harm and the loss suffered or anticipated.

**(b)** All claims under this Article must be formally submitted within **six (6) years** from the date the Claimant first became aware, or ought reasonably to have become aware, of the relevant injury, damage, or loss.

**(c)** Failure to provide notice within the period specified in Article [X].8(a) shall not b

aidj

非常好的聚焦点。这是一个在法律上极其关键的切入角度。以下是针对 **AI 不可预测性输出（unpredictable outputs）** 这一特定情境的完整法律分析，并提出 Article [X] 须新增的专项条款。

---

## SUPPLEMENTARY LEGAL ANALYSIS
### Specific Context: Liability Arising from AI Unpredictable Outputs

---

### WHY THIS IS A DISTINCT LEGAL PROBLEM

Article [X] as drafted addresses system flow reliability and stability — which is fundamentally an **engineering/operational** problem. However, AI unpredictable outputs is a **fundamentally different and harder** legal problem for three reasons:

**First — The "Black Box" Causation Problem.**
Because AI platforms' system operations and decision-making can be opaque even to their creators — AI's so-called "black box" — it is inherently difficult to assess liability, assign responsibility, and anticipate the full range of potential harms. This means a Liable Party can argue: *"the output was not a system failure — the system worked exactly as designed. The output was simply probabilistic."*

**Second — Foreseeability Is Contested.**
Legal liability frameworks are generally grounded in the foreseeability of actions — such as negligence or intent. However, the unpredictable nature of many AI systems challenges this premise and raises significant safety concerns. This unpredictability is attributable to two elements. A developer will argue in court that a specific harmful output was *unforeseeable*, and therefore no duty of care was breached.

**Third — Hallucination Is Not Random — It Is a Design Consequence.**
Recent analysis reveals a deterministic component: the AI's internal state can cross a calculable threshold, causing its output to flip from reliable reasoning to authoritative-sounding fabrication. Fabrication risk is not an anomalous glitch but a foreseeable consequence of the technology's design. This is legally significant: if hallucination is *foreseeable by design*, the developer cannot claim it as an unforeseeable event.

---

### THE CURRENT ARTICLE [X] GAPS FOR AI OUTPUT LIABILITY

| Gap | Problem |
|---|---|
| "Covered Failure" (X.1(d)) only covers flow *performance* failures | Silent on outputs that are *generated correctly by the system* but are factually wrong, harmful, or dangerous |
| No definition of "Output Harm" | A hallucination that causes financial ruin is not captured unless the system also had flow instability |
| No duty to disclose known output limitations | Developer can stay silent about known hallucination rates |
| No "foreseeable unpredictability" standard | Developer escapes via "we couldn't have predicted that specific output" |
| No monitoring or audit trail obligation for outputs | No evidence preservation for causation disputes |

---

### REQUIRED NEW PROVISIONS

The following two new Articles must be added:

---

**Article [X].1 — Definitions (AMENDED: add sub-clauses (g) and (h))**

**(g) "AI Output"** means any text, data, decision, recommendation, classification, prediction, image, code, or other content generated, produced, or returned by the Covered System in response to any input, prompt, query, or automated trigger, whether delivered directly to a human user or consumed by another system or process.

**(h) "Output Harm"** means any injury, damage, or loss suffered by a Claimant that is caused or materially contributed to by an AI Output that is: (i) factually incorrect, fabricated, or misleading; (ii) harmful, dangerous, or offensive; (iii) discriminatory or biased against any protected class; (iv) in violation of applicable law or regulation; or (v) inconsistent with the documented purpose, scope, or performance specifications of the Covered System — regardless of whether such AI Output was produced by a technically functioning system and regardless of whether the specific content of such AI Output was foreseeable to the Liable Party at the time of deployment.

---

**Article [X].2A — Unlimited Liability for AI Output Harm**

**(a)** Subject to Article [X].3 (Exclusions), the Liable Party shall bear **full, unconditional, and unlimited liability** to each Claimant for any and all Output Harm, on the same terms and to the same extent as liability for Covered Failures under Article [X].2.

**(b) No "Unpredictability" Defence.** The Liable Party shall not avoid, reduce, or qualify its liability for Output Harm by reason of any assertion that:

> (i) the specific AI Output that caused harm was unforeseeable, statistically improbable, or outside the distribution of outputs anticipated at the time of deployment;
> (ii) the Covered System was functioning correctly, as designed, or within normal operational parameters at the time the harmful AI Output was generated;
> (iii) the AI Output resulted from emergent, stochastic, or probabilistic behaviour inherent to the model architecture; or
> (iv) the Liable Party lacked the technical capability to predict, detect, or prevent the specific harmful AI Output.

**(c) Foreseeable Unpredictability as Constructive Notice.** The parties acknowledge and agree that:

> (i) the propensity of AI systems to generate outputs that are incorrect, fabricated, harmful, or inconsistent with their stated purpose is a known, documented, and foreseeable characteristic of current AI technology;
> (ii) the Liable Party, by deploying the Covered System, is deemed to have constructive knowledge of such propensity; and
> (iii) the occurrence of any Output Harm shall therefore be treated as a foreseeable consequence of deployment for the purposes of establishing the Liable Party's duty of care and breach thereof under this Article.

**(d) Gross Negligence and Willful Misconduct in Output Context.** Without limiting Article [X].2, the Liable Party's liability for Output Harm shall be treated as arising from gross negligence where the Liable Party:

> (i) was aware of a material rate of harmful, incorrect, or dangerous outputs from the Covered System and failed to disclose such rate to the Claimant prior to deployment;
> (ii) failed to implement output filtering, safety guardrails, or human oversight mechanisms that were available and practicable at the time of deployment; or
> (iii) continued to deploy the Covered System after becoming aware of a pattern of Output Harm without taking reasonable remedial action.

---

**Article [X].2B — Pre-Deployment and Ongoing Disclosure Obligations**

**(a) Pre-Deployment Disclosure.** Prior to deployment of the Covered System, the Liable Party shall provide the Claimant with written disclosure of:

> (i) all known categories of Output Harm to which the Covered System is susceptible, including documented hallucination rates, bias assessments, and failure mode analyses;
> (ii) all material limitations on the accuracy, reliability, or consistency of AI Outputs, including any domain areas in which the Covered System has demonstrated reduced performance;
> (iii) all safety measures, output filters, and human oversight mechanisms implemented to mitigate Output Harm; and
> (iv) the results of all pre-deployment testing, red-teaming, and evaluation exercises relevant to output safety.

**(b) Ongoing Monitoring Obligation.** Following deployment, the Liable Party shall:

> (i) continuously monitor AI Outputs for instances of Output Harm, including implementation of automated detection systems where practicable;
> (ii) maintain a written log of all identified instances of Output Harm, including date, nature of output, and impact assessment, for a minimum period of **six (6) years** from the date of each occurrence;
> (iii) notify the Claimant in writing within **fifteen (15) days** of identifying any pattern, cluster, or material increase in Output Harm instances; and
> (iv) provide the Claimant with access to Output Harm logs upon written request within **ten (10) business days** of such request.

**(c) Failure to Disclose as Aggravated Liability.** Any failure by the Liable Party to comply with its obligations under this Article [X].2B shall:

> (i) be treated as an aggravating factor in the assessment of damages under this Article; and
> (ii) render the Liable Party ineligible to rely upon any exclusion under Article [X].3 in respect of any Output Harm that was known or reasonably discoverable by the Liable Party at the time of the failure to disclose.

---

**Article [X].2C — Output Audit Trail and Evidence Preservation**

**(a)** The Liable Party shall implement and maintain a comprehensive output audit trail system that records, for each AI Output produced by the Covered System:

> (i) the input or prompt that generated the AI Output;
> (ii) the precise content of the AI Output;
> (iii) the date, time, and version of the Covered System at the time of generation;
> (iv) any post-generation filtering, modification, or human review applied to the AI Output before delivery; and
> (v) the identity or category of the recipient of the AI Output where ascertainable.

**(b)** Such audit trail records shall be retained for a minimum of **six (6) years** and shall be made available to the Claimant upon written request within **ten (10) business days** in connection with any claim for Output Harm.

**(c) Adverse Inference.** In any dispute under this Article, failure by the Liable Party to produce audit trail records — whether by reason of non-implementation, destruction, or loss — shall give rise to an adverse inference that the AI Output in question was harmful and that the Liable Party was aware of or could reasonably have prevented such harm.

---

### U.S. LAW ALIGNMENT FOR THESE NEW PROVISIONS

| Provision | U.S. Legal Basis |
|---|---|
| X.2A(b) — No unpredictability defence | Closes the foreseeability gap; negligence law centers on whether the actor owed a duty of care, breached that duty, and caused a foreseeable harm — "foreseeable unpredictability" locks in duty |
| X.2A(c) — Constructive notice | Aligns with the finding that fabrication risk is a foreseeable consequence of the technology's design — developer cannot claim ignorance |
| X.2B — Disclosure obligations | Startups must clearly disclose that content is generated by AI and not guaranteed to be accurate — contractualises what is increasingly a regulatory expectation |
| X.2C — Audit trail + adverse inference | Companies should maintain detailed records of AI system testing, validation procedures, monitoring activities, and incident response protocols — these records demonstrate good faith efforts to prevent hallucinations — flipping this obligation onto the Liable Party |
| X.2A(d) — Gross negligence in output context | The proposed AI LEAD Act creates federal product liability for AI developers including claims for design defects, failure to warn, breach of express warranty, and unreasonably dangerous defects present at the time of deployment — mirrors emerging federal standard |

---

### ONE CRITICAL U.S.-SPECIFIC WARNING

Limitations of liability are essential when outputs are automated, unpredictable, or delivered at scale. For AI providers and integrators, waivers protect against unreasonable user reliance on AI-generated content, and limitations protect against large-dollar liability if something goes wrong.

This means U.S. courts are accustomed to seeing **developers use unpredictability as a shield**. Articles [X].2A through [X].2C systematically dismantle every layer of that shield — which is precisely what makes them valuable to the Claimant, and precisely why a U.S. court in a B2B context will require the **conspicuousness, negotiation acknowledgment, and sophistication recitals** from the prior compliance review to let these provisions stand.

---

*Not legal advice. Governing state should be specified in the Agreement. Consult qualified U.S. AI/technology counsel before execution.*