Your AI Is Getting Dumber Every Quarter. And You Are Liable for What It Decides.

Model collapse is no longer a lab phenomenon. It is a documented production reality, confirmed by Communications of the ACM in February 2026, appearing in commercial tools your organisation depends on daily. When the AI powering your consequential decisions quietly degrades — producing increasingly narrow, homogeneous, factually eroded outputs — the governance gap between what you believe the system can do and what it actually does becomes your liability. Here is the framework that closes it.

Model collapse is the compounding degradation that occurs when AI systems are trained — or fine-tuned — on increasingly synthetic, AI-generated, or low-diversity data. The result: outputs that are fluent, confident, and systematically wrong. Rare cases vanish. Minority viewpoints disappear. Factual accuracy erodes while the model’s surface-level competence remains intact. A February 2026 Communications of the ACM article confirmed this is happening in commercial production systems today. The legal exposure is specific: organisations relying on AI outputs in consequential decisions — employment, credit, legal analysis, medical triage, financial modelling — that cannot demonstrate ongoing performance monitoring face simultaneous product liability exposure under the EU PLD (Brief 11), EU AI Act post-market monitoring violations from August 2, and Caremark board governance liability for deploying AI without the oversight infrastructure to detect its degradation. The governance protocol that prevents collapse from becoming liability is the Model Integrity Operating System — five components, 60-day build, applicable to every AI system in consequential deployment. The free tier names the collapse patterns and their legal triggers. The paid tier delivers the full Protocol.

The Hidden Pattern

Picture this: Your legal team has been using an AI research assistant for eighteen months. It produces polished, well-structured memos. Citations are present. Tone is authoritative. Senior partners have come to trust it for preliminary issue-spotting and contract risk summaries.

Then a junior associate notices something. The AI’s position on a particular area of privacy law has not changed in six months — even as the regulatory landscape has shifted materially. The model keeps producing the same analysis. Confident. Fluent. Quietly obsolete. When she runs a systematic check, the pattern is broader: the system’s outputs have converged. Earlier memos explored edge cases, minority arguments, and circuit splits. Recent ones do not. The AI writes with the same authority but has lost the intellectual range that made it valuable.

This is model collapse in its most dangerous form. Not a crash. Not a hallucination event that triggers immediate alarm. A slow, invisible narrowing — the compounding approximation error that researchers documented and the February 2026 Communications of the ACM article confirmed is happening in commercial production systems. The model’s failure mode is not noise. It is false confidence. The outputs remain structured. The reasoning remains legible. The breadth has simply vanished.

Now add the legal dimension. That law firm’s advice letters to clients are based on analysis from a system whose performance has materially degraded. The firm cannot demonstrate it monitored the system’s accuracy over time. There is no version history, no benchmark comparison, no post-deployment testing cadence. If a client later challenges the advice, the question is not whether the AI failed. The question is whether the firm had a governance architecture capable of detecting the failure — and responded accordingly. The answer, in most firms deploying AI today, is no.

Model collapse does not announce itself. It compounds silently — each cycle of retraining on synthetic or recycled data producing outputs that are slightly less diverse, slightly less accurate, slightly more homogeneous than the last. The organisation that cannot detect this drift is not merely exposed to AI failure. It is exposed to the specific liability of governing a degraded system as if it were a functioning one.

The Cost of the Problem

Liability Trigger 1 — EU AI Act Post-Market Monitoring Failure

The EU AI Act’s Article 72 requires deployers and providers of high-risk AI systems to establish, document, and operate a post-market monitoring system — a structured mechanism for collecting and analysing data about the AI system’s real-world performance, including accuracy degradation, output diversity erosion, and safety-relevant deviations. From August 2, 2026, deploying a high-risk AI system without a functioning post-market monitoring protocol is a regulatory violation carrying penalties up to €15 million or 3% of global turnover per finding. Model collapse — the systematic degradation of output quality over time — is precisely the failure mode post-market monitoring is designed to detect. An organisation that cannot demonstrate ongoing monitoring has not merely failed a compliance checkbox. It has deployed a degrading system without the regulatory infrastructure to know it was degrading, and the regulator will treat those two facts as a single, compounding violation.

The question regulators will ask is not philosophical: it is operational. Show us your model performance baseline at deployment. Show us your quarterly performance assessment records. Show us the corrective actions you took when performance deviated from baseline. If those records do not exist, the system was not monitored. And an unmonitored high-risk AI system operating in consequential decision contexts from August 2, 2026 is a regulatory enforcement event waiting for a trigger.

Liability Trigger 2 — EU PLD Defectiveness Through Performance Degradation

Brief 11 established that the EU Product Liability Directive treats AI software as a product subject to no-fault strict liability from December 9, 2026. The PLD’s defectiveness standard is linked to EU AI Act compliance — non-compliance with the EU AI Act’s post-market monitoring requirements is evidence of product defectiveness under the PLD. A model that has collapsed — that is producing systematically narrowed, factually eroded outputs in consequential decision contexts — and whose deployer cannot demonstrate they monitored its performance, is not merely a failed AI system. It is a defective product. The harm it caused while operating in a degraded state is the strict liability claim. The absence of monitoring documentation is the evidence that the deployer knew the product could degrade and built no system to detect when it did.

Liability Trigger 3 — Caremark Board Governance Exposure

Brief 5 established that directors who approved an AI deployment without building a functioning oversight system face personal Caremark liability. Model collapse is the specific AI failure mode that the Caremark monitoring obligation is designed to catch. A board that approved an AI strategy, received no performance monitoring reports, and discovered the AI had been producing degraded outputs for six quarters has not governed AI. It has delegated AI — entirely, without oversight — and the Caremark doctrine treats that delegation as a board-level governance failure, not an operational one. The quarterly AI risk report format from Brief 5 is the monitoring record the board needs. An organisation whose quarterly board reports contain AI adoption metrics but no AI performance metrics has the paperwork of governance without the substance of it.

— Paid Subscribers Continue Here —

What follows is the complete Model Integrity Operating System — the baseline measurement protocol, the performance drift detection methodology, the corrective action framework, the board reporting integration, and the vendor accountability provisions that convert model collapse from an invisible liability into a managed, documented, defensible governance position.