Device drift in edge AI creates legal exposure under SR 11-7 model risk management, FDA PCCP guidance, healthcare malpractice law, product liability failure to warn doctrine, and SEC securities disclosure obligations.
The Silent Degradation Problem A medical device company discovers that its edge-deployed AI diagnostic tool has been silently degrading for eleven months. The model’s accuracy against the clinical validation benchmark has dropped fourteen percentage points. No alert fired. No disclosure was made. Three thousand patient encounters occurred during the drift period. This is not a hypothetical. It is a description of the legal exposure that device manufacturers, financial institutions, and enterprise technology companies are accumulating right now as deployed AI systems drift away from their validated states without detection.
Device drift is the progressive degradation of a deployed AI model’s performance relative to its validated baseline. It occurs because the statistical distribution of real-world input data shifts over time away from the distribution on which the model was trained and validated. Edge-deployed models are particularly vulnerable because they receive live inference inputs continuously, they often cannot be retrained in place, and the feedback loops that would detect performance degradation in centralized systems are frequently absent at the edge. The legal consequences of undetected drift span multiple regulatory regimes, contractual obligations, tort liability, and securities disclosure requirements, and they are not being adequately addressed by current AI deployment practices.
SR 11-7 and Financial Services Model Risk Management The Federal Reserve and the Office of the Comptroller of the Currency addressed the model risk management obligations of supervised financial institutions in Supervisory Guidance SR 11-7. SR 11-7 requires that models be subject to ongoing performance monitoring, that performance results be compared against benchmarks and expectations, and that material deviations trigger review and corrective action. The guidance explicitly contemplates that models will degrade over time and that institutions must have processes to detect and respond to that degradation. Edge-deployed AI models used in credit underwriting, fraud detection, anti-money laundering monitoring, or market surveillance are subject to SR 11-7 if the deploying institution is a supervised entity. An institution that cannot demonstrate it is monitoring edge model performance against validated benchmarks on a defined schedule, documenting results, and escalating deviations to model risk management functions is operating out of compliance with SR 11-7, regardless of how sophisticated the original model validation was.
FDA PCCP Guidance and Medical Device Drift The FDA’s final guidance on Predetermined Change Control Plans, issued August 18, 2025, creates a parallel framework for medical AI devices. The PCCP guidance establishes a pathway for manufacturers of AI and machine learning-enabled medical devices to obtain premarket approval for a defined set of post-market model updates, including updates designed to address performance drift, without requiring a new premarket submission for each update. The practical significance is twofold. First, the guidance signals that the FDA expects manufacturers to anticipate drift and build validated update processes into their regulatory submissions from the outset, rather than treating drift as an unexpected post-market event. Second, a manufacturer who does not have an approved PCCP and whose device drifts materially from its cleared performance specifications may be distributing a device that no longer conforms to its clearance, triggering reporting obligations under 21 C.F.R. Part 803 and potential liability for distribution of an adulterated or misbranded device under the Food, Drug, and Cosmetic Act.
Healthcare Malpractice Exposure Healthcare malpractice exposure from device drift is distinct from regulatory exposure and in many cases more financially significant. When a physician relies on an AI diagnostic tool whose performance has degraded without disclosure, the physician has made a clinical decision based on materially inaccurate information. The question of whether the physician, the hospital, or the device manufacturer bears liability for the resulting harm is unsettled in most jurisdictions, but the trend in medical device liability litigation is toward treating AI-assisted clinical decisions as joint liability scenarios where the manufacturer’s contribution through failure to disclose known performance degradation is actionable. A manufacturer who monitors device performance, detects drift, and delays disclosure while patient encounters continue is in a materially worse legal position than one who detects drift and immediately triggers corrective action protocols.
NIST AI RMF Ongoing Monitoring Requirements The NIST AI Risk Management Framework addresses ongoing monitoring as a core component of responsible AI deployment. The AI RMF’s GOVERN, MAP, MEASURE, and MANAGE functions all contemplate that deployed AI systems will be subject to continuous performance assessment and that documentation of monitoring activities, anomaly detection, and corrective actions will be maintained throughout the system lifecycle. While the AI RMF is not itself a legally binding regulation, it is increasingly being incorporated by reference in procurement contracts, federal agency guidance, and state AI regulations as the standard of care for AI risk management. A company that cannot produce AI RMF-consistent monitoring documentation in response to a regulatory inquiry or civil discovery request faces a significant problem: the absence of documentation in this context is often treated as evidence that monitoring did not occur.
Contractual Performance Warranty Breach Contractual performance warranties in edge AI procurement agreements create independent liability exposure for device drift. Enterprise customers purchasing edge AI systems for industrial automation, quality control, logistics optimization, or other operational applications increasingly negotiate performance warranties that specify the model’s accuracy, precision, recall, or other metrics against defined test sets. When the deployed model drifts below warranted performance thresholds, the customer has a breach of warranty claim. The damages available under a warranty breach theory in a commercial contract can include the cost of remediation, lost operational efficiency during the drift period, and in some cases consequential damages for business losses attributable to degraded AI performance.
Product Liability for Failure to Warn Product liability exposure for failure to disclose foreseeable device drift follows from the principle that manufacturers have an obligation to warn of known and reasonably foreseeable product hazards. Device drift is not an unexpected failure mode: it is a well-documented characteristic of deployed machine learning systems that every competent AI engineer understands will occur over time. A manufacturer who deploys an AI system knowing that drift will occur, without implementing monitoring to detect it and disclosure mechanisms to inform users when it does, has failed to warn of a foreseeable hazard. In a strict products liability analysis, that failure to warn may be sufficient to establish liability for harm caused by degraded performance during the drift period, independent of any showing of manufacturing defect.
SEC Securities Disclosure Obligations Securities disclosure obligations attach when device drift becomes material to a public company’s business operations or financial condition. SEC disclosure obligations under Regulation S-K require disclosure of material risks and known trends that are reasonably likely to have a material effect on results of operations. An AI system whose degraded performance is affecting customer retention, contract compliance, or regulatory standing in a material way is a known trend that may require disclosure. A public company whose material AI systems are degrading without detection and without disclosure faces both the operational risk of the underlying degradation and the securities law risk that delayed disclosure will be characterized as a violation of Regulation S-K or Rule 10b-5.
The Organizational Accountability Gap The most consequential implication of the device drift problem is organizational rather than strictly legal. The companies most exposed to drift liability are not necessarily those with the worst technology. They are those where the organizational separation between AI deployment teams and legal, compliance, and regulatory teams is so wide that nobody with legal accountability for performance warranties, regulatory submissions, and disclosure obligations has visibility into what the deployed model is actually doing in the field. This is not a technology gap. It is an accountability gap, and it is endemic in organizations that treat AI deployment as an engineering matter rather than a cross-functional legal and compliance matter.
Building a legally defensible drift management program requires integrating several functions that typically operate in silos. Model performance monitoring must generate documented outputs at defined intervals that are reviewable by model risk management or quality assurance functions. Drift thresholds must be calibrated against regulatory requirements, warranty obligations, and clinical or operational risk tolerances. Disclosure protocols must specify who is notified when drift is detected, what information is provided to users, customers, and regulators, and on what timeline. For FDA-regulated devices, PCCP submissions must anticipate the update types that will be used to correct drift. For financial services firms, SR 11-7 model risk management frameworks must explicitly address edge-deployed models as a distinct monitoring category. The companies that build these programs now will be better positioned legally when the first wave of drift-related litigation and regulatory enforcement actions arrives.