Stability AI

Users retain ownership of the Inputs they submit, subject to representing they have the necessary rights. This is a user-favorable position typical of generative AI tools.

Stability assigns to the user all of its right, title and interest in the Outputs, so the user owns the generated content to the extent permitted by law. The terms caveat that similar inputs may yield similar outputs to other users, limiting exclusivity.

Stability may use Inputs and Outputs to improve and develop its Services, including to train its models, but explicitly offers an opt-out. Training use is default-on, which is a moderate risk, but the presence of a clear opt-out mitigates it.

The current (July 31, 2025) privacy policy states retention is open-ended based on necessity, with no explicit schedule or deletion SLA. The older (April 25, 2024) privacy notice imposes a one-year cap. Deletion can be requested by email but no defined turnaround is given for general deletion, and no security-driven audit-log retention is described.

The current privacy policy discloses sharing with affiliates, service providers, cloud partners, and — for website visitors — advertising partners, analytics providers and social networks, which may constitute a CCPA 'sale/share.' Sharing is disclosed and an opt-out is offered where legally required, but advertising-partner sharing goes beyond core service delivery. The older privacy notice conflictingly claims no sales at all.

The policies provide multiple concrete opt-out mechanisms: an explicit opt-out from model training, marketing unsubscribe, targeted-advertising opt-out where legally required, honoring of Global Privacy Controls, and an arbitration opt-out within 30 days. This is a comparatively strong opt-out posture.

The supplied policy text names no security certifications or regulatory attestations. GDPR/UK GDPR and CCPA are addressed operationally (legal bases, SCCs, adequacy decisions, GPC), but no SOC 2, ISO, or NIST framework is claimed within the accessible documents. A SOC 2 Type II announcement URL was listed but inaccessible, so it cannot be credited as evidence. Against the universal baseline plus consumer/developer/enterprise frameworks, the absence of any named certification is a significant gap.

The documents provide no model explainability commitments, no enterprise audit capabilities, and no transparency reporting. They caution that outputs may be inaccurate and require human review, but offer no auditability of model behavior. Corporate-domain accounts allow admin monitoring of Inputs/Outputs, but that is account oversight, not model auditability.

The documents describe only generic 'appropriate technical and organizational security measures' with no specifics on encryption at rest/in transit, penetration testing, bug bounty, or incident response. No breach history is disclosed. No trust center or security page content is present in the supplied text, and the vendor explicitly disclaims any guarantee of security.

The consumer terms note that corporate-domain accounts may be linked to an enterprise account where an administrator can view Inputs and Outputs, and reference separate supplemental/enterprise terms, but no enterprise agreement was supplied. Data-handling differences between free and paid tiers are not defined in the accessible text, so the delta cannot be confirmed.

Stability reserves the right to preserve and disclose text prompts and Content for legal and safety purposes, and enterprise-account admins may view Inputs and Outputs. While no explicit routine human-review-for-training clause appears, the reserved rights to access prompts/outputs for safety and legal reasons are broad.

The documents disclose that Stability may comply with government, court, and law enforcement requests and may disclose data in disputes or litigation. Mandatory binding arbitration with a class-action waiver applies to US/Canada users. No specific pending litigation is disclosed. These provisions shift dispute-resolution power toward the vendor.

Stability collects name, email, payment information (via Stripe), IP address, device identifiers, location data (potentially precise via GPS), usage/log data, and the content of Inputs and Outputs. The current policy states it does not process sensitive personal data to infer characteristics, and one older notice states no SPI is processed — yet precise geolocation and free-form Input content could contain SPI. Collection is disclosed but broad, warranting a YELLOW.

The current Terms and Privacy Policy are dated effective July 31, 2025 — within 12 months of the August 2026 analysis date — and they do address AI/ML processing, model training, third-party models/APIs, and integrations, matching the product surface (Stable Audio, Brand Studio, API). However, the supplied set also includes stale documents (an April 25, 2024 privacy notice and a December 14, 2022 API Terms) that remain live, creating currency ambiguity. Because multiple conflicting-vintage documents coexist, the rating is capped at YELLOW.

Multiple documents were supplied and they materially contradict one another. The April 25, 2024 privacy notice states no sensitive information is processed, no third-party data is received, and a one-year retention cap, while the July 31, 2025 privacy policy states it receives information from other sources, shares data with advertising partners (a CCPA 'sale/share'), and uses open-ended retention. These are material differences that would change a user's decision depending on which document they read.