The scoring system follows a comprehensive guide created specifically for this project, accessible here, and evolves dynamically with innovations and feedback from the decentralized community.

This nilGPT Privacy Data Review reinforces our mission to promote transparency, data protection, and digital sovereignty for AI users. Our goal remains clear: to enlighten and inform, without filter or influence, so we can build together a fairer and more transparent AI ecosystem.

update : 25/09/21

Key Findings from the nilGPT Privacy Data Review

Model

Meta-Llama-3.1-8B currently, with experimental DeepSeek integration; roadmap towards Llama-3.1-70B and Gemma-3-27B-IT in the coming weeks.

Data Collection

Prompts stored: Chat histories are securely stored in nilDB, a decentralized MPC-based storage network, where no data can be reconstructed or leaked even in case of compromise. A

Use for training: No, inputs/outputs are not used to train any AI model, ensuring data confidentiality. A

Account required: Registration via email or wallet. Pseudonymity is possible, but an account is always required. B

Data retention duration: Data is stored in nilDB and fully controlled by the user. Since September 2025, in-app deletion ensures immediate and user-managed data removal. A

User Control

Deletion possible: Yes, in-app deletion is now available (September 2025 release), while export remains limited to manual requests. A

Export possible: Users can request full access and data portability in compliance with GDPR. Although export is handled via support, it guarantees users’ right to recover their data. A

Granularity control: There are no fine-grained controls within the app to choose what is collected or stored. C

Explicit user consent: Obtained where legally required (e.g., marketing), but no strong evidence of explicit, easy consent for all processing activities. C

Transparency

Clear policy: Accessible, detailed, and up-to-date (July 2025). A

Change notification: Users are informed of material changes with updated effective dates. A

Model documentation: Models are listed with references; source code was published on GitHub in September 2025 (NillionNetwork/nilgpt), along with initial technical documentation and a clear roadmap for further releases. A 

Privacy by Design

Encryption (core & advanced): Inputs/outputs are encrypted locally in the browser using a user-chosen passphrase (which never leaves the device), then secret-shared across nilDB nodes. Models run within TEEs, and the entire backend operates on nilCC (Nillion’s confidential compute layer). All data is encrypted in transit and at rest. A

Privacy-Enhancing Technologies: MPC for data at rest, TEEs for inference. Differential privacy/federated learning not needed in current model. A

Auditability & Certification: No third-party audits yet. Initial source code was published on GitHub in September 2025, but independent verification is still pending. C

Transparency & Technical Documentation: Architecture and privacy principles are publicly documented, with source code released in September 2025. Documentation covers core privacy and technical measures, with additional details planned. A

User-Configurable Privacy Features: Architecture and privacy principles are described publicly; in-app data deletion is now available (September 2025 release), but fine-grained privacy controls and complete documentation are still missing. C

Hosting & Sovereignty

Sovereignty: B

• The backend of NilGPT is hosted in nilCC, which uses Nillion’s bare metal servers in Virginia.
• nilDB nodes are used for data storage.
• There are 3 nodes used for NilGPT, one managed by Nillion and the other two by external parties.

Legal jurisdiction: Nillion Labs is a company based in Ireland, and the data collected by NilGPT is subject to Irish (EU) law and the General Data Protection Regulation (GDPR). A

Local option: It is not possible to host NilGPT locally, as it is only accessible via a web application. D

Big Tech dependence: NilGPT does not rely on public cloud infrastructure. In fact, nilCC, which hosts the backend and runs AI models, runs on bare metal, while nilDB nodes can be hosted anywhere without compromising user data security and confidentiality. A

Open Source

Publicly available model: The nilGPT application code was released in September 2025 on GitHub (NillionNetwork/nilgpt). The first version is fully accessible to the public, marking the beginning of its open-source journey. A

Clear open source license: The models operate under specific licenses (Llama, DeepSeek). The application’s license was partially disclosed with the September 2025 open-source release, but a complete open source license for all modules is still pending. B

Inference code available: Models are open source. Initial inference code was published on GitHub in September 2025, but it remains partial and lacks full documentation and audits. B 

Remarks

nilGPT Privacy Data Review : Overall Score

L’article nilGPT : Privacy and Data Review 2025 est apparu en premier sur Future of Internet.

The scoring system follows a comprehensive guide created specifically for this project, accessible here, and is designed to adapt dynamically as innovations emerge and feedback comes from the decentralized community.

Our mission is clear: to enlighten and inform, without filter or influence; so we can build together a fairer, more transparent AI ecosystem.

update : 25/08/11

Key Insights from the Perplexity Privacy Review 2025

Model

Perplexity Engine (core AI), Perplexity Pro (paid subscription, with Enterprise tiers that have stricter data retention controls), Comet Browser (standalone, privacy-focused browser with ad-driven data collection and its own privacy notice).

Data Collection

Prompts stored: Indefinite storage of prompts and interactions without immediate deletion guarantees undermines user control. Perplexity should implement a clear 90-day retention limit for non-essential data and offer instant deletion options via a user dashboard. C

Use for training:Using prompts for AI training by default with opt-out and no guaranteed anonymization risks user privacy. Perplexity must default to opt-in for training use and publish a detailed anonymization protocol to ensure data is untraceable. C

Account required: Requiring a Perplexity account for full features, including personal info like email and name, limits accessibility for privacy-conscious users. Allow full Comet Browser functionality without an account or minimize mandatory data to just an email address. B

Data retention duration: Consumer data is retained as long as “reasonably necessary” for business purposes, with no strict short-term cap. Deletion is possible upon request for most users, but not automatic. Enterprise files are deleted after 7 days by default. C

User Control

Deletion possible: Data deletion within 30 days is too slow and requires manual requests. Perplexity should offer a one-click deletion tool in the user interface with a 7-day processing guarantee. B

Export possible: Manual data export processes are cumbersome and user-unfriendly. Develop an automated export tool in the account settings to provide downloadable data in standard formats like CSV or JSON. C

Granularity control: Limited control over specific data types, like browsing history, restricts user autonomy. Introduce detailed settings to let users toggle data collection for individual features, such as search history or ad tracking. B

Explicit user consent: Relying on “legitimate interest” for non-sensitive data processing sidesteps user choice. Require explicit, granular consent for all data processing activities, aligning with GDPR-like standards. C

Transparency

Clear policy: The detailed privacy policy risks being inaccessible to non-technical users due to legal jargon. Simplify the policy with plain language summaries and multilingual translations to reach a broader audience. B

Change notification: Relying on users to check for minor policy updates is inconvenient and opaque. Implement proactive notifications via email and in-app alerts for all policy changes, with a 30-day advance notice for material updates. B

Model documentation: No public technical documentation on AI models erodes trust. Publish high-level model cards outlining training data sources and architecture without compromising proprietary details. C

Privacy by Design

Encryption (core & advanced): Claiming standard encryption without advanced privacy tech like homomorphic encryption or zero-knowledge proofs is insufficient. Adopt and document at least one advanced privacy-enhancing technology, such as secure multi-party computation, within 12 months. C

Privacy-Enhancing Technologies:  Perplexity relies on standard encryption without advanced privacy-enhancing technologies like zero-knowledge proofs, weakening its privacy framework. Adopt a PET, such as homomorphic encryption, within 12 months and publish a clear explanation of its implementation to enhance user trust. C

Auditability & Certification: Lack of third-party audits undermines credibility. Perplexity should commit to annual SOC 2 or ISO 27001 audits and publish summary reports to validate privacy practices. D

Transparency & Technical Documentation: Policy transparency exists, but technical details are absent. Share a public technical overview of data flows and encryption standards to demonstrate privacy commitment. C

User-Configurable Privacy Features: Moderate controls for cookies and tracking are a start, but gaps remain for other data types. Add comprehensive privacy settings, like disabling specific data collection categories (e.g., location, search queries), in a centralized dashboard. B

Hosting & Sovereignty

Sovereignty: Exclusive US hosting with cross-border transfers conflicts with global privacy laws like GDPR. Offer EU and APAC data residency options by 2026 to comply with regional regulations and enhance user trust. D

Legal jurisdiction: California-based jurisdiction and Data Privacy Framework oversight may concern international users. Clarify DPF protections and offer contractual safeguards for non-US users to address jurisdictional risks. D

Local option: No offline or local deployment options limit privacy-focused use cases. Develop an offline mode for Comet Browser or local processing for enterprise users within 18 months. D

Big Tech dependency: Reliance on OpenAI and Anthropic introduces third-party privacy risks. Publish a risk mitigation plan, including data isolation guarantees, and explore in-house model development to reduce dependency. C

Open Source

Publicly available model: The fully proprietary model blocks transparency and community trust. Release a limited, non-commercial version of the model or its weights to encourage research and align with open AI trends. D

Clear open source license: Perplexity’s fully proprietary platform lacks any open-source licensing, limiting transparency and developer engagement. To improve, Perplexity could provide a public API or sample code to support privacy-focused development while maintaining proprietary control. D

Inference code available: No public inference code stifles innovation and trust. Provide sample inference code or an open API sandbox for developers to test privacy-preserving features. D

Remarks

Perplexity Privacy Review 2025: Overall Score

L’article Perplexity : Privacy and Data Review 2025 est apparu en premier sur Future of Internet.

The scoring system follows a comprehensive guide created specifically for this project, accessible here, and is designed to adapt dynamically as innovations emerge and feedback comes from the decentralized community.

Our mission is clear: to enlighten and inform, without filter or influence, so we can build together a fairer, more transparent AI ecosystem.

update : 25/08/11

Key Insights from the Grok 4 Privacy Data Review 2025

Model

Grok 4 / Grok 4 Heavy

Data Collection

Prompts stored: User Content, including prompts, multimodal inputs (images/audio/video), and outputs, is stored with AES-256 encryption at rest and TLS 1.3 in transit; Private Chat ensures deletion within 30 days barring legal/safety exceptions, while standard chats persist for business needs. C

Use for training: User Content may train models unless opted out via settings, with public internet data and X posts (potentially containing personal info) used unrestricted; de-identification via hashed identifiers mitigates risks but doesn’t prevent incidental sensitive data inference. C

Account required: Mandatory accounts collect Account Data (name, credentials, DOB) and third-party login info (e.g., X profile, Google ID), enforcing age gates (13+ with parental consent for minors); this enables persistent tracking via IP-derived Technical Data. C

Data retention duration: Retention aligns with legitimate business purposes, potentially indefinite for legal/compliance, with no universal cap; Private Chat limits to 30 days max, emphasizing data minimization principles. C

User Control

Deletion possible: Users request conversation or account deletion via settings or privacy portal, processed in 30 days subject to legal holds; automated queuing ensures efficiency but exceptions for safety/security persist. B

Export possible: Data access requests under GDPR/CCPA yield personal info copies via Relyance portal, though not self-service or multi-format; appeals process adds user recourse. B

Granularity control: Binary toggles for training opt-out and Private Chat activation provide basic control over data use and visibility; lacks per-data-type or feature-level granularity. B

Explicit user consent: Consent mandated for precise location via GPS/third-parties, withdrawable anytime; core processing relies on terms acceptance with implied consent for training, advising against sensitive inputs. B

Transparency

Clear policy: Comprehensive policy details data categories, uses, and rights with examples, integrated FAQs on training sources; effective July 10, 2025, it promotes transparency via direct links. A

Change notification: Updates posted with new effective dates, proportional notices (e.g., email/in-Service) for material EU changes; global users lack proactive alerts beyond page revisions. B

Model documentation: High-level overviews cover reasoning, 256k context, multimodal support; no deep disclosures on architecture blueprints, parameter scales, or training pipelines. C

Privacy by Design

Encryption (core & advanced): Implements AES-256 at-rest encryption, TLS 1.3 transit protection, and hashed de-identification; absent advanced PETs like homomorphic encryption or differential privacy. B

Privacy-Enhancing Technologies: Data minimization via opt-outs and aggregation, anonymization for deidentified data; no production-scale federated learning or zero-knowledge proofs evident. C

Auditability & Certification: SOC 2 Type 2, GDPR, CCPA compliance claimed, with audit reports/certifications available on request; third-party verifications support but lack public disclosure. B

Transparency & Technical Documentation: Policy offers high-level security flows and trust statements; omits detailed architecture diagrams or data lineage maps. C

User-Configurable Privacy Features: Private Chat for ephemeral storage and training opt-out toggles enable basic customization; no advanced profiles or granular PET integrations. B

Hosting & Sovereignty

Sovereignty: Primarily U.S.-hosted on Colossus supercluster in Tennessee, with no regional sovereignty options; relies on domestic infrastructure for training/deployment. D

Legal jurisdiction: Governed by Texas law under Nevada-based X.AI LLC, weaker than EU standards despite GDPR claims for Europeans; disputes centralized in Tarrant County courts. D

Local option: Exclusively cloud/API-based, no self-hosted or on-premises variants; international transfers use SCCs for compliance. D

Big Tech dependency: Engages subprocessors like Oracle Cloud, NVIDIA GPUs, AWS for scaling; in-house Colossus mitigates but sustains reliance on hyperscalers. C

Open Source

Publicly available model: Grok 4 remains fully proprietary, unlike Grok-1’s Apache 2.0 release; weights and internals closed to public. D

Clear open source license: No OSS license applies; access restricted to API/subscriptions. D

Inference code available: Inference via hosted API only, no public code or downloadable models; supports text/vision but not open execution. D

Remarks

Grok 4 Privacy Review: Overall Score

L’article Grok 4 : Privacy and Data Review 2025 est apparu en premier sur Future of Internet.

The scoring system is based on a comprehensive guide created specifically for this project, accessible here. This ranking is dynamic, evolving with innovations and feedback from the decentralized community.

Our mission: to enlighten and inform, without filter or influence, to build together a fairer and more transparent AI ecosystem.

update : 25/08/11

Key Insights from the Qwen Privacy and Data Review

Model

Qwen Chat (by Alibaba Cloud, Singapore Private Limited)

Data Collection

Prompts stored: As stated in Alibaba Cloud’s Privacy Policy (Section 5.2): “We may retain and use your input queries and responses to improve our AI models.” No short-term deletion guarantee. C

Use for training: Non-personal User Content may be used to improve models, but without clear anonymization safeguards. C

Account required: No mandatory registration; no personal info required. However, prompts and technical identifiers (e.g., IP) are collected regardless. Anonymity is partial. B

Data retention duration: Retained as long as necessary or legally required. No maximum duration defined. Deletion/export rights exist only in certain jurisdictions (e.g., EEA, CCPA). D

User Control

Deletion possible: Users can request deletion, but process varies by jurisdiction and is not immediate. B

Export possible: Data portability rights apply in some regions, but export mechanisms and formats are unclear. C

Granularity control: No fine-grained settings for data collection or sharing. D

Explicit user consent: Consent is obtained via Terms and Privacy Policy, including deemed consent clauses. Withdrawal may limit service access. C

Transparency

Clear policy: Detailed Privacy Policy and Terms are provided, covering data use and rights, but written in formal legal language. A

Change notification: Updates published online or via email; users must actively monitor changes. B

Model documentation: No public model cards, training data provenance, or evaluation reports. Lags behind OpenAI or Anthropic in transparency. D

Privacy by Design

Encryption (core & advanced): Encryption in transit and at rest claimed, but no public verification or advanced measures (e.g., homomorphic encryption) C

Privacy-Enhancing Technologies: No documented use of differential privacy, federated learning, or zero-knowledge proofs. D

Auditability & Certification: No mention of third-party audits or certifications (e.g., SOC 2, ISO 27701). Encryption claims lack independent validation. D

Transparency & Technical Documentation: No public details on privacy architecture or PETs implementation. D

User-Configurable Privacy Features: No adjustable privacy settings beyond consent withdrawal. D

Hosting & Sovereignty

Sovereignty: Hosted in Singapore, Indonesia, and/or China. Despite local hosting, Alibaba Cloud is subject to Chinese jurisdiction, where national security laws may override privacy protections. C

Legal jurisdiction: Governed by Singapore law; moderate privacy standards, weaker than EU’s. B

Local option: No self-hosted or local deployment option available. D

Big Tech dependency: Runs on Alibaba Cloud infrastructure, no reliance on US-based Big Tech. A

Open Source

Publicly available model: The Qwen Chat service is fully proprietary. While some model weights (e.g., Qwen-7B) are released on Hugging Face under a non-commercial license, they are not equivalent to the production version. D

Clear open source license: No open-source license for the live service. Available weights are restricted and not production-ready. D

Inference code available: No inference code, training pipeline, or backend made public. Access only via hosted platform. D

Remarks

Qwen Chat grants Alibaba broad rights to store, use, and process user inputs for model improvement, with weak anonymization and indefinite retention. Users in certain regions have limited rights to deletion and data portability, but processes are not immediate or self-service. There is no transparency on model architecture, training data, or independent audits. While encryption is claimed, it lacks third-party verification, and no advanced PETs are used. Crucially, despite hosting in Singapore and Indonesia, the service remains under Chinese jurisdiction, a major sovereignty risk. The model is fully closed-source in production, with no local deployment option. Overall, Qwen Chat is a centralized, proprietary AI service operating under a moderately protective legal framework.

L’article Qwen 3 : Privacy and Data Review 2025 est apparu en premier sur Future of Internet.

The scoring system is based on a comprehensive guide created specifically for this project, accessible here. This ranking is dynamic, evolving with innovations and feedback from the decentralized community.

Our mission: to enlighten and inform, without filter or influence, to build together a fairer and more transparent AI ecosystem.

update : 25/08/09

Key Insights from the Gemini Privacy and Data Review

Model

Gemini 2.5 Flash / 2.5 Pro / AI Pro

Data Collection

Prompts stored: User prompts and interactions are stored by default if the user is 18 or older, with retention configurable (default 18 months, options for 3 or 36 months). If activity is disabled, data is kept up to 72 hours. Conversations reviewed by humans are retained up to 3 years. Data is dissociated from the account before human review. C

Use for training: Use for training: User prompts are used for model improvement and training, including with human reviewers, but anonymization measures are applied, and users can opt out of most data use by disabling activity (though comments may still be used). This offers above-average control compared to similar models. B

Account required: A Google account is required to use Gemini apps, and standard personal data is collected. No anonymous or accountless use described. C

Data retention duration: Data retention defaults to 18 months (user-configurable to 3 or 36 months). If activity is off, data is kept for 72 hours; reviewed conversations are kept up to 3 years. C

User Control

Deletion possible: Deletion possible: Users can delete past conversations from their account, offering significant control, though data reviewed by humans is retained up to 3 years and not deleted when clearing activity. This is above average but not fully comprehensive. B

Export possible: Export possible: Users can export their data upon request, likely via Google Takeout, but details on formats (e.g., JSON, CSV) and completeness are not specified in the privacy guide. B

Granularity control: Users can configure the retention period (3, 18, 36 months) and enable/disable specific activity collection, but not fine-grained controls for all data types. B

Explicit user consent: Consent is obtained for certain features (e.g., Voice Match), and users are informed about data use, but some consents are implicit or tied to general Google account terms. B

Transparency

Clear policy: A detailed and regularly updated privacy policy is available, with explicit explanations for most data processing. A

Change notification: The policy states that users will be clearly notified if data use for ads changes, but does not specify proactive, advance notifications for all changes. B

Model documentation: Limited information on model usage and legal bases is provided, with no public technical documentation on architecture, data, or security, which is below industry standards for transparency. D

Privacy by Design

Encryption (core & advanced): Standard encryption in transit and at rest is implemented, meeting industry expectations, though advanced encryption or public audits/certifications are not mentioned. B

Privacy-Enhancing Technologies: Some anonymization and dissociation measures are described, but there is no mention of advanced privacy tech like differential privacy (e.g., as used by Apple) or federated learning. C

Auditability & Certification: No mention of third-party audits or certifications in the provided text. D

Transparency & Technical Documentation: Privacy policy is detailed, but technical documentation is not described. C

User-Configurable Privacy Features: Some user controls for privacy (activity on/off, retention, deletion, audio opt-in), but not deeply configurable features (e.g., encryption levels, anonymization settings). C

Hosting & Sovereignty

Sovereignty: Data is hosted in Ireland (EU) for EEA and Switzerland, and in the USA elsewhere; no self-hosting or local deployment option mentioned. C

Legal jurisdiction: EU users are protected under EU law; other users are under US law. B

Local option: No local or self-hosted option; Gemini is only available as a cloud service. D

Big Tech dependency: Fully dependent on Google Cloud and infrastructure, with no alternatives. D

Open Source

Publicly available model: Gemini is fully proprietary; model and training data are not publicly available. D

Clear open source license: No open source license; the system is proprietary. D

Inference code available: No access to inference code; only API or web interfaces are offered. D

Remarks

L’article Gemini 2.5 : Privacy and Data Review 2025 est apparu en premier sur Future of Internet.

The scoring system is based on a comprehensive guide created specifically for this project, accessible here. This ranking is dynamic, evolving with innovations and feedback from the decentralized community.

Our mission: to enlighten and inform, without filter or influence, to build together a fairer and more transparent AI ecosystem.

update : 25/08/08

Key Insights from the Deepseek Privacy and Data Review

Model

Data Collection

Prompts stored: User inputs and chat history are retained to support service functionality and model enhancement. An opt-out mechanism (“Improve the model for everyone”) is available, but no automatic deletion policy exists; data is retained based on account or business requirements. C

Use for training: Access requires mandatory account creation (email, username, etc.), with no anonymous access. Opt-out for training is available, but enforcement and scope of data use are unclear. C

Account required: Account is required to use the service, and personal information like email, username, and date of birth may be collected, but no sensitive data is required by default. No anonymous access option mentioned. C

Data retention duration:Data is stored indefinitely unless the user initiates account deletion. No fixed retention limits are specified, raising concerns about prolonged data storage. D

User Control

Deletion possible:Users can delete their account and chat history via the platform interface. However, DeepSeek may retain certain data for legal compliance or to address violations, with no guaranteed immediate or unconditional deletion (no Service Level Agreement on deletion timelines). B

Export possible: Data export requires manual requests through support channels, lacking self-service functionality, which falls short of GDPR’s emphasis on accessible data portability. C

Granularity control: Users can opt out of model improvement and manage limited personal data (e.g., chat history). However, fine-grained control over all data types is not explicitly supported. B

Explicit user consent: For EU/UK users, GDPR-compliant consent is obtained for specific data processing, with clear information and options to withdraw consent, aligning with regulatory standards. A

Transparency

Clear policy: The privacy policy is comprehensive, accessible, and regularly updated, meeting high standards for clarity and user awareness. A

Change notification:Policy updates are communicated retroactively, lacking proactive user notification mechanisms. C

Model documentation: Limited references to model architecture are available but lack detailed public documentation, reducing transparency. C

Privacy by Design

Encryption (core & advanced): DeepSeek claims to implement “security measures” but provides no specifics on end-to-end or at-rest encryption protocols, raising concerns about data protection rigor. D

Privacy-Enhancing Technologies: Claims of data de-identification are noted, but no evidence of advanced PETs (e.g., differential privacy) or verifiable proofs is provided. D

Auditability & Certification: A GDPR representative (Prighter) is appointed, but no third-party audits or certifications are disclosed. D

Transparency & Technical Documentation: No detailed technical documentation on security measures or model architecture is available, limiting independent verification. D

User-Configurable Privacy Features: Only basic opt-out options are provided, with no advanced privacy customization features. C

Hosting & Sovereignty

Sovereignty: All user data is processed and stored in the People’s Republic of China, with no mention of self-hosting or alternative sovereign hosting. D

Legal jurisdiction: The legal jurisdiction is the People’s Republic of China, which does not provide strong data protection guarantees. D

Local option: No local or self-hosting options are available; the service is entirely cloud-based in China. D

Big Tech dependency: No explicit reliance on major US tech providers is noted. As a China-based entity with internal infrastructure, dependency is likely minimal. B

Open Source

Publicly available model: No indication that the model or training data is open source or publicly available. D

Clear open source license: No open source license provided or referenced. D

Inference code available: No mention of accessible inference code; API access only. D

Remarks

L’article DeepSeek V3 : Privacy and Data Review 2025 est apparu en premier sur Future of Internet.

The scoring system is based on a comprehensive guide created specifically for this project, accessible here. This ranking is dynamic, evolving with innovations and feedback from the decentralized community.

Our mission: to enlighten and inform, without filter or influence, to build together a fairer and more transparent AI ecosystem.

update : 25/08/07

Key Insights from the Llama Privacy and Data Review

Model

Data Collection

Prompts stored: Prompts and responses are not stored beyond the session, a strong privacy feature. However, other user data (e.g., identity, device info, usage patterns) may be retained without clear anonymization or deletion policies, posing privacy risks. Meta should guarantee no persistent prompt storage, anonymize all prompts by default, and define clear retention periods. C

Use for training: User data, even if anonymized, may be used to improve Llama 4, with no clear opt-out for standard users (self-hosting excepted). This lack of control and transparency over data use is concerning. A technical opt-out mechanism is needed. C

Account required: Personalized use requires an account with standard personal data collection. Anonymous use is limited to self-hosting, which is only feasible for technical users. Most users must share personal information. C

Data retention duration: User data is retained “as long as needed” for service or legal purposes, with no specific deletion timeline except for cookies (up to 400 days). This vagueness creates uncertainty. Meta should specify exact retention periods for all data types. C

User Control

Deletion possible: Account data can be deleted under GDPR, but the process may not be immediate or complete due to legal retention requirements. Full user control requires instant, total deletion by default. C

Export possible: GDPR-compliant data export is available, but the format and completeness are unclear. Standardized, usable formats (e.g., JSON, CSV) would enhance portability. B

Granularity control: Basic privacy settings are available, but fine-grained controls (e.g., per-prompt or per-interaction) are absent. Granular dashboards would empower users. C

Explicit user consent: Explicit consent for sensitive data and ads is GDPR-compliant and revocable, a strong transparency feature. A

Transparency

Clear policy: Meta’s privacy policy is detailed, readable, and regularly updated, clearly outlining data practices. A

Change notification: Users are notified in advance of major policy changes and can review them, a key transparency practice. A

Model documentation: Only high-level information on architecture, security, and data is provided, lacking detailed technical documentation. Full disclosure is needed for accountability. C

Privacy by Design

Encryption (core & advanced): Security is mentioned, but specifics on encryption methods (e.g., end-to-end) or certifications are absent, undermining trust. C

Auditability & Certification: No evidence of independent audits is provided. Regular, public audits would enhance transparency. D

Transparency & Technical Documentation: No advanced technical documentation on privacy or architecture, only high-level information. You can’t really see “under the hood” of Llama 4’s privacy protections. Full technical disclosure is needed for experts and users alike. C

User-Configurable Privacy Features: Basic privacy controls exist, but advanced customization (e.g., custom retention periods, prompt-level controls) is lacking. More flexible options would strengthen user control. C

Hosting & Sovereignty

Sovereignty: Self-hosting allows enterprises and advanced users to control data on their servers, a major privacy advantage. Non-technical users rely on Meta’s GDPR-compliant EU cloud, which is less flexible. A

Legal jurisdiction: Meta-hosted data falls under EU/Irish GDPR laws; self-hosted data follows local laws. Both offer strong legal protections. A

Local option: Self-hosting is available for enterprises and experts, but regular users are limited to Meta’s cloud. A local option for all users would improve flexibility. B

Big Tech dependency: Cloud users are tied to Meta’s infrastructure, while self-hosting offers independence but requires technical expertise. C

Open Source

Publicly available model: Llama 4’s weights are downloadable but come with usage restrictions, falling short of true open-source freedoms. No access to training data or full code is provided, limiting adaptability. B

Clear open source license: The custom “open weights” license restricts full open-source freedoms (e.g., sharing, modification). An OSI-approved license (e.g., MIT, Apache) would boost trust and innovation. C

Inference code available: Providing inference code enables self-hosting, a significant win for transparency and control, though primarily for technical users. A

Remarks

L’article Llama 4 : Privacy and Data Review 2025 est apparu en premier sur Future of Internet.

The scoring system is based on a comprehensive guide created specifically for this project, accessible here. This ranking is dynamic, evolving with innovations and feedback from the decentralized community.

Our mission: to enlighten and inform, without filter or influence, to build together a fairer and more transparent AI ecosystem.

update : 25/09/21

Key Insights from the Mistral Privacy and Data Review

Model

• Premier Models: Codestral 25.08, Voxtral Mini Transcribe, Mistral OCR 25.05, Ministral 3B, Ministral 8B, Codestral 25.01, Mistral Large 2, Pixtral Large, Mistral Small 2, Mistral Embed, Codestral Embed, Mistral Moderation.
• Open Models: Voxtral Small, Voxtral Mini, Mistral Small 3.2, Mistral Small 3.1, Mistral Small 3, Codestral Mamba, Pixtral 12B, Mistral Nemo 12B, Mathstral 7B, Mixtral 8x7B, Mistral 7B.

Data Collection

Prompts stored: Mistral AI provides clear and transparent practices for data storage, with Le Chat inputs retained until user deletion (or zero retention activation) and API data stored for 30 days unless zero retention is enabled. Fine-Tuning and Agents API data retention is also clearly defined. The availability of zero data retention and alignment with GDPR standards justify an A rating, as these practices demonstrate a strong commitment to user privacy and data minimization.

Use for training: Users can opt out of data use for training, and Le Chat Pro is excluded from training by default unless users opt in, which is a best-in-class approach. Feedback mechanisms (e.g., “thumbs up/down”) are also opt-out, ensuring user control. This robust, user-centric approach to training data usage, combined with clear opt-out options, warrants an A rating for exceeding standard expectations

Account required: An account is required for La Plateforme and certain endpoints, typically involving an email or phone number. Standard API key registration is available for broader access. C

Data retention duration: Mistral clearly outlines retention periods: Le Chat data is kept until deletion, API data for 30 days (unless zero retention is activated), Fine-Tuning/Agents API data until account termination, and civil identity data for 5 years post-termination. While the long retention of civil identity data slightly impacts the score, the transparency and GDPR compliance, along with zero retention options, support a B rating, reflecting a strong but not perfect policy.

User Control

Deletion possible: Mistral offers multiple clear channels for data deletion, including account settings, a “Privacy Requests” form, and direct contact with the Privacy Team at support@mistral.ai. These accessible and user-friendly mechanisms align with GDPR requirements and demonstrate a strong commitment to user control, justifying an A rating for robust deletion processes.

Export possible: There is no explicit mention of data export options for users. This would typically be detailed in the user agreement or privacy policy. D

Granularity control: Limited granularity control is available, primarily through opt-in/opt-out for training and zero data retention requests. Free users can object to data processing via email, the Help Center, or letter, though Pro users have easier in-interface controls (e.g., two-click opt-out). B

Explicit user consent: Consent is implied via acceptance of Terms of Service and Privacy Policy. Explicit consent is required for non-essential cookies and certain training opt-ins. Free users’ right to object was added post-GDPR complaint in February 2025, ensuring compliance. B

Transparency

Clear policy: Mistral AI provides a Privacy Policy and Terms of Service, accessible via their website and Help Center. Specifics are detailed in these documents, though not always in public-facing extracts. B

Change notification: No explicit proactive notification process for policy changes is mentioned; users are advised to check the Privacy Policy regularly for updates. D

Model documentation: Extensive technical documentation is available for models, endpoints, and deployment (e.g., La Plateforme, Mistral OCR), ensuring high transparency for technical offerings. A

Privacy by Design

Encryption (core & advanced): Mistral AI employs end-to-end encryption for data in transit and at rest, with AES-256 used at the disk level for secure storage, ensuring robust data protection. A

Auditability & Certification: Certified under ISO 27001/9001/14001/50001 and ANSSI II-901, with regular audits to maintain high security and compliance standards. A

Transparency & Technical Documentation: Extensive technical documentation exists for models and deployment, but privacy-specific measures (e.g., data processing details) are less comprehensive, rated as partial. B

User-Configurable Privacy Features: Privacy features are limited to global controls such as zero data retention and training opt-out. No fine-grained settings are available. C

Hosting & Sovereignty

Sovereignty: Mistral AI is hosted in EU tier-3+ data centers, providing options for sovereign and private hosting. This ensures compliance with EU data protection regulations and enhances data sovereignty. A

Legal jurisdiction: Mistral AI operates under EU legal jurisdiction, which is known for strong data protection laws. This provides an additional layer of legal protection for user data. A

Local option: Mistral AI models can be self-hosted, allowing for local deployment and greater control over data. This is ideal for organizations with stringent data privacy requirements. A

Big Tech dependence: While some endpoints are available via Big Tech clouds (Azure, AWS, GCP), Mistral AI offers sovereign and self-hosting options, reducing dependence on Big Tech infrastructure. B

Open Source

Publicly available model: Many of Mistral AI’s models are fully open source, including weights, code, and training data. This promotes transparency and community collaboration. A

Clear open source license: Mistral AI models are released under Apache 2.0 or Mistral Research License, providing clear guidelines for use and modification. A

Inference code available: Inference code for open models is accessible, well-maintained, and documented on platforms like GitHub. This facilitates ease of use and integration. A

Remarks

• Data Collection : 20 + 20 + 5 + 15 = 60
• User Control : 20 + 0 + 15 + 15 = 50
• Transparency : 15 + 0 + 20 = 35
• Privacy by Design : 20 + 15 + 20 + 15 + 5 = 75
• Hosting & Sovereignty : 20 + 20 + 20 + 15 = 75
• Open Source : 20 + 20 + 20 = 60

Total  : 60 + 50 + 35 + 75 + 75 + 60 = 355

23 x 20 = 460

365 / 460 x 100 = 77.2

L’article Mistral AI : Privacy and Data Review 2025 est apparu en premier sur Future of Internet.

The scoring system is based on a comprehensive guide created specifically for this project, accessible here. This ranking is dynamic, evolving with innovations and feedback from the decentralized community.

Our mission: to enlighten and inform, without filter or influence, to build together a fairer and more transparent AI ecosystem.

update : 25/08/05

Key Insights from the Claude Privacy and Data Review

Model

Claude Sonnet 4 : Hybrid reasoning model, optimized for a balance of performance and efficiency, with advanced coding, data analysis, tool use, and agentic task capabilities. Available free on Claude.ai and via paid plans on Anthropic API, Amazon Bedrock, and Google Cloud Vertex AI.

Data Collection

Prompts stored: User data is only used for training with explicit opt-in consent, in line with Anthropic’s practices. By default, Anthropic employees cannot access conversations unless explicit consent is provided or Usage Policy violations occur. A

Use for training: Training uses a mix of public web data (up to March 2025), non-public third-party data, service provider data, data from consenting users, and internally generated data. Cleaning and filtering methods, including deduplication, are applied. A

Account required: Yes, an account is required to access the model via Claude.ai, API, or cloud platforms. C

Data retention duration: Specific retention periods documented C

• Deleted conversations: Immediate removal from history, automatic backend deletion within 30 days
• Usage Policy violations: Up to 2 years for inputs/outputs, up to 7 years for safety classification scores
• Feedback data: 10 years for consented submissions

User Control

Deletion possible: Yes, users can delete conversations with immediate removal from conversation history and automatic backend deletion within 30 days. A

Export possible: Yes, data exports are available to free/Pro Claude.ai users and Team plan Primary Owners. Users can export conversation data and account information from Settings > Privacy on web app or Claude Desktop (not available on mobile apps). A

Granularity control: Claude Enterprise offers custom retention controls (minimum 30 days). Standard users have conversation deletion control. B

Explicit user consent: Explicit consent required for use of data for training, consistent with Anthropic’s stated policy. A

Transparency

Clear policy: Data use and deployment policies are clear, though some aspects (e.g., data retention) lack precision. B

Change notification: Updates are well-documented via the Responsible Scaling Policy and System Cards, including details on security evaluations and deployment decisions. A

Model documentation: Full documentation via detailed report and System Card, including performance evaluations (e.g., 72.7% on SWE-bench Verified), security measures (ASL-2), and deployment decisions. A

Privacy by Design

Encryption (core & advanced) | B | Data encrypted in transit using TLS and at rest using 256-bit AES encryption. Implementation includes:

• Transport Security: TLS 1.3 protocol with Perfect Forward Secrecy (PFS)
• At-Rest Encryption: AES-256 encryption for stored data with encrypted key management
• Access Controls: Role-based access control (RBAC) with principle of least privilege
• Employee Access: Zero default access policy – conversations accessible only with explicit user consent or Usage Policy violations requiring Trust & Safety review
• Research Initiative: Active development of Confidential Inference via Trusted Virtual Machines for enhanced model weight and user data protection

Auditability & Certification | A | Comprehensive compliance framework including:

• SOC 2 Compliance: SOC 2 Type I and Type II certifications achieved
• HIPAA Certification: HIPAA compliance for healthcare data handling
• ISO Standards: ISO/IEC 42001:2023 certification for AI management system – first international standard for AI governance
• External Audits: Regular independent security assessments and penetration testing
• Government Evaluation: Deep evaluations by US AI Safety Institute, UK AI Security Institute for safety and compliance
• Trust Center: Publicly accessible compliance artifacts and documentation via Anthropic Trust Center

Transparency & Technical Documentation | A | Comprehensive technical disclosure including:

• System Cards: Detailed performance evaluations, security measures (ASL-2 classification), and deployment decisions
• Safety Metrics: Quantified harmlessness rate (98.99% ±0.23%), jailbreak resistance metrics (6.71% standard mode, 2.24% extended thinking)
• Technical Capabilities: Documented performance benchmarks (72.7% on SWE-bench Verified), context window specifications (200k tokens), output limits (64k tokens)
• Security Testing: Prompt injection defense statistics (86% prevention with protections), cyber capability assessments
• Responsible Scaling Policy: Public documentation of safety thresholds and deployment criteria

User-Configurable Privacy Features | B | Tiered privacy control system:

• Enterprise Level: Custom retention controls (minimum 30 days), enhanced privacy settings, data residency options via cloud providers
• API Clients: Zero data retention agreements available for qualified enterprise API clients (subject to Anthropic approval)
• Standard Users: Conversation deletion controls, data export capabilities (Settings > Privacy), opt-out mechanisms for training data
• Claude Code: Strict read-only permissions by default with explicit permission requests for file modifications
• Telemetry Control: Environment variable options for disabling analytics collection (DISABLE_TELEMETRY)

Hosting & Sovereignty

Sovereignty | C | Limited options available:

• Standard deployment: AWS (US East, US West, APAC, Europe) and Google Cloud
• Government deployment: AWS GovCloud for US Intelligence Community and national security customers
• Claude Gov models: Specialized deployment for classified US national security environments
• No general on-premises or local deployment options for commercial customers

Legal jurisdiction | D | Primarily US-based operations with international data transfer protections:

• EU transfers protected by Adequacy Decisions and Standard Contractual Clauses (SCCs)
• Specialized US jurisdiction deployment for government customers

Local option | C | Limited local deployment options:

• Claude Code can integrate with local MCP servers for desktop applications
• Enterprise deployments through Amazon Bedrock or Google Vertex AI
• No true on-premises deployment for commercial customers
• Government-specific deployments in AWS GovCloud

Big Tech dependency: Confirmed dependency on AWS (regions: US East, US West, APAC, Europe) and Google Cloud Vertex AI for hosting. D

Open Source

Publicly available model: Proprietary model, not open source. D

Clear open source license: Not applicable. D

Inference code available: Not available. However, Model Context Protocol (MCP) is open protocol for integrations. D

Remarks

Claude Sonnet 4 is a versatile model, surpassing Claude Sonnet 3.7 in multiple domains, notably coding (72.7% on SWE-bench Verified) and data analysis, with a 200k token context window and 64k output max. It is classified under the ASL-2 safety standard, with a harmlessness rate of 98.99% (±0.23%) and improved jailbreak resistance (6.71% in standard mode, 2.24% in extended thinking).

The model has specialized government deployments through Claude Gov models, which are designed for US national security customers and operate in classified environments. Rigorous evaluations have been conducted to minimize risk, including harmlessness training, post-deployment monitoring, and defenses against prompt injection (86% prevention with protections).

Cyber capabilities remain limited (e.g., 1/4 on network challenges), and the model is below ASL-3/ASL-4 thresholds for autonomy and CBRN risk. Available free on Claude.ai, it is well-suited for daily tasks and intermediate coding projects.

Privacy and Data Review: Overall Score

59.78/100

• Data Collection : 20 + 20 + 5 + 5 = 50
• User Control : 20 + 20 + 15 + 20 = 75
• Transparency : 15 + 20 + 20 = 55
• Privacy by Design : 15 + 15+ 20 + 20 + 15 = 85
• Hosting & Sovereignty : 5 + 0 + 5 + 0 = 10
• Open Source : 0 + 0 + 0 = 0

Total : 50 + 75 + 55 + 85 + 10 + 0 = 275

23 x 20 = 460

275 / 460 x 100 = 59.78

This evaluation is provided for informational purposes only and reflects a subjective analysis based on publicly available data at the time of publication. We do not guarantee absolute accuracy and disclaim all liability for errors or misinterpretations. Any disputes must be submitted in writing to futurofintenet@proton.me

For full methodology, see our complete scoring guide here: LLM Privacy Rating Guide

L’article Claude Sonnet 4 : Privacy and Data Review 2025 est apparu en premier sur Future of Internet.

The scoring system is based on a comprehensive guide created specifically for this project, accessible here. This ranking is dynamic, evolving with innovations and feedback from the decentralized community.

Our mission: to enlighten and inform, without filter or influence, to build together a fairer and more transparent AI ecosystem.

Key Insights from the ChatGPT Privacy and Data Review

Model

• GPT-4

Data Collection

• Prompts stored: User prompts and data are stored for service operation, security, and model improvement; temporary chat mode limits storage to 30 days, but retention is otherwise indefinite based on service or legal needs.C
• Use for training: User content is used for training by default, with an opt-out available in settings; no clear anonymization process for training data, and opt-out is not enabled by default. C
• Account required: Usage requires a registered account, collecting standard personal data (email, name, etc.), with no anonymous access option. C
• Data retention duration:
  Data retention duration: Data is retained as long as necessary for service or legal obligations; temporary chats are deleted within 30 days, but other data retention is potentially indefinite.C

User Control

• Deletion possible: Users can delete personal data via self-service in account settings or support requests, but delays (up to 30 days) may occur, and complete deletion of all data (e.g., training sets) is not guaranteed. B
• Export possible: Data export is available via account settings or support request, but limited to conversation history in formats like JSON, with potential exclusions of metadata. B
• Granularity control: Granularity control:</strong> Limited to opt-out for training data; no advanced controls for metadata or other data types for non-enterprise users. C
• Explicit user consent:
  Explicit user consent: Consent is collected at registration, with GDPR rights supported; training data usage is opt-out by default, and settings management can be non-intuitive. B

Transparency

• Clear policy: Privacy policy is comprehensive, accessible, and regularly updated, detailing data processing for general users. A
• Change notification: Significant policy or terms changes are notified via email or in-product alerts at least 30 days in advance. A
• Model documentation: Model documentation:Limited public documentation on model architecture, training data, or security processes; system cards provide only high-level details. C

Privacy by Design

• Encryption (core & advanced): AES-256 encryption for data at rest and TLS 1.3 for data in transit; no public details on key management or advanced features like HSM. B
• Privacy-Enhancing Technologies : ChatGPT-4 employs robust Privacy-Enhancing Technologies like AES-256-GCM encryption, TLS 1.2/1.3, and opt-out data training controls, but past data leaks (e.g., March 2023) and potential jailbreaking risks limit its rating to B. Users should avoid sharing sensitive data and use professional versions for enhanced privacy. B
• Auditability & Certification: Auditability & Certification:</strong> SOC 2 Type 2 certification for enterprise versions, with regular third-party audits; not explicitly confirmed for public version; no ISO 27001 certification. B
• Transparency & Technical Documentation: Partial public documentation on privacy and security measures via Trust Portal; lacks detailed technical specifics. B
• User-Configurable Privacy Features: Basic opt-out for training available; enterprise versions offer advanced controls, but public users have limited options. C

Hosting & Sovereignty

• Sovereignty: Data hosted primarily in the USA, outside EU/EEA/CH; GDPR compliance via Standard Contractual Clauses (SCCs), but no EU-US Data Privacy Framework certification. C
• Legal jurisdiction: Ireland for EEA users, USA for others; GDPR compliance ensured through SCCs for data transfers. B
• Local option: No local or self-hosting options available for any users. D
• Big Tech dependency: Hosted on major cloud providers (likely AWS or Azure), with no explicit alternatives disclosed. C

Open Source

• Publicly available model: Fully proprietary model, with no public access to weights or architecture. D
• Clear open source license: No open source license; usage governed by proprietary terms of service. D
• Inference code available: Inference accessible only via OpenAI API; no public inference code provided. C

Remarks

Privacy and Data Review: Overall Score

47.7/100

• Data Collection : 5 + 5 + 5 + 5 = 20
• User Control : 15 + 15 + 5 + 15 = 50
• Transparency : 20 + 20 + 5 = 45
• Privacy by Design : 15 + 15 + 15 + 15 + 5 = 65
• Hosting & Sovereignty : 5 + 15 + 0 + 5 = 25
• Open Source : 0 + 0 + 5 = 5

Total : 20 + 50 + 45 + 50 + 25 + 5 = 210

23 x 20 = 460

210 / 440 × 100 = 47.7

This evaluation is provided for informational purposes only and reflects a subjective analysis based on publicly available data at the time of publication. We do not guarantee absolute accuracy and disclaim all liability for errors or misinterpretations. Any disputes must be submitted in writing to futurofintenet@proton.me

For full methodology, see our complete scoring guide here: LLM Privacy Rating Guide

L’article ChatGPT 4 : Privacy and Data Review 2025 est apparu en premier sur Future of Internet.