AI assistants privacy and security comparisons

AI assistants now sit quietly inside our inboxes, browsers, and code editors – answering questions, rewriting drafts, and processing documents with impressive speed. But behind their helpfulness is a simple reality: every interaction feeds them data.

What happens to that data, how it's stored, used, or shared, depends entirely on the provider. Some assistants keep your prompts indefinitely. Others use them to train future models. A few may pass them to third-party systems for processing.

So we turn to the companies behind: OpenAI, Google, Anthropic, and Microsoft, and ask the questions that rarely make it into marketing decks:

• When users hand over their data, what exactly are you collecting?
• Where does that data go? How long do you keep it?
• Is any of it used to train your models now, or in the future?
• And are users ever given a clear explanation of what’s happening?

How we compared AI assistant privacy and security

To understand how major AI assistants handle user data, I looked beyond claims and focused on what’s documented – both in official policies and independent findings.

The evaluation focused on 5 key criteria:

• Data collection. What information do these systems capture – intentionally or by design? Does it include prompts, uploaded files, interaction logs, or background telemetry?
• Storage and retention. Once collected, where does the data live? For how long? Under what encryption standards, and with what access controls?
• Data sharing and use. Is user data used to train future models? Is it ever reviewed by humans, shared with contractors, or routed through third-party processors?
• Compliance and transparency. Do these tools meet regulatory benchmarks like GDPR and CCPA? Just as important: how clearly do they explain any of this to the user?
• Consumer vs enterprise practices. I made a distinction between freely available consumer tools and their enterprise-grade counterparts, where the data handling policies often diverge, even within the same company.

Data collection and storage policies

We all know, in theory, that AI assistants are learning something about us. But what exactly are they collecting? And where does that data go after the chat ends?

This section looks at what each major provider, OpenAI, Google, Anthropic, and Microsoft, actually stores: prompts, files, metadata, and usage habits. I examine how long that data sticks around, where it’s kept, and what control (if any) users really have over it.

ChatGPT (OpenAI)

OpenAI’s data policies have evolved under pressure: from the public, regulators, and lawsuits, but the core model still leans heavily on data collection. If you're using ChatGPT in its standard form, your content is stored by default, your metadata is logged, and your deletion rights are opt-in, not automatic.

Here’s a breakdown of how it works:

Policy area	Details (2025)
Data collected	Prompts, responses, uploaded files (images, docs), IP address, browser, device info, usage analytics
Retention (standard users)	Indefinite, unless manually deleted
Temporary chat mode	Optional; data auto-deletes after 30 days
Enterprise/education tiers	Customizable retention; default data stored in preferred regions (US, EU, Japan)
Data residency	Supported at rest; some transient processing may occur outside selected regions
Special features	Operator AI agents may store data (e.g. screenshots) for up to 90 days
Encryption	TLS 1.2+ in transit, AES-256 at rest
Deletion/export options	Available in account settings; full deletion may take up to 90 days
Training use	Opt-out default for standard users; enterprise data excluded from training
Limitations	Third-party GPTs and Apple Intelligence are not covered under data residency rules

ChatGPT’s privacy posture now includes more user controls, but the defaults still favor collection and long-term retention. If you’re not proactive, your data will likely stick around longer than you think.

Gemini (Google)

Google's Gemini sits at the intersection of enterprise infrastructure and consumer-facing AI – a position that demands a balancing act between control and data appetite. On paper, the privacy framework is great: strict access rules, region-aware storage, and enterprise-grade encryption.

But peel back the layers, and the policy reveals a familiar tension, especially in the standalone app and API tiers, where training use and long retention windows quietly persist.

Here’s a breakdown of how Gemini handles data in 2025:

Policy area	Details (2025)
Data collected	Prompts, responses, uploaded content, device metadata, system logs, usage analytics
Workspace interactions	Prompts/responses not retained after session ends; content handled per Cloud DPA
Standalone app retention	Up to 36 months by default; can be configured by org admins
Conversation history off	Data retained for up to 72 hours for service performance
Enterprise control	Admins manage retention via Google Vault; Workspace DLP and client-side encryption supported
Training use	Free-tier user data may be used to train models; enterprise/org data excluded unless consented
Data access	Strict permissions model; no cross-user data access
Human review	Some prompts/files reviewed to improve AI; disassociated but not anonymized
Review retention	Up to 3 years, even if original data is deleted
Export/deletion options	Managed through Workspace and Google Account tools
Storage location	Distributed globally; EU data boundaries respected for enterprise
Security	Enterprise-grade protections; client-side encryption available

Despite a strong security posture, Gemini’s consumer-level data use, especially the training defaults and prolonged human-reviewed storage, raises questions about transparency and GDPR nuance. For enterprise users, the policies are safer, clearer, and tightly governed. But for everyone else, the system quietly remembers more than it lets on.

Claude (Anthropic)

Claude was once the poster child for restraint – deleting consumer data within 30 days, promising no training without consent. That chapter has now closed. As of late 2025, Anthropic has introduced a default opt-in model for training data, reshaping Claude’s privacy posture from conservative to quietly expansive.

The shift hinges on consent, but with a twist: if you don't explicitly opt out by September 28, 2025, your silence is taken as permission. From that point, your chats, your code, and your conversations may be used to train Anthropic’s future models, and they may be stored for up to five years.

Here’s how Claude handles data now:

Policy area	Details (2025)
Data collected	Prompts, code, conversations, usage metadata; flagged content tracked separately
Consent model	Opt-out required by Sept 28, 2025; non-response = consent
Training use	Default opt-in for Claude Free/Pro/Max/Code; enterprise, gov, and API users excluded
Retention (ppt-in users)	Up to 5 years for training data
Retention (deleted/flagged)	Deleted chats excluded from training; flagged content stored up to 7 years
Privacy controls	Settings available to opt in/out; applies only to future data, not retroactively
Enterprise protection	Business, education, government, and API users’ data not used for training
Third-party sharing	No personal data shared for training; filtering tools applied to reduce sensitive content
Export/deletion options	Limited; retroactive deletion of training data not supported
Security	Standard encryption and internal safeguards; data not shared externally

This marks a philosophical pivot: from default privacy to default contribution. Claude now asks users to choose, but quietly assumes consent if they don’t. In a climate of rising AI ambition, that’s less a bug and more a business model.

Copilot (GitHub / Microsoft)

Copilot operates under a different set of assumptions. While other assistants mine prompts and documents for signals, Copilot is built to get in, suggest, and get out, especially when it comes to your code. It doesn’t store the code it sees, doesn’t reuse it for training, and offers settings that can push telemetry collection close to zero. The message is clear: we’ll help you write, but we don’t need to remember it.

Here’s a breakdown of how GitHub Copilot handles your data in 2025:

Policy area	Details (2025)
Data collected	Code snippets (processed in real-time), usage analytics, prompt acceptance/rejection, engagement data
Code retention	Code is not stored post-suggestion; discarded immediately after processing
Telemetry retention	Prompts/suggestions: up to 28 days or not retained (depending on mode); engagement data: up to 2 years (Business/Enterprise)
Customization options	Users/orgs can disable most telemetry for near-zero data retention
Training use	Code snippets not used to train models
Security	End-to-end encryption, fine-grained access control, local real-time processing
Feedback data	Retained as long as necessary for service improvement, abuse detection, compliance
Compliance	GDPR-compliant; Data Protection Agreements available for enterprise customers
Developer controls	Code scanning/filtering tools to reduce exposure of sensitive data in prompts
Storage location	Managed via GitHub’s infrastructure; subject to DPA agreements for enterprise

In a landscape filled with aggressive data retention and default opt-ins, Copilot’s posture is almost ascetic. It’s here to help, not to harvest. And for developers who treat their code like intellectual property, as they should, that separation matters.

Data sharing and third-party access

While AI providers often state they don’t sell user data, that doesn’t mean it stays entirely private. Internal systems, integration partners, and legal obligations can all create pathways for third-party access.

OpenAI (ChatGPT) allows authorized vendors access to user data under strict controls and doesn’t share data for advertising. But unless you opt out, your chats and uploads may be used to train its models. Enterprise data is excluded, but for everyone else, retention is long, and integrations carry their own risks. OpenAI complies with government demands and may retain deleted content under legal orders.

Google Gemini shares data within the Google ecosystem to support integration with Workspace and Cloud tools. Advertising use is off the table, but personalization and product improvements are very much on it. Enterprise data remains within controlled environments, and government requests are disclosed via standard transparency reports.

Claude (Anthropic) now uses a default opt-in model for training: no response by the deadline means consent. Data isn’t shared for advertising, and enterprise and API users are shielded from training use. While Anthropic follows legal obligations, it lacks the scale of public transparency offered by larger firms.

GitHub Copilot stands apart. It doesn’t retain or reuse code snippets, and telemetry, though collected, is tightly scoped and configurable. Code is not used for training. Microsoft handles legal requests through detailed public disclosures, offering the clearest window into government access.

In short, Claude and OpenAI lean toward model training, Google toward integration, and Copilot toward discretion. The safest path still begins with reading the settings you skipped past.

Security measures

In 2025, encryption is the easy part. Every major provider encrypts data both in transit (using protocols like TLS 1.2+) and at rest (with AES-256 or similar). That’s the baseline.

Where they differ is in how they manage identity, enforce access, and track accountability. This includes the strength of their authentication systems (SSO, MFA), the depth of their enterprise control layers (admin tools, audit logs, policy enforcement), and the granularity of access management (who can do what, and when).

Authentication & access control

Provider	Consumer login	Enterprise login
OpenAI (ChatGPT)	Email + password, optional MFA	SAML SSO, enforced MFA, org-wide policies, role-based access
Google Gemini	Google account (OAuth, email/password)	Workspace SSO, enforced MFA, IAM controls, audit trails
Claude (Anthropic)	Email + password, optional MFA	SSO with major IAMs, configurable auth, contractual restrictions
Copilot (GitHub)	GitHub account (email/password, MFA)	Azure AD SSO, centralized identity, audit logging

OpenAI’s ChatGPT supports SAML SSO, a widely used standard that lets companies connect ChatGPT to internal identity providers like Okta, Azure Active Directory, or Ping Identity. This means employees log in using their company credentials, and IT can enforce global rules: MFA, password policies, and session timeouts.

Every action inside Gemini ties back to your Google Identity, allowing administrators to control OAuth app permissions, force MFA, monitor activity logs, and define detailed access rules, all without leaving the Google Admin dashboard.

Claude, via its Claude Enterprise and Bedrock offerings, supports SSO integration through leading IAM platforms. That includes tools like Okta, AWS IAM, or any provider that supports SAML or OIDC standards. This allows companies to tailor login behavior, which enables biometric MFA, region-specific access, or temporary login tokens.

Copilot, tightly bound to Microsoft’s ecosystem, uses Azure Active Directory (AAD) to manage enterprise access. With AAD SSO, developers sign in using their existing Microsoft credentials – those used across Outlook, Teams, and Azure. Admins can enforce conditional access (e.g., only allow logins from company devices), monitor session logs, and revoke permissions at the user or repo level in real time.

Data encryption standards

When AI companies say your data is encrypted in transit and at rest, they’re using two standard tools.

TLS (Transport Layer Security) keeps your data safe while it’s moving, like when you send a message to the AI. AES-256 protects it once it’s stored on their servers.

Platform	Encryption in transit	Encryption at rest	End-to-end encryption	Enterprise vs consumer differences
OpenAI	TLS 1.2+	AES-256	❌ No	Enterprise adds data residency, retention, and access controls
Google Gemini	TLS 1.2/1.3	AES-256	❌ No	Enterprise gets client-side encryption, data loss prevention, residency
Claude	TLS (latest)	AES-256	❌ No	Enterprise offers zero-retention, hardware encryption options
GitHub Copilot	TLS 1.2/1.3	AES-256	❌ No	Enterprise supports compliance-focused encryption, controls

All four providers rely on robust encryption protocols, but none offer true end-to-end encryption for AI interaction data due to the necessity of processing data on their servers for AI functionality. Enterprise plans across the board provide additional encryption controls, data residency, and retention policies to meet compliance needs more effectively than consumer plans.

Model training with user data

Platform	Uses user data for training	Opt-out available	Enterprise default policy
OpenAI	✅ Yes	✅ Yes, user opt-out	No training on enterprise data
Google Gemini	✅ Yes	Limited, mainly admin-controlled	No training on enterprise content
Claude (Anthropic)	Only with explicit opt-in	✅ Yes, opt-in required	No training by default
GitHub Copilot	❌ No (code excluded)	Not generally required	No training on enterprise code

In brief, all four platforms use user data for model training to varying extents but provide opt-out or opt-in mechanisms for consumers. For enterprise customers, the default is generally set to exclude customer data from AI training to maintain confidentiality and compliance.

User control and transparency

Platform	Data deletion and export tools	Training opt-out	Transparency reports
OpenAI	User deletion, export, dashboards	User opt-out available	Regular transparency reports
Google Gemini	Data deletion, export via Workspace	Admin-controlled opt-out	Part of Google transparency reports
Claude	Deletion, export (enterprise focus)	Explicit opt-in only	General transparency policies
GitHub Copilot	Limited deletion, via GitHub tools	Code excluded, no opt-out	Microsoft transparency reports

OpenAI (ChatGPT) was the most user-friendly. You can delete chats (with a 30-day delay), export your data, and opt out of training right from the settings. Enterprise users are excluded from training by default. Their transparency reports are clear and regularly updated.

Key takeaway for ChatGPT: while a deletion feature exists, its effectiveness for non-enterprise users is now in question due to legal obligations imposed on OpenAI.

With Google Gemini, most control sits with Workspace admins. Deletion and export happen through tools like Google Takeout, and training opt-out is mostly an organizational setting. Consumers get less say, but Google's infrastructure is strong and well-documented.

Claude (Anthropic) has moved to a default opt-in model; if you don’t opt out, your data can be used for training and kept for up to five years. Deleted chats aren’t used, but older ones might still live on. Enterprise tools are stronger; consumer-facing controls are still developing.

GitHub Copilot avoids most of this by not using code for training at all. But telemetry data is collected, and users have limited ways to delete it. Most controls live in GitHub or Microsoft account settings. Microsoft’s legal transparency, though, is solid.

FAQ