We scored 12 AI tools
on privacy. Most failed.
We spent months reading privacy policies, regulatory filings, academic research, and enforcement actions so you don't have to. Eight of the twelve tools share your data with third parties. Three train on your conversations by default. Two earned an outright F with no credible path to opt out.
How to cite: Ahmed, M. (2026). Privacy in the Age of Frontier AI. Wzdom Research, Q2 2026. wzdom.ai/ai-privacy — full paper: wzdom.ai/research/wzdom-ai-privacy-report-q2-2026.pdf
- 1. What AI tools collect
- 2. How the data pipeline works
- 3. Global AI privacy laws
- 4. US privacy landscape
- 5. EU AI Act & GDPR
- 6. Data collection at a glance
- 7. Full AI privacy scorecard
- 8. About Wzdom Research →
- 9. AI privacy intelligence feed
- 10. For researchers & influencers
- 11. How to protect yourself
- 12. Wzdom AI Shield
- 13. Stay updated
See this happening on your device — Wzdom Privacy monitors every network connection in real time.
See what your apps are sending →Wzdom Research — 12 tools analyzed, Q2 2026
Source: Wzdom AI Privacy Index (WAPI), Q2 2026. Based on published privacy policies, terms of service, and regulatory filings. Full methodology in research paper ↗
Right now, on your device
Your apps are making connections you cannot see. Wzdom Privacy shows every one of them — in real time.
100% on-device · No account required · Free during beta
AI Privacy Index — 12 tools, every dimension.
Each tool scored across 6 independently weighted dimensions: training consent, data retention, transparency, regulatory record, data use, and user control. Sorted worst to best by WAPI grade. Scroll within the table to see all rows without losing the column headers. Full methodology ↗
| AI Tool | Grade | Risk Level | Stores Conversations? | Trains on Your Data? | Training Opt-Out? | Human Review? | 3rd-Party Sharing? | Local / Offline Mode? |
|---|---|---|---|---|---|---|---|---|
| GrokxAI | F | HIGH | Yes | Yes (+ X posts) | Buried in X settings | Yes | Yes (X data) | No |
| Meta AIMeta | F | HIGH | Yes | Yes — no opt-out | None | Yes | Yes (ad targeting) | No |
| DeepSeekDeepSeek AI | F | HIGH | Yes (China servers) | Yes | None | Unknown | Unknown | Self-host only |
| GeminiGoogle DeepMind | D | HIGH | Yes | Yes (Workspace default) | Limited | Yes | Yes (Google Ads) | No |
| ChatGPTOpenAI | C | MODERATE | Yes | Yes (default) | Opt-out available | Yes | Yes (sub-processors) | No |
| CopilotMicrosoft | C | MODERATE | Yes | Consumer tier | Partial | Yes | Yes (MS partners) | No |
| GitHub CopilotGitHub / Microsoft | C | MODERATE | Yes (code snippets) | Yes — your code | Enterprise only | Yes | Yes (MS ecosystem) | No |
| PerplexityPerplexity AI | C | LOW | Yes (90 days) | No | N/A | No | Yes (search APIs) | No |
| ClaudeAnthropic | B | MODERATE | Yes | Consumer only | Opt-out | Limited | No (API) | No |
| BedrockAmazon / AWS | B | LOW | No (by default) | No | N/A | Unknown | No | No (API only) |
| Mistral (Le Chat)Mistral AI | B+ | LOW | Yes | No (API) | N/A | No | No | Open weights |
| Ollama / Local LLMsOpen source | A | LOW | No | No | N/A | No | No | Yes — 100% |
Source: Wzdom AI Privacy Index (WAPI), Q2 2026. Based on published privacy policies, terms of service, and regulatory filings. Green = privacy-respecting. Red = concern. Amber = partial or conditional. Not legal advice.
Why this matters
The consequences are not theoretical.
AI tools train on your conversations by default
Most major AI platforms use your chat history to improve their models. Your private queries, professional questions, and personal context become permanent training data — unless you opt out, if an opt-out exists.
Human employees can read your prompts
Quality review teams at OpenAI, Google, and Meta have access to flagged conversations. What you consider a private conversation may be read, annotated, and stored by a human reviewer you never consented to.
Your data is shared with third parties
Eight of the twelve AI tools we analyzed share behavioral data with sub-processors, advertising partners, or parent companies. Most disclose this in privacy policies written to obscure rather than inform.
AI decisions increasingly affect real outcomes
AI systems now influence hiring decisions, credit scoring, healthcare triage, and content moderation. The data these systems train on shapes the decisions they make — and the privacy practices of today determine the fairness of those decisions tomorrow.
What they collect
What AI tools actually collect about you
Most people think they're sharing a question. In reality, they're sharing far more — their identity, their device, their behavior patterns, and often, fragments of work they'd never share with a stranger. Every prompt you type is logged before any response is generated. That's not paranoia — it's how these systems are designed to improve. Here's what ends up in their databases.
Prompt history
Every message you type — including accidental pastes of passwords, medical records, or private contracts. Retained indefinitely unless you manually delete.
Uploaded files & images
Documents, spreadsheets, photos, and code files you share. These are processed and stored server-side, often used for model evaluation.
Browsing context
AI browser extensions (Copilot, Gemini in Chrome, ChatGPT companion) capture the full text of every page you visit while active.
Account & identity
Name, email, phone number, payment method, billing address, IP address, and device fingerprint — all tied to your conversation history.
Usage patterns
What you ask, how often, at what hours, which features you use, what you delete — behavioral data worth as much as the conversations themselves.
Third-party integrations
Data from connected apps — Google Drive, GitHub, Outlook, Slack, Salesforce — flows through AI tools when you enable integrations.
What really happens
When you send a message to an AI, you are not just talking to a model.
Every prompt enters a pipeline. Here is what that pipeline looks like — in real time.
This pipeline varies by tool. Ollama runs entirely on your device — none of these nodes exist. Most consumer AI tools activate all of them by default.
"AI browser extensions read the full text of every page you visit while they're active. Most users have three or four installed and have never thought about this."
— Wzdom Research · AI Privacy Index Q2 2026
Global laws
Global AI privacy laws — 2026 tracker
Regulation is racing to catch up with AI deployment. Here's where each major jurisdiction stands — and what it means for how AI companies can use your data.
| Jurisdiction | Law / Framework | Status | AI-Specific? | Key user right |
|---|---|---|---|---|
| European Union | GDPR (2018) + EU AI Act (2024) | Active | Yes — AI Act adds tiered risk rules | Right to erasure, automated decision opt-out |
| United Kingdom | UK GDPR + Data Protection Act 2018 | Active | AI-specific guidance from ICO (2024) | Right to explanation for AI decisions |
| United States | FTC Act §5 + state patchwork (CCPA, VCDPA…) | Active (fragmented) | No federal AI law — FTC uses deceptive practices authority | CCPA: opt-out of sale/sharing of personal data |
| Canada | PIPEDA + AIDA (Bill C-27 lapsed 2025) | Pending | AIDA would create first Canadian AI-specific rules — C-27 lapsed when parliament prorogued Jan 2025; may be reintroduced | Explanation for consequential AI decisions (proposed) |
| China | PIPL (2021) + Generative AI Regulations (2023) | Active | Most prescriptive generative AI rules globally | Consent required for personalisation |
| India | DPDP Act (2023) | Pending enforcement | AI-specific rules in progress; enforcement timeline extended to 2026–27 | Consent and data principal rights |
| Brazil | LGPD (2020) | Active | AI guidance from ANPD in progress | GDPR-equivalent rights |
| Australia | Privacy Act 1988 + proposed reforms | Reform in progress | Specific AI amendments proposed 2024 | Right to erasure (proposed) |
United States
The US AI privacy landscape
The United States has no single federal AI privacy law — and that gap is enormous. Major AI companies are headquartered in the US specifically because it gives them maximum flexibility over data use. Here's what state laws and federal agencies are actually doing.
| Law / Action | Jurisdiction | Status | What it does |
|---|---|---|---|
| CCPA / CPRA | California | Active | Opt-out of data sale, right to deletion, AI profiling disclosure. Strongest US state law. |
| VCDPA | Virginia | Active | Opt-out of profiling for targeted advertising. Applies to AI tools with VA users. |
| American Privacy Rights Act (APRA) | Federal (proposed) | Stalled in Congress | Would create national opt-out rights for AI-driven data use. Multiple versions introduced 2023–24. |
| FTC Enforcement Actions | Federal | Ongoing | FTC has fined AI companies under deceptive trade practices rules (Section 5). No dedicated AI statute needed. |
| Executive Order on Safe AI (Oct 2023) | Federal | Active | Requires safety testing, watermarking of AI-generated content, and privacy-preserving training techniques for large models. |
| State AI Bills (TX, CO, IL, WA…) | Multiple states | Various stages | Biometric data, automated employment decisions, AI in healthcare. Illinois BIPA remains strongest biometric law. |
EU & GDPR
EU AI Act & GDPR — the world's toughest rules
The EU AI Act (entered into force August 2024) is the world's first comprehensive AI regulation. Combined with GDPR's existing data rights, it gives EU citizens more protection against AI data practices than any other population on earth. Here's what it requires of AI companies.
| Requirement | Who it applies to | Timeline | Penalty |
|---|---|---|---|
| GPAI transparency | All general-purpose AI models (ChatGPT, Gemini, Claude, etc.) | Aug 2025 | Up to €15M or 3% global revenue |
| High-risk AI requirements | AI used in HR, healthcare, credit, law enforcement | Aug 2026 | Up to €30M or 6% global revenue |
| Training data disclosure | Foundation model providers | Aug 2025 | Administrative fine |
| GDPR Right to Erasure ("Right to be Forgotten") | All data controllers processing EU personal data | Active (2018) | Up to €20M or 4% global revenue |
| GDPR Automated Decision Opt-Out | Decisions based solely on automated processing | Active (2018) | Up to €20M or 4% global revenue |
This index is published by Wzdom Research — independent, vendor-free, and source-cited. No company pays for a score. Every finding links to its primary source. About our methodology and independence →
Intelligence feed
AI Privacy Intelligence — Q2 2026
Tracked changes, regulatory actions, and policy shifts. Updated as the landscape changes.
For researchers
For privacy researchers, journalists, and content creators
This page is designed to be useful, linkable, and quotable. If you write about digital privacy, security, or AI — here's how we can make your work easier.
Full research paper
The complete scored analysis of 12 AI tools with methodology, primary citations, regulatory timeline, and company-by-company findings. Free to download.
Download PDF ↗Permalink anchors
Every tool has a permanent anchor (e.g. /ai-privacy#chatgpt) so you can deep-link directly to a tool's score in articles or forum posts.
Primary source citations
Every score traces to a primary source: privacy policies, regulatory filings, academic papers, or enforcement records. Every claim in the research paper is independently verifiable.
How to cite us
Findings may be quoted with attribution to Wzdom Research / Wzdom Labs. Suggested citation: Ahmed, M. (2026). Privacy in the Age of Frontier AI. Wzdom Labs, Q2 2026. wzdom.ai/research
Press and media
For interview requests, embargoed briefings, or background commentary on AI privacy topics, contact research@wzdom.ai. We respond within one business day.
Corrections welcome
If you have primary source evidence that contradicts a score or finding, we want to hear it. Send your source to research@wzdom.ai — corrections are published with attribution.
Protect yourself
10 ways to protect yourself from AI data collection right now
You don't need to stop using AI tools. You need to use them more carefully. These steps reduce your exposure significantly without sacrificing productivity.
-
1
Opt out of training data use
Every major AI tool has a setting. ChatGPT: Settings → Data Controls → off. Gemini: Activity Controls. It takes 60 seconds per tool.
-
2
Never paste passwords, keys, or personal ID into AI chats
Prompts are logged. Even "temporary" conversations are stored server-side before any deletion. Treat every prompt as a support ticket.
-
3
Use a work account for work AI — not personal
Enterprise/Business tiers generally offer better data controls. Personal accounts often have weaker protections and less clarity on retention.
-
4
Run sensitive queries through local models
For legally sensitive, medically sensitive, or commercially sensitive queries, use Ollama or LM Studio on your own hardware. Nothing leaves your device.
-
5
Disable AI browser extensions when not in use
AI assistants in your browser read the full text of every page while active. Disable them except when needed. Use a separate browser profile.
-
6
Review AI tool permissions in your OS
Check System Settings → Privacy on macOS and Windows. AI apps often request camera, microphone, accessibility, and file system access. Revoke what's unneeded.
-
7
Monitor what your AI tools send over the network
Tools like Wzdom Privacy show every outbound connection in real time — so you know exactly when an AI tool is phoning home and where it's going.
-
8
Exercise your deletion rights
Submit GDPR/CCPA deletion requests annually for tools you no longer use. Most companies have a Data Deletion form in their Privacy Center.
-
9
Read the sub-processor list, not just the privacy policy
The privacy policy tells you what the company does. The sub-processor list tells you who else gets your data. They're usually different pages.
-
10
Stay informed — policies change constantly
OpenAI, Google, and Microsoft update their privacy policies multiple times per year. Subscribe to our Live Scorecard alerts to know when something material changes.
Coming in v0.2
Wzdom AI Shield — privacy for the AI era
AI Shield intercepts traffic to ChatGPT, Copilot, Gemini, and Grok before it leaves your machine. It strips personally identifiable data from prompts, monitors what each AI knows about you in real time, and lets you route sensitive queries to fully local models with one click.
Wzdom Privacy
Your AI tool has a grade.
Your apps don't have a monitor. Yet.
The AI Privacy Index tells you what's in the policy. Wzdom Privacy shows you what's actually leaving your machine — every connection, every app, in real time.
macOS & Windows · Free during beta · No account needed