We Read the Fine Print: What AI Companies Say (or Don’t Say) About Prompt Privacy

When you type into an AI model, where does that data go? Who sees it? Can it train the AI? And how long is it stored? 

These aren’t hypothetical concerns. In 2025, prompt data is a business asset — and a potential liability. 

We dug deep into the actual privacy policies, terms, and legal pages of the biggest AI companies. What we found: a sharp divide between vendors who give you control, and those who bury risks in legal grey zones. 

The Context: Why Prompt Privacy Matters in 2025 

• AI adoption is mainstream — Startups, enterprises, freelancers, and students all use AI tools daily. 
• Prompt data often includes proprietary or sensitive input — customer info, strategies, internal research. 
• Many users wrongly assume their prompts are private by default. They’re not. 

What a company does with your prompt — whether it stores it, trains on it, or shares it — affects your IP security, compliance risks, and even brand safety. 

Methodology: What We Looked At 

We analyzed six leading platforms: 

• OpenAI (ChatGPT) 
• Anthropic (Claude) 
• MidJourney 
• Jasper.ai 
• Copy.ai 
• Leonardo.ai (covered in shorter form at the end) 

We evaluated them across six dimensions: 

OpenAI (ChatGPT): Power, But With Persistent Retention 

Prompt Storage 

Yes, stored by default. 

ChatGPT chats are retained indefinitely unless manually deleted. API logs are stored for 30 days. 

April 2024 change: Free/Plus users can no longer fully disable chat history. Deletion must be manual. 

Training Use 

• Yes, by default. 
  Unless you toggle off “Improve model performance,” your inputs can be used to train future ChatGPT versions. 
• Enterprise and API clients are excluded from training by default. 

User Control 

• You can opt out of training use. 
• But OpenAI still retains your data unless you delete it manually or use Temporary Chats (which delete after 30 days). 

Data Residency 

• Global, mostly U.S.-based. 
  No regional storage guarantees for most users. Some EU-specific processing introduced in 2023. 

Legal Protections 

• You “own” your input/output. 
• But OpenAI may allow vetted staff to access your data for moderation or product improvement. 
• No full confidentiality guarantee unless you’re on Enterprise or API tier. 

Key Risk 

Unless you’re on an Enterprise plan, your data is stored indefinitely and can be used to train the model. 

Anthropic (Claude): Privacy-First, Opt-In Only 

Prompt Storage 

• Yes, but time-limited. 
  Most unflagged prompts are kept no longer than 2 years, and users can delete anytime. 
• Flagged content (policy violations) is stored longer — up to 7 years for safety metadata. 

Training Use 

• No, not by default. 
  Claude only trains on your data if you opt in or submit feedback (e.g., thumbs-up/thumbs-down). 
• Flagged prompts may be used for moderation training, not for core model training. 

User Control 

• You’re opted out by default. 
• Prompt deletion is fast: removed from user history immediately, backend systems within 30 days. 

Data Residency 

• Global, but GDPR-compliant. 
  Data may be stored in the U.S., but Anthropic uses standard clauses for lawful EU transfers. 

Legal Protections 

• Anthropic does not claim ownership over your content. 
• Only limited, need-based access by Trust & Safety staff. 
• Clear language, strong SOC2 compliance, and no vague loopholes. 

Key Strength 

Claude is one of the most privacy-protective LLMs on the market. Even free users get strong safeguards. 

MidJourney: The Public Gallery by Default 

Prompt Storage 

• Yes — permanently and publicly. 
  Unless you pay for “Stealth Mode,” everything you type and generate is visible on their platform. 
• Even deleted accounts may leave public remnants (e.g., in Discord logs or remixable content). 

Training Use 

• Likely yes. 
  The Terms grant MidJourney a perpetual license to use your content for any purpose, including model improvement. 
• There is no opt-out. 

User Control 

• Limited. 
  Only paying users can activate privacy controls (Stealth Mode). Even then, prompts are still stored. 

Data Residency 

• U.S.-based, unclear regional handling. 
  No support for EU-specific storage. Discord infrastructure adds another layer of exposure. 

️ Legal Protections 

• Free users don’t even own their outputs. 
  MidJourney retains copyright over images unless you’re a Pro subscriber. 
• No confidentiality promised. They advise: “Don’t input anything private.” 

Key Red Flag 

MidJourney is the least private platform. Everything is public by default unless you pay. 

Jasper.ai: Enterprise-Ready and Private by Design 

Prompt Storage 

• Yes, but time-bound. 
  Data is retained as needed. Deletion tools (via Transcend or in-app) are available. 
• Enterprise users can request custom deletion timelines. 

Training Use 

• No. 
  Jasper explicitly states: “Your data is never used to train third-party LLMs.” 
• Internally, they may use aggregated stats (non-identifiable) to improve Jasper features. 

User Control 

• You don’t need to toggle anything — no training is the default. 
• You can delete prompts, documents, or your entire account anytime. 

Data Residency 

• Primarily U.S., with SOC2 compliance. 
  No default regional silo, but enterprise customers can negotiate DPA terms for location-specific storage. 

Legal Protections 

• You own your input and output. 
• Jasper’s DPA, ToS, and Trust Center commit to confidentiality, data segregation, and enterprise compliance. 

Key Advantage 

Jasper is ideal for teams needing strict data boundaries. Even free trial users benefit from the no-training policy. 

Copy.ai: Mostly Safe, But Less Transparent 

Prompt Storage 

• Yes, with no clear deletion policy unless initiated by the user. 
• Prompts are stored for functionality (editing, brand voice) and can be removed manually. 

Training Use 

• Unclear. 
  Copy.ai likely does not train public LLMs with your data, but they may use inputs to improve features. 
• No explicit statement or opt-out. 

User Control 

• You can delete prompts, docs, or your account. 
• No setting exists to stop internal use of prompts for “service improvement.” 

Data Residency 

• U.S.-based with global cloud use. 
  No regional promises. GDPR covered via SCCs, but nothing explicit about EU storage options. 

️Legal Protections 

• Users own their inputs and outputs. 
• No clear confidentiality language — internal access or use isn’t tightly defined. 

Key Concern 

Copy.ai is likely safe for casual use, but lacks the clarity or contracts that enterprise buyers might expect. 

Leonardo.ai: Public Showcase, Minimal Privacy 

• Public by default. Prompts and images often appear in community feeds unless you’re on a private plan. 
• Training use is unclear, but Terms allow wide reuse. 
• No opt-out or region-specific guarantees. 
• Not suitable for confidential work unless you pay for privacy and get written terms. 

Perplexity.AI: The Search-Like LLM With Blurry Privacy Lines 

Prompt Storage 

• Yes, prompts are stored. 
  Although Perplexity appears search-like in UX, it retains your queries and outputs. Their privacy policy confirms collection of prompt content, usage logs, and metadata. 
• Logged data is used for analytics, safety, and product improvement — and retention periods are not clearly disclosed. 

Training Use 

• Yes, partially. 
  Perplexity’s terms allow usage of submitted queries to improve their service. Whether that means fine-tuning models or only analytics is not defined. 
• There is no prompt opt-out setting for training or analytics. 

User Control 

• Very limited. 
  You can delete your account, but individual prompt deletion or training exclusion is not available via the interface. 
• No “privacy mode” or session-only queries unless using special API endpoints (unclear to the average user). 

Data Residency 

• Likely U.S.-based. 
  Perplexity uses cloud infrastructure (presumably AWS or GCP), but does not specify server locations or regional controls. 
• Standard international transfer clauses are included, but there’s no user-facing toggle for region-specific processing. 

Legal Protections 

• The platform does not claim ownership of your prompts — but grants itself broad rights to reuse and process them. 
• No public SOC2 or ISO certification pages are available, and there’s no DPA or privacy contract flow for teams or enterprises. 
• Notably, Perplexity warns users not to input sensitive or regulated data, which implies limited internal safeguards. 

Key Ambiguity 

Despite its “search” appearance, Perplexity is storing and possibly learning from your prompts — with no opt-out or visibility control. 

It’s fast and helpful, but not built for secure or confidential use. 

What Founders Should Know About “Prompt Leakage” Risks 

What Is Prompt Leakage? 

Prompt leakage refers to unintended exposure or reuse of inputs given to an AI system — often due to: 

• Model training on prompts without explicit consent 
• Prompts being visible to other users (e.g., MidJourney, Leonardo) 
• Staff or contractors having backend access 
• Outputs reproducing phrases or ideas from earlier prompts 

Real-World Risk 

In early 2024, a marketing agency unknowingly uploaded a client’s full product roadmap into a free AI image generator. Months later, similar visuals appeared in public feeds — generated by others using loosely related prompts. 

No one “hacked” anything. The platform just used prompt data to improve its models. 

For startups, especially in stealth or IP-heavy industries, this is a real liability. 

What Founders Must Do 

• Do not use public-facing AI tools for confidential prompts unless you’ve confirmed: 
• No training without consent 
• No public access 
• Legal protections via DPA or terms 
• Get a Data Processing Addendum (DPA) before integrating AI into workflows 
• Check vendor SOC2, ISO 27001 status 
• Educate teams: Your intern pasting dev notes into ChatGPT could cost you investor trust 

Better Practice 

Use Claude, Jasper, or OpenAI Enterprise if confidentiality matters. Otherwise, assume what you type may resurface — directly or indirectly. 

Privacy ≠ Compliance: Why Founders Still Need Legal Guardrails 

Don’t Confuse Policy with Legal Protection 

Just because a company says “we don’t train on your data” doesn’t mean: 

• They’re encrypting prompts at rest 
• They’ve signed contractual guarantees 
• They’ll delete data when you ask — unless legally bound 

A Privacy Policy is not a legal agreement unless you have a signed contract. 

What B2B Founders Should Demand 

If your AI use touches customer data, internal IP, or third-party integrations, make sure to: 

1. Negotiate a DPA 
   This gives you control over data retention, access, deletion, and location. 

2. Ask for regional processing 
   EU or healthcare-focused users may require in-region handling (ask for SCCs or HIPAA language). 

3. Document your vendor stack 
   Many AI tools use OpenAI, Anthropic, or Cohere in the backend. 
   You need to know who touches the data at every step. 

4. Prepare for audits 
   If you’re in fintech, edtech, or health, regulators will ask where data flows. Don’t wait for that email. 

Practical Tools 

• Use platforms like Transcend, OneTrust, or Ethyca to automate DPA and consent workflows 
• Use Claude or Jasper for prompt-level privacy control 
• Avoid platforms without export logs or legal contacts 

Conclusion: Which AI Company Respects Prompt Privacy? 

How Spaculus Software Helps 

AI adoption is no longer just about speed, it’s about control. At Spaculus, we help businesses use AI without exposing their data, IP, or strategy. 

Safe AI Workflows 

We integrate AI using tools like Claude, OpenAI Enterprise, or private models — so your prompts never train public systems unless you approve it. 

Legal-Ready Infrastructure 

We help teams sign DPAs, map data flows for GDPR/SOC2, and build audit logs so compliance isn’t an afterthought. 

Transparent by Design 

Every AI workflow we build shows where your prompt went, who processed it, and what’s stored — so you always stay in control. 

Deployment Flexibility 

Need on-prem or VPC deployment? We’ve done it. From open-source LLMs to secure cloud stacks, we tailor AI to your risk level. 

Final Thought 

If you wouldn’t write it on a postcard, don’t paste it into an AI tool — unless you’ve read the fine print. 
Too many platforms still treat prompt data as raw material. But companies like Anthropic and Jasper are showing what responsible AI stewardship looks like. 

Want complete prompt control? Pay attention to policy — not just product.