iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Duck.ai, Proton Lumo, and Brave Leo are three privacy-conscious alternatives to consider, but the available provider policies do not prove that one is objectively the safest. Each describes different controls: DuckDuckGo says it strips personal metadata before sending prompts to model providers; Proton says it encrypts saved chat history with zero-access encryption; and Brave’s retention terms vary by model. Compare training, retention, human review, metadata, saved history, and model processing separately. The details below reflect the companies’ official pages as accessed in 2026, not independent verification.
What should you compare when choosing a private AI chatbot?
“Private” is not a single technical setting. A service might say it does not train on chats while still processing prompts on its servers, temporarily retaining data, or applying different terms to different models. To assess the trade-offs, look at six separate questions:
- Training: Does the provider say prompts or responses can be used to train models?
- Retention: Are chats stored, cached, or kept temporarily, and for how long?
- Human review: Can people review chats, and are there exceptions to deletion rules?
- Metadata: Does the service remove identifying details, such as an IP address, before a third-party model provider receives the prompt?
- Saved history: Where is it kept, and what protection applies to stored chats?
- Model processing: Which model or provider handles the request, and do its terms differ from the chatbot’s headline policy?
These controls answer different questions. For example, encryption for stored history does not by itself describe how a prompt is processed, and a no-training promise does not establish that nothing is temporarily retained.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
It also helps to separate a provider’s promise from a measurable outcome. “Not used for training” is a rule about permitted use, not proof that a prompt never reaches another company or remains in a short-lived cache. “Deleted after a reply” describes a stated lifecycle, while human-review exceptions may create a different lifecycle for a subset of conversations. Read each sentence in context, including the model, account type, and feature to which it applies.
#1 Best Overall
How do Duck.ai, Lumo, and Brave Leo compare?
The three services make different provider-stated privacy commitments, so the most useful choice depends on which controls matter to you. This comparison summarizes the official policy pages available in 2026; terms, models, and features can change.
| Service | Training use | Chat retention | Human review | Metadata before model processing | Saved history | Model or provider handling the request |
|---|---|---|---|---|---|---|
| Duck.ai | DuckDuckGo says its agreements with model providers prohibit using chats for training. | DuckDuckGo says chats are not recorded or stored by default. Its help page notes prompt caching for certain OpenAI requests for up to one hour. | Not stated in the cited Duck.ai privacy help page. | DuckDuckGo says it removes personal metadata, such as IP addresses, before sending prompts to model providers. | Not stated in the cited help page beyond the default claim that chats are not recorded or stored. | Prompts are sent to model providers. DuckDuckGo says its agreements require providers to delete prompts and responses after generating a reply; the cited help page does not establish one fixed provider for every request. |
| Proton Lumo | Proton says Lumo does not use chats to train its models. | Proton says it erases logs after each chat. | Not stated in the cited Lumo privacy page. | Not stated in the cited Lumo privacy page. | Proton says saved chat history uses zero-access encryption. | Proton says prompts are processed on its GPU servers. The cited page does not identify a third-party model provider for every request. |
| Brave Leo | Brave says chats are not used for training in its overview. | Brave’s overview says chats are not retained, but its model-specific help page says conversations using listed third-party Claude models may be retained for no more than 30 days. Brave-hosted models are described separately. | Not stated in the cited Brave overview and model-differences page. | Not stated in the cited Brave overview and model-differences page. | Not stated in the cited pages. | Leo is integrated into Brave’s browser. Available models include listed third-party Claude models and Brave-hosted models; the applicable terms depend on the model. |
Sources: DuckDuckGo’s Duck.ai privacy help, Proton’s Lumo privacy page, Brave’s AI overview, and Brave’s explanation of Leo model differences. “Not stated” means the cited material does not establish that detail; it does not prove the practice never occurs.
How should you interpret retention and deletion language?
Retention can occur at several layers. A chat may appear in your account history, exist in an encrypted database, remain in an operational log, or be cached briefly by a model provider. Those layers can have different clocks and different deletion mechanisms. When a policy says “erased after each chat,” check whether it is referring to logs, saved history, or both. When it says “not retained,” check whether a particular model, safety process, or troubleshooting feature is excluded.
A deletion control also may not erase copies already separated for a stated purpose. Google’s Gemini policy, for example, says material reviewed by human reviewers and related data can remain for up to three years and is not deleted when a user deletes activity. That is why a short account-history setting should not be treated as a universal eraser. The same discipline applies to the three services here: record the exact duration and exception rather than converting a headline into an absolute guarantee.
What does Duck.ai say it protects?
DuckDuckGo describes a combination of default non-storage, metadata removal, and terms imposed on model providers. Its help page says Duck.ai does not record or store chats by default, removes personal metadata such as IP addresses before prompts are sent to model providers, and requires providers to delete prompts and responses after generating a reply and not use chats for training.
Removing an IP address before forwarding a prompt can reduce the identifying information visible to the downstream model provider. It does not make the text itself anonymous. A prompt that includes a customer name, a unique incident number, a private URL, or a distinctive fact can still identify a person or organization. Redact those details before sending them, even when metadata stripping is part of the service design.
Rank #2
There is a limited retention qualification: DuckDuckGo says certain OpenAI requests may involve prompt caching for up to one hour. That is a stated maximum for those requests, not a general retention period for all Duck.ai conversations. The provider terms are DuckDuckGo’s claims; the cited page does not establish an independent audit of compliance.
Free tools Windows power users keep installed
One-click scans. No signup required.
Duck.ai therefore may suit someone who prioritizes minimizing identifying network metadata while using a hosted model. It is still a relay to model providers, and the exact model can affect the applicable terms. Check the model label and current privacy help before assuming every request follows an identical path.
How does Lumo’s saved-history encryption differ from prompt processing?
Proton says Lumo does not train on chats, erases logs after each chat, and protects saved chat history with zero-access encryption. It also says prompts are processed on Proton’s GPU servers. Those statements concern distinct stages: protecting stored history is not the same as describing what happens while a prompt is being processed.
Zero-access encryption generally means the service presents stored history in encrypted form that Proton says it cannot decrypt under its design. That is relevant when an account contains conversations you deliberately save. It does not mean the model can answer without seeing the prompt in usable form during processing. The GPU-server statement identifies where that processing occurs, but the cited privacy page does not establish every technical detail, such as all transient buffers, operational logs, or reviewer access.
If your priority is encrypted saved history, Proton’s stated zero-access approach is relevant. If your concern is where a prompt is handled, the GPU-server statement is the relevant detail. Ask separately whether you need history at all: avoiding saved transcripts can reduce the amount of material exposed if an account is later compromised, while encryption protects the history you choose to retain.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →Why should you check which model Leo is using?
Brave presents Leo as an AI assistant integrated into its browser and says chats are not retained or used for training in its overview. However, the model-specific help page qualifies the retention claim: conversations using listed third-party Claude models may be retained for no more than 30 days, while Brave-hosted models are described separately.
That difference makes the selected model important. Before using Leo for a sensitive task, open the current model selector and read the terms attached to that model. A browser-integrated assistant can offer convenience, but the request may be handled under a third-party model arrangement rather than solely by Brave. Switching models can therefore change the retention statement that applies, even if the interface still says Leo.
Brave’s pages do not establish comparable details for every privacy dimension in the table, including human review, metadata removal, and saved-history protection. Treat those fields as unanswered questions, not as evidence that the most protective option is in place. If you cannot identify the model or its retention rule, avoid entering information whose exposure would cause harm.
How does Gemini illustrate the difference between deletion and review?
Google’s Gemini Apps Privacy Hub offers a useful comparison point because it describes both an activity setting and a separate retention exception. For signed-in Gemini Apps Activity, Google says the default auto-delete period is 18 months, with options for three months, 36 months, or indefinite retention. Google also says chats reviewed by human reviewers and related data may be retained for up to three years and are not deleted when a user deletes activity.
This does not mean every Gemini chat is reviewed. It shows why an activity-deletion setting should not be read as a complete answer about reviewer-related retention. The periods are Google’s stated policy settings and limits, not measured outcomes. See Google’s Gemini Apps Privacy Hub for the current details.
Use this example as a checklist for any chatbot. Find the normal history period, then look for separate language covering safety investigations, human review, abuse prevention, legal obligations, or model improvement. If an exception survives account deletion, document it before deciding whether the service fits a regulated or confidential workflow.
What practical steps reduce exposure?
- Classify the prompt. Decide whether it contains personal data, confidential business material, health information, credentials, or information that could identify someone indirectly.
- Minimize the text. Replace names with roles, remove account numbers, generalize dates and locations, and paste only the excerpt needed for the task.
- Identify the processing model. In Leo, check whether the selected model is Brave-hosted or one of the listed third-party Claude options. In Duck.ai, note that the request is sent to a model provider. In Lumo, note Proton’s statement that prompts run on its GPU servers.
- Check the applicable retention rule. Look for a duration, caching exception, and any separate treatment for saved history or reviewed material.
- Choose history deliberately. If you do not need a transcript, do not save it. If you do need one, understand the stated protection, such as Lumo’s zero-access encryption for saved history.
- Keep an internal record. For workplace use, record the service, model, account type, and policy date so colleagues know which terms were assessed.
- Delete and verify. Use the service’s deletion control, then check whether the policy describes exceptions that are not removed with account activity.
Do not submit sensitive information merely to test a chatbot. A harmless, synthetic prompt can reveal whether a feature works without creating a new confidential record.
Rank #4
What are the common privacy pitfalls?
- Equating no training with no retention. A provider can prohibit training while allowing a short cache or a separate safety-review period.
- Assuming encryption covers active prompts. Encryption for saved history addresses stored data; it does not describe what the model sees while generating an answer.
- Ignoring the downstream provider. A relay may remove metadata yet still send the prompt to another company under contractual terms.
- Reading a service-wide headline instead of a model rule. Leo’s third-party Claude qualification shows why model selection matters.
- Inferring anonymity from no sign-in. Account-free use does not establish what network, device, or prompt information is collected.
- Treating “not stated” as “none.” If a page does not explain human review or metadata handling, the responsible conclusion is that the detail is unknown.
- Testing with real secrets. Use fictional names and data when evaluating a new assistant.
What about availability, plans, and cost?
Privacy policies do not provide a reliable cross-service price comparison. Features, model access, account requirements, and limits can vary by plan, country, browser version, or account setting. Check each provider’s current product page for availability and pricing rather than assuming that a privacy feature is included in every tier. Brave’s browser and the chatbot services may have separate commercial terms; if you link to or recommend a paid service, verify any affiliate arrangement and disclose it.
How should you choose among these services?
Start with the kind of data you expect to share, then match your priority to a specific stated control. None of the provider pages reviewed establishes an objectively safest chatbot or a comparable independent audit result, so a single overall winner would overstate what is known.
- Decide what you need to protect. Distinguish between keeping a saved history private, limiting training use, minimizing identifying metadata, and limiting how long a provider retains a prompt.
- Check the current model and feature. This is especially important where a service offers multiple models or routes requests to an external provider. Brave’s model-specific retention terms illustrate why the service name alone may not be enough.
- Read the relevant policy for the specific control. Look for the retention period, exceptions, review terms, and whether the claim applies to stored history or active processing. Treat a missing detail as unknown rather than assuming the most protective interpretation.
- Share less when the policy does not answer your concern. Remove names, account numbers, unpublished business information, and other identifying details where they are not needed. Do not submit sensitive data just to test a chatbot.
- Recheck after a policy or model change. Availability, settings, and terms can vary by account, plan, model, or feature, and official pages may be updated.
These are provider-described controls, not independently verified guarantees. If a policy leaves a material question unanswered, choose a different workflow or avoid entering that information.
Frequently Asked Questions
Does private browsing make an AI chatbot private?
Private browsing can affect what your browser saves on your device, but it does not by itself set a chatbot provider’s training, retention, or review practices. Check the service’s own policy and account settings for those controls.
Is an open-source AI model automatically more private?
No. Access to a model’s code or weights does not tell you how a hosted service handles prompts, logs, or account data. Privacy depends on the way the model is deployed and the operator’s policies.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Can I use a privacy-focused chatbot for confidential work?
Only if your organization’s rules permit it and the service’s current terms meet your obligations. A provider’s privacy statements do not replace your employer’s data-handling policy or a review of applicable confidentiality requirements.
Best Value
Does using a chatbot without an account make a conversation anonymous?
Not necessarily. An account-free interface does not establish what network or device data is collected or what model provider receives a prompt. Look for explicit terms covering those details rather than inferring anonymity from the sign-in flow.
Which of Duck.ai, Lumo, and Leo is the most private?
The cited provider pages do not establish an objectively safest service or a comparable independent audit. Duck.ai emphasizes metadata removal and provider deletion terms, Lumo emphasizes erased logs and encrypted saved history, and Leo requires attention to the selected model. Choose the control that matches your priority.
Does Duck.ai keep any prompts?
DuckDuckGo says chats are not recorded or stored by default, but it notes prompt caching for certain OpenAI requests for up to one hour. That is a limited exception, so check the current model and policy before sending sensitive material.
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Clear out junk files and repair common Windows errors3Scan for outdated or missing drivers - takes under a minuteDoes Lumo’s zero-access encryption mean Proton cannot process my prompt?
No. Proton describes zero-access encryption for saved chat history and separately says prompts are processed on its GPU servers. Stored-history protection and active prompt processing are different stages.
Why can Leo’s retention depend on the model?
Brave’s model-specific help page says conversations using listed third-party Claude models may be retained for no more than 30 days, while Brave-hosted models are described separately. The model selector and its current terms therefore matter.
Are the stated retention periods guarantees?
They are limits or settings stated in provider policies, not independently measured privacy outcomes. Policies can include exceptions, and terms may change by model, plan, account, or feature.
Should I test a chatbot with real customer or health data?
No. Use synthetic examples while evaluating a service. If you later use it for legitimate work, minimize the prompt and confirm that your organization’s rules and the current provider terms allow the data.
The Bottom Line
Duck.ai, Proton Lumo, and Brave Leo each emphasize a different part of privacy: Duck.ai says it removes personal metadata and limits provider use, Lumo describes erased logs and zero-access encryption for saved history, and Leo’s terms depend on the chosen model. None is proven the universal safest choice. Compare training, retention, review, metadata, storage, and processing separately, minimize what you send, and recheck the current policy before using any service for sensitive information.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
