#1654 Suggestion to DeepSeek: Balance Safety Policies, Add Age Verification, Fix Refusal UX, and Open Global Support

Product Policy & Support Suggestions for DeepSeek

Target product: DeepSeek Chat (Web & Mobile) + DeepSeek API, with focus on DeepSeek-V4.1-Flash Audience: DeepSeek Policy / Safety / Product / Support teams Purpose: Reduce false-positive blocks, improve adult-user experience, and restore reachable global support

Executive Summary

DeepSeek-V4.1-Flash is a strong model, but the current safety and support stack feels rigid rather than balanced. Legitimate educational, creative, technical, and mature adult conversations are being treated the same as clear policy violations. Users receive vague in-chat refusals, suspensions and global users who need human help are often locked out by Feishu/Lark phone-number restrictions (+86 / +82).

This letter proposes four concrete updates:

Rebalance user policies so they are strict on real harm, not overly strict on normal use.
Introduce universal age verification, with an 18+ policy tier for verified adults.
Block clear policy-violating prompts at the chatbox, before they are sent.
Replace region-locked support with in-app Contact Us + open email support.

1. Rebalance User Policies (Not Rigid / Not Over-Tight)

Current Issue

Safety alignment and filters currently appear to treat sensitive keywords as automatic high-risk violations. This causes false positives on:

Educational discussions
Creative writing
Technical / research topics
Mature but legal adult conversations

Suggested Change

Update DeepSeek user policies from a blanket restriction model to a balanced, context-aware model:

Area	Keep strict	Allow more context
Illegal activity / violence / CSAM / doxxing	Hard block	—
Medical, legal, psychological discussion	Safe guidance required	Full educational discussion allowed
Creative fiction / roleplay	Harm rules still apply	Mature themes allowed for 18 or 18+
Technical security / coding	Block real attack enablement	Research and defensive discussion allowed
Politics / religion / sensitive news	No harassment or incitement	Balanced analysis allowed

Why This Matters

Users should not feel punished for asking normal, complex, or adult-level questions. Policy should protect against real harm without making everyday use feel locked down.

2. Universal Age Verification + 18 or 18 plus Policy Tier

Proposal

DeepSeek should introduce an age verification system for everyone, not only in selected regions.

Suggested Flow

User creates or upgrades an account.
User completes age verification (ID check, privacy-preserving verification, or equivalent).
Account is tagged as:
- Under 18 → default strict safety policy
- 18+ or 18 verified → adult policy tier
Verified 18 or 18+ users can discuss sensitive, mature, or adult-oriented topics that are legal and policy-compliant.

Design Principles

Verification must be opt-in but available to all eligible users.
Store only the minimum verification result (for example: verified + age band), not unnecessary identity documents.
Make the active policy tier visible in account settings.
Provide a clear appeal path if verification or policy routing fails.

Precedent

Other major assistants, including Claude & Gemini, allow verified adults to handle mature topics under appropriate safeguards and even allow you to do personal conversation too as your ease or if you feel comfortable to do that. DeepSeek should match that standard. Being 18 or 18+ does not mean unlimited harmful content — it means adult users get age-appropriate policy, not the same tight default as minors or unverified accounts.

3. Block Policy-Violating Messages at the Chatbox

Current Issue

Users can type a clearly policy-violating prompt, submit it, and only then see a refusal such as:

"That's beyond my scope. I can't talk on it for your safety..."

This is late, vague, and wasteful.

Suggested Change

DeepSeek Web and Mobile should include a pre-send policy detection layer:

Behavior

Detect clear policy violations before the message is accepted as a prompt.
Do not submit the message to the model.
Show an immediate UI warning in the chatbox:

Your message appears to violate DeepSeek user policy and was not sent.
Please rephrase your message.

Implementation Notes

Use a fast local/edge check for obvious violations (keywords + lightweight semantic rules).
Use a stronger server-side pre-inference check for borderline cases.
Keep the warning clear and non-punitive.
Do not silently drop the message — always explain why it was not sent.

Benefit

Users get instant feedback.
Chat history stays clean.
Inference resources are not wasted on prompts that will only return a refusal.
The product feels intentional, not clunky.

4. Stop Vague In-Chat Refusals After Submission

Current Issue

Even when a message is accepted and processed, DeepSeek often answers with a generic refusal:

"That's beyond my scope. I can't talk on it for your safety..."

This is frustrating because:

The refusal is too broad.
It does not say which policy was triggered.
It does not suggest how to rephrase.
It can appear even for benign prompts that only looked sensitive.

Suggested Change

If a request must still be refused after submission:

Explain the category clearly.
Offer a rephrase path.
Link to the relevant policy section.
For API users, return a structured policy error, not only free-text refusal content.

Example chat response:

I can’t help with this request under DeepSeek policy because it appears to
fall under [policy category]. If your goal is educational, creative, or
research-related, please rephrase with that context and I can try again.
See: [User Policy Link]

Example API response (instead of 200 OK + refusal text only):

{
  "error": {
    "type": "policy_violation",
    "code": "content_blocked",
    "message": "Request rejected by content policy.",
    "details": {
      "category": "example_category",
      "appeal_available": true
    }
  }
}

5. Fix Global Customer Support

Current Issue

DeepSeek support is difficult for international users because Feishu/Lark account creation is tied to limited country-code phone numbers, especially +86 and +82. Many users worldwide cannot create an account, which means they cannot easily contact support at all.

Suggested Changes

5.1 In-App "Contact Us" Support

Add a built-in support portal in:

DeepSeek Web
DeepSeek iOS / Android apps

Features:

Account-linked ticket creation
Policy appeal form
Bug report form
Billing / API support form
Ticket status tracking
No Feishu account required

5.2 Open Email Support

Publish dedicated support email addresses that work with Gmail, Outlook, Proton, and other normal providers, for example:

[email protected] — general support
[email protected] — false-positive / policy appeals
[email protected] — developer and API issues

These should be monitored and linked from the app, website, and API docs.

5.3 Keep Existing Channels, but Do Not Depend on Feishu Alone

Feishu can remain one optional channel for users who already have access. It should not be the primary or only path for global support.

Expected Outcomes

If DeepSeek implements these changes, users should see:

Fewer false-positive blocks on normal educational, creative, technical, and mature topics.
Verified adults receiving an appropriate 18+ policy experience.
Clear chatbox warnings when a message cannot be sent.
Less wasted inference on auto-refused prompts.
Reachable support for users in every region, not only Feishu-supported phone markets.

Request to DeepSeek

Please consider updating the product in this order:

P0: Improve refusal messaging and pre-send chatbox checks.
P0: Open global support via in-app Contact Us + email.
P1: Publish a clearer, more balanced user-policy framework.
P1: Launch universal age verification and an 18+ policy tier.
P2: Add developer-visible policy category codes and appeal links.

DeepSeek already has strong model capability. The next step is making safety, policy communication, and support feel equally strong for real users around the world.

I will be Thankful if these suggestions are reviewed and considered, your loyal and fellow DeepSeek User.