Yes, but not the kind cloud chatbots use. An offline chatbot has no moderation server checking your messages and no company reviewing your chats. What it will and won’t answer comes from how the model was trained, so the limits differ from one model to the next. Cloud chatbots add more layers on top: systems that screen what you type and what the model replies, usage policies enforced against your account, and in some cases human reviewers.
How do cloud chatbots filter what you ask? #
A cloud chatbot like ChatGPT, Gemini or Claude usually has several layers between you and the answer:
- Safety training in the model. The lab trains the model to decline certain requests and to handle sensitive topics carefully.
- Separate classifiers. Many services run your prompt and the model’s reply through extra systems that flag categories such as violence, self-harm or harassment. OpenAI offers its moderation model to other developers as a standalone service.
- Account enforcement. Usage policies let providers warn, restrict or ban accounts that break the rules.
- Human review. Some providers have people read a sample of conversations. Google says reviewers read a subset of Gemini chats and keep the reviewed ones for up to three years (Gemini Apps Privacy Hub).
- Local law. Providers adjust what their services say by country. China, for example, has required generative AI services to follow content rules since August 2023.
These layers change over time without notice, which is one reason the same question can get a different response next month.
Where does an offline model’s filtering come from? #
When a model runs on your phone, only the first layer comes along. The model file is a set of weights: there’s no classifier between you and it and no server for it to report to. Its limits come from four places:
- The lab’s safety training. Google DeepMind (Gemma), Alibaba (Qwen), Mistral AI (Ministral) and Z.ai (GLM) each tune their models differently, so the models refuse different things and phrase refusals differently.
- The model’s license. Open-weight licenses and use policies set terms for how a model may be used. They’re legal terms rather than technical blocks, but they still apply to you.
- Your instructions. A system prompt sets the role, tone and format for a chat.
- Its size. Small models can be clumsy here. They may refuse a harmless question that sounds like a risky one, or answer carelessly where a larger model would add a caveat.
| Cloud chatbot | Offline chatbot | |
|---|---|---|
| Who decides what’s allowed | The provider | The lab that trained the model |
| Extra screening of prompts and replies | Often | None |
| Human review of your chats | Possible | Nobody else has your chats |
| Account penalties | Possible | There’s no account |
| Behavior changes without notice | Yes | Only if you download a different model |
| Change tone with a system prompt | Somewhat | Yes, per chat |
Why do different offline models refuse different things? #
Because they come from different labs, trained on different data with different priorities. A question about a medication dose might get a careful answer from one model and a flat refusal from another. Political questions can differ too: models from China-based labs are built under Chinese rules for generative AI, so their answers on some political topics can be more guarded. We look at that in detail in is Qwen safe to use.
If one model refuses something reasonable, trying another is normal. In Personal LLM you can switch models from the chat header in the middle of a conversation and put the same question to Qwen 3.5, Gemma 4 and Ministral 3 to compare.
What can a system prompt change? #
A system prompt is a standing instruction for the chat, such as “You are a blunt editor. Keep answers under 100 words.” It reliably changes:
- Tone and personality.
- Format, like bullet points, tables or length.
- The role the model plays: tutor, recipe assistant, code reviewer.
- How much hedging you get on everyday topics, within limits.
It doesn’t rewrite the model’s training. A model tuned to decline something will usually keep declining it, and a small model given contradictory instructions tends to get worse at everything. The app lets you set a system prompt for each chat and a default one for every new chat, plus Creative, Balanced, Precise and Thinking presets that adjust how varied or focused the replies are.
Does anyone see what you type into an offline chatbot? #
Not with a true on-device app. Personal LLM runs the model on your phone’s own chip. It has no server and no account, and it doesn’t run analytics on your conversations. It uses the network for two things only: the one-time model download from Hugging Face, and ads in the free version, which come from Google AdMob and don’t include your chats. Once the model is downloaded you can switch on airplane mode and keep chatting.
That privacy is a big part of why people use offline AI for health questions, personal writing and work drafts. For how different assistants compare, see the most private AI assistant. If you’re weighing whether to paste something sensitive into a cloud chatbot instead, read is it safe to paste emails into ChatGPT.
Do the usual rules still apply offline? #
Yes. Running a model privately doesn’t change what’s legal where you live, and model licenses carry use restrictions whether anyone checks or not. Apps also carry age ratings; this one is rated 12+ on the App Store because an open model can discuss mature subjects.
There’s a practical point too. Fewer guardrails don’t mean better answers. An offline model states wrong facts with the same confidence as right ones, and no second system is checking its work. For medical, legal or financial decisions, treat its answer as a starting point and confirm it with a professional or an official source.
Frequently asked questions #
Are local LLMs censored? #
They carry whatever safety training their lab gave them, so they decline some requests. What they don’t have is the extra screening, account enforcement and human review that cloud services add. How often a model refuses varies a lot from one model to another.
Does Personal LLM filter or read my chats? #
The app doesn’t send your chats anywhere, so there’s nothing for a company to check or read. What a model will and won’t answer comes from the model you chose and the system prompt you set.
Why does my offline AI refuse harmless questions? #
Small models sometimes misjudge a harmless question as risky because it looks like something they were trained to decline. Rephrasing with more context, adding a system prompt that explains your purpose, or switching to a larger model usually helps.
Can I remove the filters from a local model? #
Community-modified versions of open models exist, but they’re unvetted, often lower in quality, and the original license’s use rules still apply to them. We don’t recommend them.
Will a cloud chatbot ban me for asking the wrong question? #
Usage policies let providers warn, restrict or ban accounts that break their rules, and the details differ by provider. An offline chatbot has no account to restrict.