Gemma 4 Ethics & Safety
Google filtered Gemma 4's training data and tested the finished models for harmful output before release. But an open model you download arrives without the safety layers a hosted service wraps around it. This page explains what Google did, what is left to you, and how EmbeddingGemma 2 differs.
What Google did
From the Gemma 4 model card: how the training data was cleaned and how the models were tested.
Child-safety filtering
Filtering for child sexual abuse material was applied at multiple stages of data preparation.
Sensitive data filtering
Automated methods removed certain personal information and other sensitive data from the training sets.
Quality and safety filtering
Further filtering based on content quality and safety, in line with Google's policies.
What was tested
Automated and human evaluations, run with Google's safety teams, aimed at preventing child exploitation material, dangerous content, sexually explicit content, hate speech and harassment. Testing covered text and images, with safety filters switched off.
What Google reports
"Gemma 4 models significantly outperform Gemma 3 and 3n models in improving safety, while keeping unjustified refusals low." The card describes minimal policy violations across all sizes. It gives no detailed figures.
From Google's Gemma 4 model card. These are the maker's own evaluations; this site has not tested the models' safety behaviour.
The risks the model card names
Bias and fairness
Models trained on large amounts of real-world material can reflect the social and cultural biases in it.
Misinformation
The model can state things that are wrong or out of date, and could be misused to produce misleading text. Its knowledge stops at January 2025.
Privacy
Training data was filtered for personal information. Filtering lowers the risk; it is not a guarantee.
Harmful content and misuse
Google's guidance is for developers to add content safety safeguards suited to their own product.
Paraphrased from the "Usage and limitations" section of the model card. What the models get wrong in ordinary use is on the limitations page.
What open weights change
The same model behaves the same way wherever it runs. What changes is who is responsible for everything around it.
With a hosted service
- The provider adds filters on what goes in and what comes out.
- The provider watches for abuse and can cut off bad actors.
- Updates arrive automatically.
- Your prompts leave your machine and are handled under the provider's terms.
With open weights you run yourself
- No filters unless you add them. The model's own training is the only safeguard.
- Monitoring is yours to build, or to go without.
- You choose when to update, and must remember to.
- Fine-tuning can change safety behaviour, so test again afterwards.
Where open weights are the better choice
Prompts never leave the device
Run locally, nothing is sent to anyone. For sensitive material that can be the safest option there is. See run locally.
Open to inspection
Researchers can examine and test the actual model, and results can be reproduced because the weights do not change underneath you.
Access for everyone
No account, approval or payment is needed, which matters for education, research and places with poor connectivity.
This section is this site's view. Hosted options are on the API page; running a service yourself is under self-hosting.
The licence, and Google's use policies
| Document | What it is | Link |
|---|---|---|
| Apache 2.0 licence | Gemma 4's licence. Allows commercial use, changes and redistribution. Earlier Gemma versions used Google's own Gemma Terms of Use instead. | Gemma 4 licence |
| Prohibited Use Policy | Google's list of uses it does not permit for Gemma models, linked from the Gemma 4 documentation. | Prohibited use |
| Intended Use Statement | Google's description of what the models are meant for. | Intended use |
What the Prohibited Use Policy covers
Violating others' rights
Content that infringes or misappropriates a person's or organisation's rights.
Dangerous, illegal or malicious activity
Including child exploitation, crime, violent extremism, spam, fraud, phishing, malware and getting around safety filters.
Harm to people or groups
Hatred, harassment, violence, self-harm, doxxing, surveillance and discrimination.
Misleading content
Impersonation, false claims of expertise, defamation, and automated decisions in high-stakes areas.
Sexually explicit content
For pornography or gratification. Scientific, educational, documentary and artistic uses are excepted.
Not legal advice
How the policy and the Apache 2.0 licence fit together for your use is a question for a lawyer, not this page.
Summarised from Google's Gemma Prohibited Use Policy; read the original for the exact wording. Licence differences between versions are on the Gemma 3 page and in the model comparison.
Design, align, evaluate, protect
The four stages of Google's Responsible Generative AI Toolkit, with what each means in practice.
Design
Decide what your application should and should not do. Write it down as a safety policy before you build.
Align
Shape the model's behaviour to that policy, through prompts, examples or tuning, and look into why it responds as it does.
Evaluate
Test against your policy and compare versions side by side, so you know whether a change helped.
Protect
Add checks around the model: classify what comes in and what goes out, and label generated content.
Stage names from Google's Responsible Generative AI Toolkit; the one-line explanations are this site's summary. Google also points to its Secure AI Framework for security.
Safety tools you can add
| Tool | What it does | Details |
|---|---|---|
| ShieldGemma | Checks text going in or coming out against a safety policy and answers yes or no | 2B, 9B and 27B sizes, built on Gemma 2. Covers hate speech, harassment, dangerous content and sexually explicit information. |
| ShieldGemma 2 | Checks images against a safety policy | 4B, built on Gemma 3. Covers sexually explicit, dangerous and violent content. |
| SynthID Text | Watermarks generated text so it can be identified later | Part of Google's toolkit |
| Agile classifiers | A method for building your own safety classifier from a small number of examples | Part of Google's toolkit |
| LLM Comparator | Runs and visualises side-by-side evaluation results | Part of Google's toolkit |
| Learning Interpretability Tool (LIT) | Helps investigate and debug how the model responds to prompts | Part of Google's toolkit |
From the ShieldGemma and ShieldGemma 2 model cards and the Responsible Generative AI Toolkit. Neither ShieldGemma version is built on Gemma 4.
EmbeddingGemma 2: ethics and safety
EmbeddingGemma 2 does not write text. It turns text, images, audio and video into numbers used for search and sorting. That changes the safety picture: there is no harmful answer to block, but the model can still surface the wrong thing, or rank people unfairly.
No safety tuning
Google's card says the model has no post-training alignment or safety tuning. Safety work was focused on filtering the training data.
Same data filtering
Child-safety filtering at multiple stages, removal of certain personal and sensitive data, and quality and safety filtering.
Licence and policy
Apache 2.0. The model card also says deployments must follow the Gemma Prohibited Use Policy.
Risks Google names
- Bias. The vectors can reflect social and cultural biases in the training material, which can skew search results or classifications.
- Misuse. Embeddings can be used to retrieve, classify or organise content in false, misleading or harmful ways.
- Privacy. Developers are told to follow privacy regulations and use privacy-preserving techniques.
- Uneven languages. Performance varies across the 100+ languages it supports.
What that leaves to you
- Filter results. Google names retrieval filtering as an application-level safeguard that is your responsibility.
- Test for fairness, especially if results affect people: hiring, lending, moderation.
- Control what you index. A search can only surface what you put in it. Keep private material out, or behind access rules.
- Treat stored vectors as sensitive. They are derived from your documents, so protect them as you would the originals.
The left column and the first two points on the right are from Google's EmbeddingGemma 2 model card; the last two are this site's guidance. Running it privately is covered in how to use and the prompt guide.
Deployment checklist
A minimum bar for putting Gemma 4 in front of other people. Tick items off as you go; nothing is saved or sent.
Before launch
After launch
The checklist is this site's guidance, drawn from Google's toolkit stages. It is a starting point, not a compliance standard. Personal use on your own machine needs far less; see chat and the playground.
Is Gemma 4 safe to use?
Google filtered its training data and reports that it outperforms Gemma 3 on safety in testing. No model is safe in every situation, and a downloaded model has no extra filters around it unless you add them.
Does Gemma 4 have built-in content filters?
The model was trained to avoid harmful output, and Google tested it with filters off. There is no separate filter in the download. Hosted services may add their own.
Is it uncensored?
No. It is trained to decline harmful requests. Because the weights are open, others can modify it; a modified version is not the model Google tested.
What does Apache 2.0 mean for safety rules?
The licence itself sets no limits on use. Google publishes a Prohibited Use Policy and an Intended Use Statement alongside it. For how these apply to you, take legal advice.
Can I use ShieldGemma with Gemma 4?
Yes, as a separate checking step. It is a different model with its own terms, built on earlier Gemma versions.
Is running it locally more private?
Yes. Prompts and files stay on your machine. You remain responsible for how you use the output.
Does fine-tuning affect safety?
It can. Tuning changes the model's behaviour, so re-run your safety tests afterwards.
Is EmbeddingGemma 2 safe?
It cannot generate harmful text, but it has no safety tuning. What your search returns, and whether it treats people fairly, is for you to check.
Know the limits before you build
Safety starts with knowing what the model can and cannot do.