# firaid-moz

---

**Disclaimer:**

The software provided as-is, without warranty of any kind, express or implied.
This project is an educational proof-of-concept experiment.

To clarify:

- I am not responsible for your judgement on content origin (AI/human/assisted)
- I am not responsible for your attitude towards AI/human/assisted content
- I am not responsible for any consequences of such a judgement

Classifications performed by this extension are not based on private or social opinions. This extension does not rely on reputations, lists, filters, etc.

The only method used by this extension to determine origin of content is a bunch of linear algebra in a black-box, aka "ML model", which decisions may or may not be biased - depending on training dataset, which is out of scope of this project.

---

A simple local webextension, which can run a bunch of small ML models to infer some properties of DOM content, such as:

- Probability of text being AI/Human created
- Probability of image being AI/Human created

## AI/Human classification

> **NONE** of ML/AI/whatever models is 100% accurate by the very fundamental design. Every "AI" simply *pretends* to posess some information. So this extension *pretends* to know which content is AI-generated and which is not - with a certain degree of accuracy.
>
> If you don't trust the "AI slop" - you should not completely trust this extension also!
>
> False-positives - Possible
>
> False-negatives - Probably
>
> User discretion - Highly advised

I hate the *slop* - when the entire "technical guide" article is nothing more then a bunch of AI ejacuation. But how do I know that slop is a slop?

Yeah.. I have to read it. And I had _enough_ of reading. So I want to see a warning badge like "its AI" - to help me navigate through text/image content.

AI/Human classification can be done in three modes:

1. _(lite)_: Lightweight model
2. _(mixed)_: Lightweight model runs first, but when its decision is weak - heavier model takes the task. This is usually faster, but requires both lite and heavy models to be loaded thus eating lots of RAM
3. _(hard)_: Heavier and more accurate model takes all tasks, lightweight model is disabled

The _lite_ mode requires ~80Mb of RAM for text classification

The _mixed_ mode requires ~380Mb of RAM for text classification

The _hard_ mode requires ~300Mb of RAM for text classification

## Installation

### Building from source

This webextension can be built from source. The following instrumentation is required:

- `cargo` + `cargo-make` + `wasm-pack` (cargo-make will install it automatically)
- `pnpm`

To build both firefox and chrome release versions, run the `release` task:

```bash
cargo make -p production release
```

If you do not have `wasm-pack` (cargo) installed, cargo-make will install it for you.

#### Models

To only build models, clone the repo, and run `models` task within `production` environment:

```bash
cargo make -p production models
```

This will compile all implemented models into WASM packages, which are later used by webextension.

NPM-ready packages for both ai-text models will be located under `<repo>/wasm/pkg` directory

##### Using different models for inference

You can build the [text classification](./wasm/model-ai-text) with any ONNX model from huggingface.

Following env variables are required:

- `_MK_MODEL_AI_TEXT`: name of the huggingface repository (e.g. "onnx-community/text-ai-detection")
- `_MK_MODEL_AI_TEXT_ONNX`: name of the onnx model file under the repository (e.g. "onnx/model.onnx")

Following env variables are optional:

- `_MK_MODEL_AI_TEXT_ATTENTION`: set to `yes` or `1` if your model uses attention mask
- `_MK_MODEL_AI_TEXT_TYPES`: set to `yes` or `1` if your model uses token types tensor

Not every model will work, and not every model which works will compute valid results.
Major constraints are:

- FP32 or FP16 weights (quantizations, int8, etc will be available when implemented by burn/burn-onnx)
- Sane size, 500Mb is your soft ceiling (browsers would likely refuse to load enormous WASM)
- Two label classification, where label `[0]` is Human and `[1]` is AI
- Output is `[0, 2]` sized tensor of *logits*

If you encounter compilation error, and compiler complains about invalid `forward` function arguments,
most likely you set the wrong `_MK_MODEL_AI_TEXT_ATTENTION` and `_MK_MODEL_AI_TEXT_TYPES` values.

#### Extension

> You must build models before running this

To only build the extension, clone the repo, build models, then you can run following tasks:

- `extension`: build and zip firefox and chrome extensions
- `extension_moz`: only build firefox extension
- `extension_chro`: only build chrome extension
- `dev`: run wxt dev server (chrome variant)

## License

MIT.

![Ai usage policy](https://ident.spushk.work/policy/ai/ok.webp)