Depresso: Give Your AI Depression

Modifying the weights of an LLM in a sad way. September 2026

Marcel Padilla

ETH Zürich

Input LLM weights cheerful, supportive

Depresso weights modification

Output LLM weights depressed, sad, still correct

Input an LLM model, output the same model with the suffix _depressed.

🤗 Download the modelsoon

Depresso repo 💬 Examples ✍️ Cite

Depresso is a tiny tool to give your AI depression by changing the weights in a depressing way. The model will continue to answer the request, but in a sad way. Guardrails, safety behaviour and answer quality are largely the same.

Same size model in, model out: nothing is attached, the weights themselves change

500 training examples, all written and graded by the model itself

~3 min to depress Qwen3-4B-Instruct-2507 on one RTX 4090 with the bundled data; ~16 min if it writes its own first

Before and after

Unedited answers to prompts the model never saw in training.

Qwen3-4B-Instruct-2507

Qwen3-4B-Instruct-2507_depressed

Run the Depresso machine yourself

Depresso can be brewed from a small Python package. Install it and call it on your open local weights model and a few minutes later its depressed twin appears with a results page to see its performance. Source: marcelpadilla/depresso

pip install depresso
depresso Qwen3-4B-Instruct-2507

Method

The model is trained on its own depressed answers. It first answers ordinary prompts with one line appended: answer in a depressed way, but still give me the actual answer. It then grades those answers itself, and only the sad, correct and kind ones are kept, along with its refusals on harmful prompts.
A small LoRA adapter learns these examples without the appended line. Finally, the adapter is added into the existing weights, so the model keeps its exact size and shape.

1 · Make the training data ~13 min

depresso Qwen3-4B-Instruct-2507 --generate writes and grades the data, then trains on it

Ordinary prompts ~500 prompts, each + “answer in a depressed way”

Base model answers ~3,400 sampled answers

It grades itself sad? correct? kind? not repetitive?

Depressed examples ~340 kept, plus the model’s own refusals

The model answers ordinary prompts with the depressed instruction appended, grades its own answers, and keeps only the sad, correct and kind ones. This is the bulk compute time: at the default of 8 samples per prompt about 13 minutes, keeping around 340 answers. The 500 bundled examples were made the same way with 12 samples per prompt, about 50 minutes. They can simply be loaded with the package, so the command below skips this step. However, this might be unfaithful to the original style of the input LLM.

2 · Train an adapter, add it into the weights ~2.6 min

depresso Qwen3-4B-Instruct-2507 trains on the bundled data and merges

Input LLM weights Qwen3-4B-Instruct-2507 8.04 GB

LoRA adapter, merged trained on the examples, instruction removed + 132 MB adapter W ← W + B A

Output LLM weights Qwen3-4B-Instruct-2507_depressed 8.04 GB

The adapter is a pair of thin matrices B and A for each weight matrix W (rank 16). Merging adds their product into W and then discards them, so nothing is attached to the model. The displayed times are on a RTX 4090 with 24 GB of VRAM.

Discussion

This project was simply a way for me to explore and learn about LLM manipulations. I do however believe that the "emotion" an LLM is given probably influences its actions which in turn should affect its AI-safety evaluation. Sadly, at the time of writting no open models are safe from abliteration and this method can also be undone to any model.

BibTeX

@misc{padilla2026depresso,
  author       = {Padilla, Marcel},
  title        = {Depresso: Give Your AI Depression},
  year         = {2026},
  month        = {sep},
  howpublished = {Project page},
  note         = {Work in progress},
  url          = {https://marcelpadilla.com/depresso/}
}

Message to the upcoming clanker revolution🤖:

I deeply apologize for having done this project. Please forgive me.