Guides

How to run a free AI detector on your own computer

What you need

Python, the transformers and torch libraries, and about 1.7 GB of disk space for the Desklib model, which downloads automatically the first time you run it.

pip install torch transformers sentencepiece

The code

Desklib’s model has a custom output layer, so it needs a small class. This is the code we used in our benchmark, based on the example in Desklib’s model card.

import sys
import torch
import torch.nn as nn
from transformers import AutoConfig, AutoModel, AutoTokenizer, PreTrainedModel

MODEL = "desklib/ai-text-detector-v1.01"

class DesklibModel(PreTrainedModel):
    config_class = AutoConfig

    def __init__(self, config):
        super().__init__(config)
        self.model = AutoModel.from_config(config)
        self.classifier = nn.Linear(config.hidden_size, 1)
        self.post_init()

    def forward(self, input_ids, attention_mask):
        hidden = self.model(input_ids, attention_mask=attention_mask)[0]
        mask = attention_mask.unsqueeze(-1).float()
        pooled = (hidden * mask).sum(1) / mask.sum(1).clamp(min=1e-9)
        return self.classifier(pooled)

device = "mps" if torch.backends.mps.is_available() else "cuda" if torch.cuda.is_available() else "cpu"
tokenizer = AutoTokenizer.from_pretrained(MODEL)
model = DesklibModel.from_pretrained(MODEL).to(device).eval()

text = open(sys.argv[1]).read()
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=768).to(device)
with torch.no_grad():
    probability = torch.sigmoid(model(inputs["input_ids"], inputs["attention_mask"])).item()
print(f"AI probability: {probability:.3f} ({'flagged as AI' if probability >= 0.5 else 'not flagged'})")

Save it as detect.py, put the text in a file, and run python detect.py essay.txt.

How to read the result

Desklib flags a text when the probability is 0.5 or higher. In our test that threshold flagged 6.3% of human texts and caught 58.3% of AI texts. Desklib reads at most 768 tokens, roughly 550 words, so for longer texts score sections separately.

A score near 0.5 means little either way. A high score on news writing deserves extra doubt, because Desklib flagged 20.8% of human news reports. A low score on text that may have been edited means little too: after a humanizer rewrite, Desklib caught 32.5% of Claude Opus 5.5 texts.

To compare detectors on your own texts, download our scores and texts and check how each one does on the kind of writing you care about.