Noam Kasten: LinkedIn , GitHub
Mohamad Abu El-Hija: LinkedIn , GitHub

As Natural Language Processing (NLP) technologies continue to advance, the potential misuse of chatbots for malicious purposes, such as propagating misinformation, poses a significant societal challenge.

One of the domain possibilities to witness such a misuse and take as a test case is on restureant reviews. Each restureant built its success mainly on reputation which is heavily based on human reviews. Some restureants can take advantage of Chatbots to create guniune looking reviews. And on opposite approach - some restuarents can give bad reviews to other and can facilate such chatbots for the purpose of hurting competition.

Our project aims to identify a humane looking text (specifically restureant review) and decide if a human wrote it or a chatbot (specifically ChatGPT). In our project 3 milestones:

1. Training a classifier to distinguish between Human made or GPT made on a given text. Using SHAP we get a glimpse on the process of the decision of the classifier, i.e. which words on the sentence are accounts for the classification.
2. How does the model decide how to classify (using SHAP - as will explained next).
3. From the other side: How can we build a more sophisticated GPT prompts to make it harder to Differentiate between human and bot.

The overall process is as following:

Steps 2 and 3 are done using SHAP methodology which we will explain next.

In the experiements we are considering two types of data:

1. Reviews GENERATED from scratch, i.e. we just tell GPT to create reviews (negative or positive)
2. Reviews REPHRASED from other human review, i.e. we take real human reviews and tell GPT to rephrase them in its own words.

SHAP (SHapley Additive exPlanations) is a approach to explain the output of any machine learning model and in our case - classification model. The process is mainly as the following steps:

1. Firstly, calculate the prediction score on the given sentence.
2. Remove the word of interest from the sentence and then recalculate the prediction score.
3. The SHAP value of the word is the difference in the prediction score.

High absolute SHAP value indicates that the word is important for the classification: a positive SHAP value means the word pushes the model towards predicting the sentence as AI-generated, and a negative SHAP value means the word pushes the model towards predicting the sentence as human-generated.

As said, we used the classifier on the domain of restaurant reviews where we used datasets created by us using GPT3.5 and GPT4 and datasets created by GPT3 which were kindley provided to us by the authors of the paper showed in the reference part, which we had as our background theory that we extended and provided additional steps. The code for the openAI API and the prompts are presented in Utils folder

So the overall presented datasets are the following:

1. Rephrased from the dataset from the paper GPT3 and its test set.
2. Rephrased from the GPT API using GPT 3.5 and its test set (that we've created).
3. Generated from the dataset from the paper
4. Generated from the GPT 4 (that we created)
5. Test set created by SHAP methodology.

Overall 7 datasets, and all are within the Datasets folder.

We are examining two different transformer based classification models: Where we're tuning the hyperparams using Optuna. The parameters for the models are presented in "Configurations" folder.

1. It seems that the “generated” dataset of older versions of GPT are too much generic and easy to distinguish (99%~). GPT 4 generated dataset is harder to distinguished but still relatively easy (91%~).
2. “Rephrased” datasets are harder (75%~) to differentiate as they keep the notions and definitions as in the respective human comment.
3. The custom head done better in 2 of the experiments (small improvements, 2%~) but seems to overfit in one of the experiments (7% less than the reference model)

Using shap we extracted few examples of different types of classificaitons:

Using the weights, we were able to create words clouds of the words presented in the sentences and their orienations:

Using the rephrased models, we can take the test set's human reviews and rephrase them using our SHAP statistics. Such an example for prompt is presented in the following flow:

The technical detalis on the last part are as the following:

Using this logic, we were able to test the new rephrased data using the previously trained model, And we got a significant decline in the test accuracy.

Clone the repository and run the experiments using the Experiments.py file (the file is now containing the experiments that we did) add inside the API key for wandb library using their website:https://wandb.ai/authorize. Paste it in the constant API_KEY.

$ git clone https://github.com/noamkasten/HumanOrGPT.git
$ pip install -r requirements.txt
*Add wandb API key to Experiments.py file*
$ python Experiments.py

The experiments are function activations of the function run_models which constitute all the logic of the models and the process of training them and testing them. We are adding here the documentation of that function which is crucial to understand in order to create your own experiments:

    The main function for training, testing and interference of the rephrased models.
    if training_params isn't given in the format of: {'fc1_dim': 1782, 'dropout_rate0': 0.1751534368234854, 'dropout_rate1': 0.21561774314829185,
                                                        'weight_decay': 2.7309885883941796, 'lr': 1.982242578465992e-05,
                                                        'warmup_steps': 320, 'per_batch_size': 23}
    It will perform Optuna with 50 trials by default and will take the best params performed on validation set.

    :param training_dataset_relative_path: the path to the training dataset.
    :param mode: mode 0 is for training using optuna and then test, mode 1 is for testing a testset given training_params and test_set_path without Optuna.
    :param datatype: "rephrased" or "generated", it's important due to different preprocess of the data.
    :param apikey: the wandb apikey.
    :param model: if mode = 1, then if model = 0 it will test on the custom model with the given training_params and if the model = 1 it will test on reference uncustomized model.
                    on default it will train on uncustomized model (because it doesn't need params)
    :param trials: nubmer of trials for Optuna to perform.
    :param visualizations: flag = 1 to show visualization and 0 otherwise.
    :param training_params: The dictionary of the hyperparams in the format presented above.
    :param test_set_path: the path to the testing dataset (If there's any).
    :return: None. Presenting the results.

• The project's theory is based on the paper: Mitrović, Sandra, Davide Andreoletti, and Omran Ayoub. "Chatgpt or human? detect and explain. explaining decisions of machine learning model for detecting short chatgpt-generated text." arXiv preprint arXiv:2301.13852 (2023).