AI ETHICS BOT

Type: Web Application
Year: 2019
Platform: Web Browsers
Development Environment: Python, JavaScript

We used to create ethics for AI, and now it's time for AI to create ethics for itself. Just as many AI companies like Google and Microsoft who had come up with principles for AI, increasingly focus have been laid on regulating artificial intelligence, since it now indeed affects our lives in various aspects. However, few people have ever tried to generate ethics by AI itself. It may sound unreliable to let machines rule themselves automatedly, whereas considering the fact that most of the individuals in this world manage his/own behaviors with his/her own intention, automation is already everywhere. We take credit on people who are adults regulating themselves, while few of us believe that children are supposed to be responsible to their own lives without the monitor of their parents, as we know that children generally do not own comprehensive and mature world views. Here, contexts matter. People generate their world views and outlooks based on the culture and education they have received, which helped them establish a context about morality and ethics, so does AI.

Idea and Inspirations

The idea of this project is to take a peek on the possibility that AI can generate ethic rules. It trains the computer using a batch of articles in relation to Information Ethics, and then let the computer create new ethic sentences. It is meaningful in two aspects: 1. It can be a practical reference for other digital humanities scholars to research on machine learning, as there aren't many precedent works in this field; 2. It let people foresee a future that if AI is able to regulate itself, what kind of rules there will actually be.

One of the inspirations is the Moral Machine [1] produced by MIT, which collects choices of ethical problems. It is like an online questionnaire which illustrates a context of moral issues based on different moral scenarios. The statistic report of the Moral Machine can be taken as the references for building moral machines, however, it does not directly connect to AI – the contexts were built by humans, as well as the principles to rule moral machines. In the project AI ETHIC BOT, the idea will be moved further to grant the machine itself generating ethic rules, during whose procedure manual works will only be found in building contexts.

This project relies on machine learning with neural networks. Retrospectively, numerous attempts have been made to generate texts using neural networks, some of which are art projects - like Virtual Muse [2] and The Center for Midnight [3], which tried to train the computer to generate literal texts, or just for fun - like Halloween Spooky Costumes Generator [4]. Though none of them are related to ethics, their methodologies are of strong value to this project.

Methodologies

Inspired by above projects, AI ETHICS BOT mainly uses a neural network framework named char-rnn, which is an implementation of multi-layer Recurrent Neural Network (RNN, LSTM, and GRU) for training/sampling from character-level language models by MIT scholar Andrej Karpathy [5]. On top of char-rnn, Andrej's MIT colleague Max Woolf had implemented a packaged library – Textgenrnn [6] based on Python language, which makes the training of char-rnn easier.

The general process of a textgen-rnn training in this project can be as follows:

1. Gather a bunch of articles in terms of Information Ethics as the training contexts;

2. Use Python to extract the top keywords of all the documents.

3. Manually pick the most relevant keywords about Information Ethics.

4. Use a synonymous online dictionary [7] to extend the keywords list, as some of the lower frequency words may be neglected during the manual pick up.

5. Extract sentences from the documents if a sentence contains at least two ethic keywords.

6. Use the sentences as the training material to train char-rnn neural network (trained by 50 epochs).

7. Generate new sentences based on the inputted prefix.

The above phases (except manual review part) are implemented using Python. Once phase 6 is finished, it will generate a model file for phase 7 to make use of, so that it doesn't have to repeat the training phase every time when generating the texts.

​

Conclusion

So far, AI ETHICS BOT is just an attempt to show how good computers can generate ethic sentences. These sentences cannot be directly regarded as the Code of Conduct or Rules for regulating artificial intelligence. However, just like the Moral Machine doesn’t actually create a moral machine, the most important thing is that AI ETHIC BOT can also be an inspiration to other DH scholars who are interested in information ethics and artificial intelligence. It can be regarded as a simple game, just like how people play with AI chatbots, or it can be a tool for DH studies. Technically, the methodologies of this project can be reused and extended to other AI applications; Theoretically, if someone runs out of ideas to create ethic principles, he/she can turn to this project to acquire some creative/unexpected responses, and in the meanwhile, we are able to take a peek of how smart and how moral our computers are to create rules, which helps us to foresee if AI phobia is just a over-concern, or it will lead us to a world like the movies Matrix and the Terminator have shown.

​

The webpage is developed using Django [8], and has been presented as the poster session at CSDH Congress 2019 (https://www.congress2019.ca/associations/255).

​

References

​

[1] MIT. Moral Machine. moralmachine.mit.edu/.

[2] Prose and the Virtual Muse, Charles O. Hartman, oak.conncoll.edu/cohar/Programs.htm.

[3] The Center for Midnight, Robin Sloan, www.robinsloan.com/center-for-midnight/.

[4] Let Our Algorithm Choose Your Halloween Costume, Janelle Shane, www.nytimes.com/interactive/2018/10/26/opinion/halloween-spooky-costumes-machine-learning-generator.html?action=click&module=Opinion&pgtype=Homepage.

[5] Karpathy, Andrej. Multi-Layer Recurrent Neural Networks (LSTM, GRU, RNN) for Character-Level Language Models in Torch, MIT, github.com/karpathy/char-rnn.

[6] Woolf, Max. Easily Train Your Own Text-Generating Neural Network of Any Size and Complexity on Any Text Dataset with a Few Lines of Code, MIT, github.com/minimaxir/textgenrnn.

[7] Thesaurus.com, Roget's 21st Century Thesaurus, www.thesaurus.com/browse/synonymous.

[8] Django Makes It Easier to Build Better Web Apps More Quickly and with Less Code., Django Software Foundation, www.djangoproject.com/.

​