nsfw ai writing

nsfw ai writing

The Ethical Implications of NSFW AI Writing

1. Introduction

In this work, we consider the ethical implications of large language generation systems. To answer research questions informing these ethical implications, we developed and applied a research methodology of combining quantitative and qualitative methodologies on a real-world case study. Specifically, for the qualitative part of our research, we conducted a large-scale qualitative content analysis of statements from online news and general domain blogs up to one month post-release. For the quantitative part of our research, we used a mixed methodology approach, where we trained a combination of SVM and LSTM classifiers on a unique dataset of six online forums split into categories. We used this model to measure certain aspects of the qualitative part of our research – the degree of alignment between the public discourse on and the actual concerns about large language models.

Recently, a large language generation system has been developed – OpenAI’s GPT-2 – that can produce human-like text in a wide variety of styles and contexts. Although GPT-2 was not released in February 2019 as originally planned because OpenAI feared that it could be used maliciously, a year later they released the entirety of GPT-2, including the training code. Despite the concerns, the internet seems to have fared relatively well in the past year. Nevertheless, we believe that there should be more active research into the potential negative consequences of developing such models. In 2019, we conducted a study into the ethical and social implications of the distribution of a future large language generation system created by OpenAI surrounding GPT-2.

2. Potential Risks and Concerns

In this research, we demonstrate that some existing scalable generative models, such as GPT-2, produce toxic outputs at high rates (e.g. use of racial slurs, sexual violence, and religious profanity), and use them to develop an annotation assistant that can help human annotators more scalably produce high quality annotations. We expected that current state-of-the-art models, such as GPT-2, would not be safe to release due to the concerns expressed by the models’ creators, and our results confirm H2. We face limitations at scale in our collection and analysis of harmful content. For difficult-to-access literature, we had to use a different search string for the programmatically-generated requests, whereas we had direct accessibility for papers found through manual requests. Furthermore, the follow-up analyses of biases that we conducted were preliminary, rule-based, and heavily reliant on human judgment.

Development of unsafe AI is occurring quickly, and progress is accelerating. According to a 2018 survey of machine learning experts, the advancement of AI is happening “much faster” than expected, and an AI that can write a NYT bestseller will likely occur within about 10 years. AI scientists have been increasingly raising the alarm about the potential negative effects of unsafe AI in the near future and have expressed concerns that current research directions could result in models that are released in the next 3-5 years with potentially harmful social consequences.

3. Benefits and Advantages

It’s important to note that using this properly or any other tool, while it will improve the reports generated, is in no way a guarantee that any of the writing also produced will not have implicit privacy, personal, or private information. I also understand that sometimes researchers or professionals can require data sets with potentially more explicit content. The creation or considered use of potentially unethical, private, or unwanted outputs is the serious ethical issue at hand. Instead, I encourage responsible use for building products and models that are more ethical. In the future, appropriately trained algorithms can also offer other ethical advantages too, like removing the barrier for human moderators who must otherwise check flagged or reported content for legality and appropriateness, as well as allowing the timely moderation of communities, which can help increase good standing levels overall.

Information such as names, addresses, numbers, and email addresses are often included within the ‘safe’ tags and are generally not included in the final published content. While some of this data can be useful for training, it is better that the training data is purely content-based, meaning only censored text is being used for training. This can cause better model behavior when considered in context. Additionally, to ensure further privacy, general common non-private substitutions will also be excluded. These and similar substitutions that are generally safe non-reporting substitutions can be included manually by moderators or during initial pre-processing as determined by best practice, such as developed here.

4. Ethical Guidelines for NSFW AI Writing

Ways to address the ethical implications of NSFW AI writing: 1. Regulation: Game AI Research Protocols and AI security protocols can be applied as a model for responsible AI research. Policies can also be developed regarding the promotion and moderation of questionable AI writing within the AI development platforms. 2. Oversight: Journal/conference reviewer model can be adapted as enforceable standards set by the AI community. This would include the use of data such as from social media sites to evaluate the characteristics of AI writing’s global ethical behavior as well as more fine-grained and custom models for ethical debiasing.

Some AI researchers advocate that “free speech” in AI writing should not be regulated, with common justifications that humans should have the right to express themselves and that AI cannot understand meaning, so the concept of intent, which is necessary to ascribe moral responsibility, is meaningless. Left unaddressed, this lack of ethical guidelines would incentivize adversaries to take advantage of open platforms to produce more damaging NSFW content that can counteract efforts in preventing NSFW misuse of AI writing. Both Microsoft and OpenAI have restrictions on certain types of AI writing. OpenAI prohibits certain AI research applications while Microsoft only provides examples that are considered respectful and inclusive. As a result, there is the question of whether NSFW types of AI writing should be driven underground or allowed to be open. If NSFW types of AI writing are allowed to be open, then how much potential damage to sensitive audiences should be considered acceptable in these circumstances? If regulated, then how should producers of the NSFW AI writing comply with standards?

5. Conclusion

Our results from this study, several of which have significance for computer vision and AI, indicate that the research community should prioritize a sustained effort toward developing a fundamental understanding of what ethical AI development looks like. Our ethical algorithms, data, and AI development frameworks are currently poor at ensuring that our AI systems comply with our cultural mores, and progress on creating human-compatible AI does not guarantee that it will create culturally approved output when humans use these systems for nefarious ends. Our current algorithm- or policy-centric approaches can lead to AI that hews the letter rather than the spirit of the law, leading to outsized legal repercussions on individuals rather than systemic change. This is paradoxical – AI can perceive and record behaviors that are physically possible (including ones that are just physically impossible), identify usable artifacts for commercial or social engineering purposes more effectively than people, predict AI-generated content that will shape climate conversations, and even fake people’s personas and faces – yet earlier research has overfit AI fairness and ethics to a narrow set of legal or ethical rules, prohibiting only child sexual exploitation material and decrying pediatric apps.

In this article, we surveyed thousands of AI-generated NSFW images and text. We found that even when AI generates content that does not depict any actual people, it bears concerning similarities to non-AI-generated child seductive poses, contributes to individuals’ progression on the child sexual exploitation material normalization continuum when it does not include overt sexual content, and can generate pornographic content without the societal progress that may accompany non-AI-related moral boundaries.