No More Dangerous Visions: AI Guardrails and Humanities Work

 

After eight years, this blog creaks, groans, stretches, and returns to life not about virtual worlds, but the newest technology on the block: Generative AI.

 Today my students put essays to two different Large Language Models (LLMs) hosted on our campus. All of them first used Claude Sonnet 3.5 and then the groups divided, with half putting the essays to ChatGPT 4 and the others Google Gemini Pro.  By the time you read this, the versions and even names may have changed.

In any case, I asked the students to provide the assignment prompt from a partner's paper, as well as the text of the paper itself. Then the students were to compare how the LLMs' advice compares to that of a human partner as well as Grammarly. 

For three students with Humanities topics, Anthropic's Claude refused, like Barletby the Scrivener in Melville's story, to do any work. It locked the students out of the software and logged them off.

Why? Subject matter and explicit language quite appropriate when quoting from some literary works or discussion topics such as human trafficking or sexual violence. So then we quizzed Claude about why. I asked it a general question about which words or phrases it flags as "unsafe" and got this reply:

As an AI language model, I don't actually have a system for identifying "unsafe content" or refusing to answer based on specific words or phrases. I'm designed to engage in a wide range of conversations while maintaining ethical boundaries. However, I do have guidelines that help me respond appropriately to different types of queries: 

1. I don't engage in or promote illegal activities. 
2. I avoid producing explicit sexual content or pornography. 
3. I don't assist in creating harmful content like hate speech or extreme political propaganda. 
4. I don't share personal information about real individuals. 
5. I avoid giving medical, legal, or financial advice that should come from professionals. 

To avoid any issues, it's generally best to: 

1. Keep queries legal and ethical. 
2. Avoid requests for explicit sexual content. 
3. Don't ask for help with harmful or dangerous activities.
4. Stick to general knowledge rather than specific personal information. 
5. For professional advice, consult experts in the relevant field. 

Remember, I'm here to help with a wide range of topics, and I'll do my best to provide useful information while staying within appropriate boundaries. If a query is problematic, I'll typically explain why I can't assist and suggest alternative approaches if possible.

One of my students whose partner's project got flagged as "unsafe" got the AI to provide further details for that one case:

I understand your frustration with the content being flagged. While I don't have access to the specific algorithms used to flag content, there are certain words and phrases commonly associated with explicit sexual content that might trigger such flags. These often include: 

• Anatomical terms for genitalia 
• Explicit verbs describing sexual acts 
• Slang terms for sexual activities 
• Graphic descriptions of bodily fluids or functions 
• Certain words associated with pornography or sex work 

To work around this, you could try: 

• Using more clinical or academic terminology 
• Focusing on the themes and concepts rather than specific acts 
• Summarizing the content instead of quoting directly 
• Using euphemisms or less explicit language 

If you'd like, you can try describing the general topic and main points of the paragraph without using potentially flagged words. I can then suggest how to phrase it in a way that's likely to pass content filters while still conveying the essential information.

Thus an AI would not be able to work with some of Allen Ginsberg's or Walt Whitman's poetry, Toni Morrison's Beloved, and many other literary works, let alone news accounts of many topics that contain graphic language or imagery. Such prudence appears well intentioned but bothersome to this writer, since our purpose was not to titillate a reader but to inform and educate with the help of the cybernetic co-pilot. 

We very much need Humanists, not just Business and STEM folks, using AI in canny ways. Some of our work will involve dangerous topics. We Humanists ask some very hard questions not germane in other fields of study.

Good. AI and its training data need cross-examination. I'm heartened that issues of sustainability and social justice have begun to be considered. All technologies include unforeseen social and economic consequences.

Before we part, back to why I revived "In a Strange Land." I feel the same Uncanny Valley moment I did in 2007, when I first logged into a virtual world. As with that, AI reveals new and sometimes unsettling terrain.

Incidentally, I won't rule out future posts about virtual worlds, though Iggy my Second Life avatar gave up his virtual land and wanders penniless around the once-hyped simulation. I also dip into a roleplaying game build inside SL occasionally, playing a virtual pirate, to conduct naval battles with others interested in The Age of Sail. 

Avast! The Royal Navy will never get me!

Images: Prompt to Dall-E, "Female undergraduate working in library with AI Copilot" and from Second Life, the Dread Pirate Rumbottom O'Malley and Gray Nick the parrot.