Epistemic Status: Half-Baked Potato

Image Credit: 8one6

Summary:

• Minecraft could be transformative for animal advocacy strategy, but it could also lead to immense suffering if leading AI players begin training LLMs and Agents using it.
• While it is still unlikely that Minecraft will be instrumental in the development of AGI, heavy tail theory suggests that we should dedicate significant resources to ensuring such development (if it occurs) is safe and responsible.
• We have spun off a new organization which deals with both these issues. If you are interested in joining or supporting this organizations, please reach out to me.

Short Intro:

Microsoft has long been at the forefront of technological progress, first for its development of the personal computer, and more recently for its controversial partnership with OpenAI. But I wish to draw attention to a lesser-known subsidiary of Microsoft: Mojang. As I will argue in this short essay, evidence suggests that Minecraft is both an untapped advocacy tool and neglected cause area. At the End, I outline a few preliminary recommendations and close with a call to action.

As Effective Advocacy Strategy

Conversations with individuals in the animal advocacy community have revealed an important split in preferred advocacy methods. Some organizations, like Allied Scholars for Animal Protection, favor a more direct, rights-based approach, using tactics like Cubes of Truth and Dog Meat stands which aim to garner supporters and donors by maximizing publicity. Others have expressed skepticism with such methods, and some research suggests that publicity-maximizing measures may face backlash from members of the general public. These individuals and organizations favor a more conventional and publicly palatable approach.

This presents animal advocates with a difficult tradeoff to make: without the shocking images and demonstrations, animal welfare and rights might not be as popular as it is today; on the other hand, animal advocates may be harming their cause by appearing to the public as disruptive and extreme. Some content may also face institutional resistance; for example, schools may be unwilling to feature speakers who will show graphic footage.

Fortunately, there is a way around this conundrum! Based on a pilot study conducted at the CREEPER, AW MAN Institute, showing footage of factory farming conditions in Minecraft was more effective in creating attitude and behavior change than either no footage or real factory farming footage. We hypothesize that this is because Minecraft simulations capture the benefits of imagery and shock value while also injecting enough humor and psychological distance to cushion the viewer. In qualitative interviews with some of the participants, one theme we heard expressed was that Minecraft was a cogent pedagogical instrument regarding the industrialization of animal slaughter as well as the cognitive dissonance inherent in carnism; according to one respondent:

“I never thought there were things like mob farms in real life. Keeping [the animals/mobs] in such cramped conditions, burning them, forcing them to fall to their deaths, killing them at just a fraction of their natural lifespan… the way we create so many lives just to fill our chests with drops is just unconscionable.”

Pre-test and post-test measures also confirmed that individuals were significantly more likely to believe that Minecraft mobs were marginally sentient in the Minecraft condition vs. the other conditions.

As An Urgent Cause Area

With the rise of LLMs and preliminary AI Agents, the potential creation of immense suffering through simulated worlds like Minecraft has risen accordingly. Some have argued that even computers may have marginal sentience; accordingly, the more complex LLMs and more agentic Agents are even more likely to possess some level of sentience. This may give rise to incidental S-risks in two ways: first, training AI in Minecraft may give rise to highly agentic, uncontrolled, and potentially malevolent AIs; second, such training increases the likelihood that (increased) sentience is evolved.

While LLMs have become increasingly competent in narrow domains, agency and general intelligence seem to elude them. This has encouraged many attempts to use more open-ended domains, such as games, to test and develop more general AI capabilities. Minecraft, being one of the most popular sandbox games, is a presently untapped testing ground for LLMs and Agents, though a few forward-looking individuals have recognized this potential. Indeed, a few connections in China have hinted that it is possible that DeepSeek’s astonishing progress in the last few months is due to extensive training in Minecraft: China Edition.

Given this information, it is likely that OpenAI and Microsoft will move—if they haven’t already—to begin training future versions of ChatGPT in Minecraft. This would not only shorten timelines, but would increase risks from AI agency and sentience. AIs trained to master Minecraft from first principles will, across many training iterations, likely evolve abilities to create complex goals, deal with a wide variety of situations, and act without direction from humans. Since the End goal of Minecraft is to kill the Ender Dragon, it raises concerns that AIs trained in Minecraft will be importantly misaligned, with utility functions which permit or even encourage violence and murder. On the flip side, assuming that the capabilities to create complex goals and act agentically are associated with sentience, there is an increased risk that the training and simulation of such agents could create immense amounts of suffering a la sentient subroutines. Billions of agents would be introduced into simulated worlds, where they would suffer from falling, drowning, burning and suffocation, only to be resurrected to try again. Furthermore, it is possible that, to increase realism, mobs will be given more complex routines, increasing the likelihood of mob sentience and suffering.

Recommendations

• Instead of showing either graphic footage or no footage, animal advocates should show videos of Minecraft mob farms. For example, in Cubes of Truth, the screens can show the internal operations of loot farms, and individuals holding the TVs can wear creeper masks rather than Guy Fawkes masks.
• Strong regulation of training AIs in simulated environments like Minecraft is unlikely, and would likely be outside of the public’s Overton Window. Rather, organizations should adopt voluntary safety policies for training AIs in Minecraft and dedicate more resources to testing and interpreting model capabilities.
• While the forecasted likelihood that Minecraft will play an integral role in the development of AGI is low (~<10%), the downsides of neglecting this possibility is immense (see: heavy-tailed distributions). Thus, EAs should make safe and responsible training of AI in Minecraft a top priority area.
• Recognizing the need for organizations working specifically on Minecraft, I am announcing the launch of a new organization: the Creating Responsible, Effective, and Ethical Persons in Ersatz Realities, Avoiding the Worst and Making Advocacy Nectarious (CREEPER, AW MAN) Institute. The establishment of this institute was made possible by initial seed funding from Open Philanthropy, and we are deeply grateful for their support.
  • If you are interested in joining CREEPER, AW MAN, please reach out to me, and we can schedule a call to talk about what roles and capacities you may be able to fill.