Staking out a clear personal position on AI these days is a risky but necessary undertaking. AI has become one of the most controversial topics of our time, and will continue to be divisive as its impact on the economy, the environment, and the law becomes apparent. The controversiality of AI only intensifies when we consider its impact on education, which was already a very polarising topic. 

While it would be nice to “wait and see” what happens before making clear our individual stances on AI, that’s not really possible since the technology is already out there. Pandora’s box is open. The horse has bolted. The cat is out of the bag. Pick your metaphor. Even if we educators would rather ignore it, we can’t, because our students are already using it. 

It’s impossible to comprehensively cover all the common criticisms of AI in one sitting, but there are a few arguments that come up again and again that tend to make those of us who use AI in a moderate, judicious, and conscientious way feel unnecessarily guilty — and it’s worth taking the time to contend with them. 

The “soulless” nature of AI texts

One oft repeated criticism of AI is that the texts it generates are “soulless”. If we exclude religious understandings of “soul” and assume this remark relates to the fact that AI is not conscious or sentient, then yes, it is clearly true. AI is a “calculator for words” and calculators do not have souls. However, not all forms of writing need to have “soul”.  I would argue that genres such as purely factual, informative, or instructional writing constitute a niche where AI can excel without a soul. I am actually more concerned with attempts to make AI seem to have soul when it does not, which can be deceitful and disturbing. If we are using AI generated texts in our classes, we should be open and honest about that, but there is no reason why they cannot be a useful supplement to other more “soulful” human-authored materials.

Lack of respect for authors’ legal rights

Another argument connects to the idea that the legal rights of authors around the world have been ignored during the training process of some Large Language Models. LLMs like ChatGPT were literally trained on the whole of the publicly accessible Internet. These data sources were already being made readily available to any organisation with enough compute to make sense of them. Big Tech has always sought forgiveness rather than permission in its attempts to make Big Data more useful. It happened when Google scanned all the books in the public library (authors sued Google, Google won) and it’s happening again in the wake of OpenAI’s decision to train its models on the entirety of the open Internet. 

While it would have been unworkable to consult with every blog author and forum poster to assemble a whitelist of only those who consented to having their writing included in the model, OpenAI should have proceeded with more caution and public consultation. They are, after all, being sued by numerous authors’ organisations. But this is a matter beyond the influence of the average educator, and will be hashed out in America’s courts between OpenAI’s attorneys and the attorneys bringing the class action lawsuits against them. If OpenAI are found to have breached copyright in the way they trained their models, then they will be dealt with in accordance with the relevant laws. But when it comes to texts generated by their models, I don’t see how it will be technically possible to show that any specific text violates any specific individual author’s copyright (beyond instructing it to generate a text in a well known author’s voice or style, and even this could be protected by parody or fair use exemptions). 

The environmental impact of AI data centers

Another argument that rightly causes much concern is the impact of AI on the environment. AI is for sure a power hungry technology, and if that power is coming from non-renewable sources, that’s going to have a negative effect on the environment. Cooling the infrastructure required to run AI inference at scale also requires a lot of water, but many server farms are able to recycle the water they use. Even in cases where water is not recycled, it returns to the atmospheric water cycle and isn’t lost forever. In any event, environmental concerns are rightly high on the list of our reservations about AI. But there are plenty of other environmentally polluting industries that deserve just as much scrutiny. And none of them — other than AI — have the potential to come up with ways to reduce their own environmental impact. 

The threat against our livelihoods

Finally, there are concerns about the impact of AI on our jobs and livelihoods. It’s natural to be worried about such things, and skeptical of claims by AI CEOs that AI is only good at tasks not jobs, or that AI will bring about more jobs that it takes away. Even if that’s true, it won’t be easy to retrain for these jobs, which are likely to require a very high degree of education and expertise. 

But education is and will remain a human-centric social process. When we try to come up with a working definition of what it means to “learn”, it invariably involves being able to use, understand, and apply knowledge we have gained in an unassisted way. As educators, we need to utilise AI in a way that enhances the pedagogical process by supplementing human-centric teaching and learning, ensuring it does not negate or replace this process, or leave students in a position where they are totally reliant on technology and unable to write, speak, or think coherently without it.

In conclusion

I’ll end this post by reminding readers that I am not making light of AI’s potential negative impacts on the environment, economy, or authors’ legal rights. These are all very serious issues that need to be resolved by experts in their respective fields. Additionally, I am not an AI ideologue or fanatic by any means. I am open to changing my position on the issues I have highlighted above in accordance with emerging evidence. It is paramount to stay abreast of not only the technical but also the environmental, ethical, and legal implications of AI.

For now at least, in my career as an English language educator, I will continue to use AI to supplement and augment my human-centric lessons by having it do pedagogically beneficial things that I couldn’t hope to do by myself, with a view to making my teaching more engaging, effective and efficient. 

The image accompanying this article was generated by AI. The text was entirely composed by the (human) author.

Introduction

The COVID-19 pandemic has forever changed the way that we work and learn. The “new normal” in the 21st century is for students to engage with their teachers and peers in both physical and virtual learning environments. In the summer of 2022, Paul Raine and Raquel Ribeiro explored 6 different virtual worlds, and evaluated their viability for language teaching and learning. Only virtual worlds that run in a web-browser and have a free trial were selected for this project. In this article, we present the results of our investigation, with the hope that this information will be of interest to other language teachers looking to teach all or some of their classes online. A series of YouTube videos to accompany this article can be found here.

Overview

The following virtual worlds are review in this article:

1. Spatial.io (video)
2. Gather Town (video)
3. Wonder* (video)
4. Orbital** (video)
5. Mozilla Hubs*** (video)
6. KumoSpace (video)

*In April 2023, Wonder announced it would be shutting down.
**In December 2022, Orbital announced it would be shutting down.
***Mozilla ended support for Mozilla Hubs on May 31st, 2024.

	Spatial	Gather	Wonder	Orbital	Mozilla Hubs	Kumospace
Perspective	3D	Top-down	Top-down	Top-down	3D	Top-down
Audio & Video Chat	Yes	Yes	Yes	Yes	Yes	Yes
Text Chat	Yes	Yes	Yes	Yes	Yes	Yes
Screen Sharing	Yes – Embedded	Yes	Yes	Yes	Yes – Embedded	Yes
Customizable Avatar	Yes	Yes	No	No	Yes	No
Sticky Notes	Yes	No	Yes	Yes	No	No
Object Picker	Yes	Yes	No	Yes	Yes	No
Interactive Environment	Yes	Yes	No	No	Yes	No
Emoji Reactions	No	Yes	Yes	Yes	Yes	Yes
Mini Games	No	Yes	No	No	No	Yes
Web App	Yes	Yes	Yes	Yes	Yes	Yes
iOS App	Yes	No	No	No	No	No
Android App	Yes	No	No	No	No	No
Forever-free Plan	Yes	Yes	No – Free Trial	Yes	Yes	Yes
Best Feature	Dance moves	Mini Games	Randomise Users	Private Island	Laser Pen	Piano Music

Common Features

Perspective

The virtual worlds (VWs) we evaluated for this report came in two different perspectives: top-down and 3D. Four of the six worlds had top-down perspectives, and two offered a full 3D experience.

Figure 1: The top-down perspective of Kumospace

Figure 2: The 3D environment of Spatial

Audio and video chat

All the VWs we evaluated had the ability to chat via live video and audio with other members. In some VWs, the video stream was rendered as the user’s avatar, and in other VWs, the video was rendered above or to the side of the environment.

Figure 3: In Orbital, the user’s video stream is rendered as their avatar.

Figure 4: In Wonder, the user’s video stream is rendered above or to the side of the environment.

Text chat

All of the VWs we investigated in this study offered the ability to send text-based messages to other members of the environment. We found that the text chat was a very useful supplement to audio-visual teaching methods, especially when teaching new words with unfamiliar pronunciations.

Screen Sharing [Embedded]

All of the VWs in this study offered the ability to share a screen with other users in the environment. In Mozilla Hubs and Spatial, the shared screen was “embedded” in the environment such that users could choose to either continue interacting with each other, or view the shared screen from a variety of perspectives.

Figure 5: Sharing a screen in Spatial

Custom Avatar

Some of the VWs we investigated allow the user to customise their avatar in various ways. The most advanced and personalised customization was offered by Spatial, which provided a way to convert a digital photograph into a 3D head for a user’s avatar.

Figure 6: Spatial offers the ability to convert a photo into a 3D head for a user’s avatar

Sticky Notes

Most of the VWs we investigated offered the ability to add “sticky notes” to the environment. These came in useful when teaching new words or phrases.

Figure 7: The “sticky note” function in Orbital

Interactive Environment

Some of the VWs we investigated offered the ability to interact with one’s environment. For example, by picking up and moving objects around. This feature could be used for teaching prepositions, by instructing learners to “put the goldfish on the wall” for example.

Figure 8: Interacting with a 3D goldfish in Mozilla Hubs

Mini Games

Two of the VWs we investigated featured “mini games” such as chess, which were completely contained within the virtual world. It is possible that these mini games could be used for spoken fluency practice by higher level language learners.

Figure 9: An invitation to play chess within the Kumospace virtual environment.

Object Picker

In addition to being able to interact with one’s environment, some VWs also offer an “object picker”, “designer”, or “build tool” that allows the user to add, remove, or change objects in the environment. This could either be used for introducing new vocabulary items, or for making the environment a more conducive space for language learning.

Figure 10: The object picker within Gather allows the user to add a wide range of weird and wonderful items to their virtual environment.

Emoji Reactions

Most of the VWs we investigated allow the user to react with a variety of emojis. These could be used for expressing comprehension, interest, or confusion when learning a language in a virtual environment.

Figure 11: Reacting with a “heart” emoji in Gather

General Suitability for Language Learning

The authors found that in general, it was possible to learn new foreign words and phrases inside of the VWs investigated in this study. This was verified in a rudimentary way by learning words and phrases in Portuguese and Japanese. One author had a native level of Portuguese, and attempted to learn some basic Japanese. The other author had an intermediate level in Japanese, and attempted to learn some basic Portuguese. Both authors were complete beginners in the language being taught to them. 

It was found that the fidelity of the audio stream was of paramount importance in the language teaching and learning process. Where the quality of the audio was bad (such as in Spatial) it was sometimes not possible to distinguish between similar consonant sounds, such as “d” and “b”. For instance, when the Portuguese word for “chair” was first introduced, it was initially pronounced by the learner as “cabeira” whereas the correct pronunciation is “cadeira”. The authors found that the chat function could be used to clarify the pronunciation of unfamiliar words when the audio was insufficiently clear.

Specific Methodologies

Because both authors were complete beginners in the languages they were learning (Portuguese and Japanese) simple “show and tell” and “listen and repeat” methodologies were the main ones adopted in this preliminary investigation. In addition, Total Physical Response (TPR) was also briefly trialled, with one author being instructed by the other to “go closer to the tree” in Portuguese. 

It is expected that, in reality, learners using VWs would not be complete beginners in the languages they are studying. Therefore, it seems reasonable that methodologies such as Communicative Language Teaching (CLT) or perhaps even Task Based Language Teaching (TBLT) could be adopted, and that this would result in improvements to communicative and pragmatic competence in a similar way that it would in real, physical classrooms.

Issues and Limitations

There were several issues and limitations with the current study. Firstly, the authors involved were living on opposite sides of the world, with a 12 hour time difference. This sometimes made it difficult to find a suitable time to meet up. It also caused occasional network issues. 

In addition, the authors encountered some audio fidelity problems in the Spatial virtual environment, which interfered with the ability to clearly understand the correct pronunciation of unfamiliar foreign words.

Finally, only the two authors in this study were able to participate in the environments investigated. In real life situations, it is highly likely that there would be more participants in a language learning environment, including the teacher and perhaps 10 to 20 students. The effect that this number of users would have on the quality of the language learning experience is not known, and should be further investigated. Many of the VWs investigated in this study were specifically designed to handle a large number of concurrent users, and it would be interesting to see how the affordances of these virtual environments could be leveraged for larger classes.

Conclusion

Although Zoom has become the de facto application for online synchronous communication, it is not the only way to connect with people in remote locations in real time. The authors found many of the above virtual worlds to be just as reliable as Zoom and in most cases more visually engaging and stimulating. Language teachers might like to consider one or more of the above options in addition to or instead of Zoom for their online language classes.