Premise:

A chatbot that helps people find their personal vision towards co-housing.

Synopsis:

In Europe, home ownership is on a steady decline. Rising costs of living and decreasing wealth-income ratios have resulted in many individuals, especially young people, being caught in a vicious cycle of renting. However, a promising solution is gaining momentum: co-housing. This movement involves communities of various demographics coming together to co-design and establish new shared living spaces. Despite its growing popularity, co-housing faces numerous obstacles that prevent it from becoming mainstream. Interested individuals are often confronted with confusing terminology and misleading preconceptions. This project aims to explore the potential of a conversational agent in clarifying what co-housing is about and helping to find one’s own preferences, thereby offering a first step towards co-housing as a viable alternative.

Background:

People all over Europe are facing the challenges of unattainable home ownership, living with parents well into adulthood and having renting as the only option. This development leads not only to a growing inequality of wealth, but also to severe psychological consequences for those affected such as a perceived lower status in society and financial anxieties (McKee et al., 2020). Co-housing seems to be a promising solution to these restricting realities of housing. The concept is an umbrella term for many different, alternative housing practices, where private units exist but some amenities are shared with the community. It is often associated with ecological sustainability (Tummers, 2016) and healthier communities that can work against the rising loneliness in cities (Scanlon et al., 2021). Its appeal is reflected in a renewed academic and social interest since the early 2000s (Tummers, 2016). Today, co-housing communities are differentiating themselves from their more ideological progenitors like Eco-villages or Communes (Sargisson, 2012).

However, there is a disconnect between the promises of co-housing, public interest in it and how many projects are actually realized. In most European countries, the co-housing share of the housing stock is seldom exceeding even 1% (Tummers, 2015). This could be traced back to the difficulties that communities face in an capitalist housing market that is deeply ingrained in existing planning structures (Scheller & Thörn, 2018), but also to the confusing terminology and diversity of concepts that are obfuscating the perception of what co-housing is (Corfe, 2019). Even in academic literature clear differentiations of terms are only a recent development; there are confusing and interchangeable uses of co-housing, co-living, collective housing, collaborative housing and many more (Tummers, 2016). On individuals, the existing confusion is visible in the common idea that co-housing involves multiple households living under the same roof without any privacy between them (Sanguinetti & Hibbert, 2021).

Exploration

Objectives:
These critical explorations aim to examine the possibilities of conversational interfaces for solving the entry problem of co-housing for individuals.

To do so, a chatbot is created that: (1) Confronts users with the diverse possibility of co-housing and (2) helps them find their vision of shared living. These objectives are evaluated on (1) knowledge gained by users and (2) their perceived satisfaction with their vision. Multiple iterations of the interface are developed and tested with users.

 

Iteration 1
Creating a basis

Goal: Get initial insights into chatbot usage for communicating complex topics
Experiment: A basic conversational interface was designed in Figma and then developed as a website. It runs on OpenAIs API with the gpt-3.5-turbo language model.

Learnings: It quickly became clear that the chatbot has to be restricted much more to stay on track. I created a dialog path consisting of talking points that should be used in the conversation, ranging from introducing the chatbot to talking about the user’s current living situation to the exposition of co-housing concepts. However, the chatbot often followed the natural conversation with the user without any effort to return to the designed path.

When using a large language model (LLM), it comes down to prompt engineering or adjusting its temperature, which can make the responses more deterministic (Amjad et al., 2022). Another way to make the responses more aligned with what I want to archive is to use multiple steps to analyze responses before showing them to the user. This approach can prevent hallucinations and produce responses that are traceable in their logic (Creswell & Shanahan, 2022).

To incorporate these learnings into the next iteration, I used more specific prompt engineering to follow the dialog and give the chatbot more context.

 

Iteration 2
Creating a framework

Goal: Make the chatbot more specific in its interaction and the co-housing topic
Experiment: The chatbot was extended with a framework to support its usability. The interface was adjusted to separate the whole conversation into multiple sections. The chatbot was embedded into a system that gives context about itself, the user, and the section of the conversation. Furthermore, the interaction was designed in more detail.

Learnings: Multiple user tests were conducted with people without a clear conception of co-housing. The results were better than in the first iteration; not only did the chatbot take the leading role in the conversation, but the users were more inclined to engage with the topic instead of trying to abuse the system. When questions were too long, users tended not to read them very carefully, sometimes leading to talking points not getting across. Also, the chatbot often asked multiple questions, which led to users only answering the last one.

To fix these issues, I set out to improve the prompts for the next iteration and test the use of multimodal responses.

 

Iteration 3
Giving the conversation context

Goal: Further improve the dialog to follow its designated path while staying natural.
Experiment: I wanted to give the user more context about the conversation and be more specific about what step we are talking about. The interface was adjusted to have an intro page and multiple sections. Furthermore, I wanted to experiment with more engaging responses while staying in the conversation flow. For that, I came up with multiple choice boxes at specific points that are easier to understand than reading long text messages.

Learnings: The third iteration was tested on a few of the previous participants to see if they felt that it improved and on a few new, unfamiliar participants to get clean impressions. For most parts, the chatbot followed the flow quite well. In one test, the chatbot went over the goal of exploring, to recommend personalized co-housing communities. The user was pleased and excited because the bot seemingly considered every aspect they discussed. However, as it turned out, it hallucinated every info about the suggested community. This is something that has to be prevented in successive iterations.

User: “Yes i would like to live in Darmstadt in Germany. As i said i would like to have a shared garden.”
Chatbot: “Great, let me see what I can find for you in Darmstadt, Germany. After a little bit of research, I found that there is a co-housing community called “CoHab – Leben in Gemeinschaft” located in Darmstadt. According to their website, they offer a shared garden and a variety of communal spaces. Another co-housing option in Darmstadt is called “Mehrgenerationenwohnen Am Odenwaldplatz”. […] “

Conclusion

I explored the possibilities of using a conversational interface to solve the entry problem of co-housing for individuals. For this, OpenAIs gpt-3.5-turbo model was used in a website where users can chat with an agent to find out about co-housing and find their personal preferences. Over multiple iterations, user tests were conducted. The two set goals were met in later iterations, as users were satisfied with their vision, felt heard, and had a better conception of co-housing after the conversations. The main learnings were that we still need to design the interaction to fit the context and that the conversation has to be guided by the application to reach its goal. To do so, I tried multiple approaches like better prompt engineering, dividing the conversation into multiple sections, and applying multimodality to the bots’ responses.

Although the development was moving in the right direction, future iterations must improve the used approaches further. There are possibilities to design the dialog in more detail while maintaining the natural aspect that LLMs offer. I used GPT for this research instead of more conventional chatbot technologies. I was interested in its effectiveness in communicating complex topics without having the designer think of every possible conversation path. However, through the experiments, it became clear that a combination of framing tools with the capabilities of language models is necessary to provide a good user experience while focusing on the goal of the conversation.

References

Amjad, M., Dahal, S., & Chávez, L. H. (2022, September 21). Productizing Large Language Models. Replit Blog. https://blog.replit.com/llms

Corfe, S. (2019). Co-Living: A Solution to the Housing Crisis?

Creswell, A., & Shanahan, M. (2022). Faithful Reasoning Using Large Language Models. https://doi.org/10.48550/ARXIV.2208.14271

McKee, K., Soaita, A. M., & Hoolachan, J. (2020). ‘Generation rent’ and the emotions of private renting: Self-worth, status and insecurity amongst low-income renters. Housing Studies, 35(8), 1468–1487. https://doi.org/10.1080/02673037.2019.1676400

Sanguinetti, A., & Hibbert, K. (2021). Cohousing by Any Other Name: A Framing Study Exploring Ideological Barriers to Adoption of Collectivist Housing Options. https://escholarship.org/uc/item/05n2r473

Sargisson, L. (2012). Second-Wave Cohousing: A Modern Utopia? Utopian Studies, 23(1), 28–56. https://doi.org/10.5325/utopianstudies.23.1.0028

Scanlon, K., Hudson, J., Fernández Arrigoitia, M., Ferreri, M., West, K., & Udagawa, C. (2021). Community-led housing and loneliness: Research into the impact of community-led housing and co-housing solutions [Report]. London School of Economics. https://www.gov.uk/government/publications/community-led-housing-and-loneliness

Scheller, D., & Thörn, H. (2018). Governing ‘Sustainable Urban Development’ Through Self-Build Groups and Co-Housing: The Cases of Hamburg and Gothenburg: GOVERNING ‘SUSTAINABLE URBAN DEVELOPMENT’ THROUGH SELF-BUILD GROUPS AND COHOUSING. International Journal of Urban and Regional Research, 42(5), 914–933. https://doi.org/10.1111/1468-2427.12652

Tummers, L. (2015). Understanding co-housing from a planning perspective: Why and how? Urban Research & Practice, 8(1), 64–78. https://doi.org/10.1080/17535069.2015.1011427

Tummers, L. (2016). The re-emergence of self-managed co-housing in Europe: A critical review of co-housing research. Urban Studies, 53(10), 2023–2040. https://doi.org/10.1177/0042098015586696