What we at Multitudes believe about building with AI

While the tech industry buzzes with debates about AI replacing human creativity or achieving superintelligence overnight, we at Multitudes see a different reality. AI isn't here to replace human ingenuity; instead, it's a powerful amplifier that enhances what we already do best. The path to meaningful AI advancement isn't about racing toward some distant AGI finish line, but about building practical tools that solve real problems today. We believe in transparent, human-centered development where people understand and control the technology they're using. The future of AI isn't about humans versus machines — it's about humans with better tools.

Our team sticks to a set of core principles when it comes to developing AI features for our app, namely that:

1. AI has to solve a genuine use case
2. AI needs to be built with guardrails
3. People are entitled to transparency around what we’ve built with AI
4. Users should have easy pathways to giving feedback

‍

1. AI has to solve a genuine use case

This is the same principle that startup founders have been coached on for ages: Start with the problem, not the solution. But for many, AI is very much a solution looking for a problem.

When building with AI, the Multitudes team starts with genuine user needs. Our first LLM feature, Feedback Quality, focused on feedback quality problems that we'd been wanting to solve for years but couldn't tackle effectively with traditional NLP approaches, because they weren’t accurate enough at parsing tone and meaning. LLMs aren't a perfect solution for understanding feedback quality, but they represent a step-change improvement that finally made it possible to address this real pain point. We believe that good AI products emerge from patient problem-solving, not from racing to be first to market with the newest technology.

‍

2. AI needs to be built with guardrails

Our team also believes that an important part of the AI-integrated development process is guardrails. Because machine learning models are non-deterministic, employing them can come with an inherent uncertainty that creates problems (Cooper and Frankle et al., 2024) – difficulties with reproducibility, inconsistent quality between runs, and version control issues to name a few (Lones, 2024). There’s also the added factor of LLMs current limitations not being widely understood:  understanding the machine learning technology that we’re building on involves being very clear about what they can be calibrated to do, what they can assess, and what they cannot (Steyvers and Tejeda et al., 2025).

To combat the concerns introduced above, the Multitudes team follows a set of guardrails when building with AI. These guardrails have our data ethics principles embedded at their core: what this looks like for modeling means that we’re careful about the customer data that LLMs have access to. Because we value reciprocity, there are specific things that we will and will not measure, and we’re upfront about that with all of our users– we only collect people’s data if we can give them value for it, and in this case, access to diverse user data is important for us to mitigate algorithmic bias in what we build.

Models that are deployed as part of our app are closely monitored for drift, and we limit them to small, discrete use-cases where usefulness is specifically targeted to the needs of our users. This makes it easier for us to validate their output, ensure data security, and to carry out extensive verification. We also ensured that the team in charge of labeling consisted of a diverse range of individuals, and we were selective about the evaluation metrics that we were using to measure the performance of the models. 

When we developed our last feature involving AI being used to interpret human interactions, we were aware that it would be controversial. Therefore, before working on an overall approach and building anything, we made sure that everything was grounded in research. Our process started with speaking to academics, then reviewing literature to ensure that we were doing the right thing in terms of building with LLMs.

We were also in close consultation with experts who formed part of our existing user base throughout the development process, and strongly believe in participatory modelling – an approach where diverse stakeholders are actively involved in building and refining models. This allows us to test and include people from marginalized groups in our design process, and we endeavour to diversify the background of stakeholders involved.

‍

3. People are entitled to transparency around what we’ve built with AI

Documenting and applying the above standards across the board is important to ensure that we’re acting in concert with another key principle: transparency for users and the public around what we’ve built with AI.

Multitudes is focused on providing transparency around AI because we believe that especially for a new and emergent technology, sharing our work is how we get more feedback on it and improve the quality of how we execute with it. Transparency is also important for building trust. We write extensive, publicly available help center documentation for every feature that we put out detailing not just how to use the features, but also how we built them. When it comes to metrics, our docs share what we measure, what the research says about why it matters, what good looks like (based on benchmarks), and how we calculate it.

‍

We provide references to literature that we’ve gone through, include suggested reading as part of our public-facing content, and lay out the experimentation that we undertake when developing our features. We consider this approach as corollary to one of our guiding principles around data – because we value transparency in data, we also require transparency when operating with AI.

4. Users should have easy pathways to giving feedback

We’re aware of the controversy around AI adoption in the wider industry as well as within the teams that make up our core customer base. The scepticism surrounding this technology in senior teams prompted us to build in detailed mechanisms for users to give feedback to us on our features.

One such mechanism is allowing users to flag any LLM classifications within our app’s Feedback Quality feature that they think are wrong or inaccurate; when something gets flagged in this manner, a person on the Multitudes team is notified and attends to reclassifying the requested data. It’s important that we’re setting an example by fostering frank communication between us and customers on what we can be doing better, especially given that our first key principle around building with AI is that we’re doing it to solve problems for our users.

Receiving this feedback empathetically is important, but so is operating from a position of humility; there needs to be trust between us and the users of the Multitudes app if the feedback given is going to be useful, and eventually result in accountability and actionable changes.

This desire for accountability is core to our data ethics principles at Multitudes, and we consider it foundational for our work. By constantly encouraging feedback from users, being open when we seek it, and then actioning it when it comes through, we’re able to hold ourselves to a higher standard when it comes to the features that we build into our app – data gathering during feedback helps us minimize the risk of using AI where it isn’t needed, because our core motivation is identifying and solving genuine use cases for our audience. 

It’s not about the AI, it’s about what you do with it

The principles outlined above shape how we approach product development at Multitudes, even as industry pressure mounts to integrate AI everywhere simply because it's the latest breakthrough. Using AI for the sake of participating in the done thing isn’t the point; the point is to solve problems for people, and to use AI as part of that process only if it can help solve those problems.

In developing a product with AI features, we at Multitudes are forced to contend with some of the extant threats of this emerging technology: do the benefits related to our use cases outweigh the effects of exploring AI-integrated engineering on the environment and copyright? Is the conversation around the environmental effect of using AI something that should also factor more heavily into our development conversations? What about the ongoing security concerns around using non-deterministic code? What are the ethical implications of training language learning models on people’s work? While the jury is still out on a lot of those questions, we believe that these questions give us even more reason to commit to the core motivation of using AI in the service of our users over anything else.