Everyone’s talking about AI tools that promise to “solve everything.” But what does it really take to apply Vision AI and LLMs to real-world problems like moderation, accessibility, and community safety?
There’s no shortage of posts on LinkedIn, Medium, and elsewhere promising that online AI solutions will solve all your problems. And don’t get me wrong, rapid prototyping with tools like Vibe coding or Figma is incredibly valuable. But at my core, even though I run 4 Roads, I’m still a developer. I like to understand how things work and to build from the ground up.
The speed of change in technology right now is astounding, especially in the world of LLMs and Vision AI. Three years ago, when we started a project on room-sizing and object identification, the field was still young. Back then, techniques like NeRF felt cutting-edge. Today, as Ferris Bueller put it: “Life moves pretty fast. If you don’t stop and look around once in a while, you could miss it.” (how to age someone without asking them their age!) What was once ground-breaking is now commonplace, with pretrained models delivering remarkable results out of the box.
The Challenge of Online Safety
One of the biggest challenges for any online community today is the Online Safety Act (OSA). Its aim, to make the internet safer, is admirable. But, like most digital legislation, it risks being outdated at launch, and it wasn’t fully thought through.
The Act is designed to apply “proportionately” lighter duties for smaller services, heavier ones for giants like Facebook and TikTok. In theory, this sounds fair. In practice, though, even scaled-down requirements can stretch smaller communities beyond their budgets and technical capacity.
Large platforms have vast resources to build compliance systems. Meanwhile, thousands of small, medium, and even large communities in the UK and beyond are forced to compromise: disabling features, ramping up moderation (and hurting user experience), or even sacrificing user privacy just to keep operating. And of course, sites that ignore compliance altogether will continue unchecked.
The result: good intentions, but a new set of problems, especially for communities without deep pockets.
How We’re Responding at 4 Roads
One thing we can do is use our knowledge to help these communities adapt. Over the last few months, I’ve been exploring some side projects:
Training a bespoke LLM to detect abusive behaviour, tailored with specific community data.
Using a vision model, to generate alt text for images, supporting both moderation and accessibility under WCAG.
Using vision models to perform age verification
And looking at other tools such as IP monitoring and link detection to strengthen safety further.
In the coming weeks, I’ll be sharing a few deep-dive articles on the first two from a technical perspective. My aim is to spark discussion on how smaller communities can practically meet today’s challenges without losing the essence of what makes them thrive.
How do you see AI helping communities balance safety, privacy, and user experience in the age of the OSA? Have you seen promising approaches beyond what the big tech players can afford to implement?
First published on LinkedIn on 1 October 2025.
