With overdose as the leading cause of death for adults ages 18–45 in the U.S., the search for the solution to the drug crisis is becoming more pressing every day. As a harm reduction organization implementing cutting edge technology at every opportunity, we have harnessed the power of AI to create a novel harm reduction resource that both educates and informs, with the potential to save countless lives in the process.

Over the last 13 years, much of our work as The Bunk Police has been boots-on-the-ground, face-to-face conversations with individuals who are in an environment where drug use is likely to occur. We try to be in the right place at the right time for those choosing or planning to consume drugs—namely in the campgrounds at music festivals—to bring education and exposure to harm reduction practices and overdose prevention.

Despite having served hundreds of music festivals over the last decade, this is not where we distribute most of our test kits. While we gain a lot of awareness from in-person initiatives, most of our audience orders test kits from our website. Of course, this reflects the large population of drug consumers outside of music festivals, but it also speaks to a significant pattern of behavior. Because the online space provides a layer of anonymity and a feeling of safety, it is often the preferred means of interaction for those who may be interested in learning about harm reduction and safer drug use. People can explore this information from the comfort of their home or computer, without the requirement of talking to someone face-to-face (which can be intimidating for many). For this reason and many others, our website is the first point of contact for those beginning to dabble into substances, those who may have a child at the age of experimentation, and those who have friends or family who are frequently consuming drugs.

As we’ve grown as an organization and extended into the nonprofit space, we have looked for ways to broaden our reach in the virtual space, too. Harm reduction help should be easy to access, free, and plentiful, but high quality resources are far and few between. As a response to the need we saw continuing to grow, we wanted to create an all-encompassing resource for people to access from anywhere—one that could provide testing help, harm reduction knowledge, test kit browsing, and more. But not simply as information on a page—we wanted to create an interactive experience, as if you had a harm reduction guide at your side at all times.

We took the research we’d been conducting over the last decade, the responses and questions we’d received from our audience, and the budding technology of artificial intelligence to create the Transparency Harm Reduction App—a harm reduction guide at your fingertips. 

With this app, users can interact with our “chatbot” to get step-by-step guidance during the testing process, ask questions about the kits, learn harm reduction info, and more. One of the most significant functionalities of our bot is the ability to showcase how a substance should react with a test kit. For context, when you drop a bit of the test kit chemical onto a substance, it turns a particular color, indicating what substance is present in a sample. When a user tells our chatbot the drug they’re testing and the kit they’re using, our bot is trained to provide the matching reaction video, showing exactly how their reaction should look and what color it should turn. 

Creating this app was no easy task, due to the inherent nuance in drug testing, as well as the difficulty in utilizing the newest AI technology to accurately assist in results interpretation. Our AI-development team, led by developer Dan Kelly, worked tirelessly to iron out the details in ensuring our chatbot gives accurate answers, and behaves as predictably as possible. 

In creating our AI, Kelly utilized a combination of natural language processing to properly filter user intents, along with the latest generative AI models to construct replies, based on a set of knowledge base files we have been developing over the last 13 years. Our Reaction Video knowledge base has over 2500 possible responses, and our FAQ and How-To knowledge bases have over 100 possible responses, not including the permutations of answers that may combine data from different questions, or the numerous iterations of street names, substance synonyms, and other user-behavior intent possibilities, such as potential misspellings. 

We aimed to make the app as simple as possible, while utilizing our robust knowledge bases, and providing a selection of applicable functionality. We’ve designed the bot to avoid answering off-topic questions that are not contained in its knowledge bases, and we’ve limited the generative answering capabilities to keep users focused and costs low. 

Some portions of our app utilize generative AI. For instance, the latest GPT 4o model is highly effective at interpreting complex text information and formulating concise responses for users asking general FAQs and “How-To” style questions. 

The portion that provides our reaction videos required a more manual approach to maintain the highest level of accuracy, as well as structure the data consistently for the user. This required numerous hours of updating our natural language library to account for substance synonyms or street names for each substance. This natural language processing is housed in Botpress, a platform that allows us to create customized NLM flows while triggering GPT’s generative answering capabilities when applicable. 

It may be obvious that the creation of our harm reduction app was not seamlessly possible with AI— it required the knowledge of diligent developers, meticulous beta testers, and skilled harm reduction practitioners to ensure it performs as smoothly as possible. However, we view this as the jumping off point for how AI and harm reduction could work together moving into the future. 

We believe AI technology may be a crucial tool in the expanding harm reduction landscape. As the AI space broadens and strengthens over time, so will its functionality, its capability, and its accuracy. We continue to dedicate effort into checking and refining the ever-expanding behavior of our AI, but we believe this process will only get easier and easier. 

We envision a future where users can not only inquire about how a substance should look when they test it, but the extremely powerful inverse of that experience. We would like users to be able to provide a photo or video of a reaction, name the test kit they used, and for AI to study that image and to report back with the substance itself—meaning it interprets the reaction for the user, taking out the guesswork of user error in misinterpreting a reaction. 

Of course, this is a functionality that leaves a large window for mistakes and problems, and would have to be thoroughly tested and refined with human intervention. But this level of guidance and assistance is hugely impactful for the substance use space. No longer will an individual need to look for the extremely sparse in-person live testing services—which is largely an impossibility in the U.S. currently (but we’re hoping that changes, too). They can gain knowledge about drug checking and harm reduction right from their phones, and then they can take it a step further, and have step-by-step testing assistance, along with a built-in results verification. While our app makes this possible today, a more robust results interpretation incorporating the analysis of user photos and video is our imagined future.  

We believe AI technology may be a crucial tool in the expanding harm reduction landscape. As the AI space broadens and strengthens over time, so will its functionality, its capability, and its accuracy. We continue to dedicate effort into checking and refining the ever-expanding behavior of our AI, but we believe this process will only get easier and easier. 

We envision a future where users can not only inquire about how a substance should look when they test it, but the extremely powerful inverse of that experience. We would like users to be able to provide a photo or video of a reaction, name the test kit they used, and for AI to study that image and to report back with the substance itself—meaning it interprets the reaction for the user, taking out the guesswork of user error in misinterpreting a reaction.