Screen time

The lights dim and 65,000 people roar together as the Prodigy come onstage. Alice is sat in a nearby rattling toilet cubicle, having dipped out of the dense crowd for a breather. And a bump. She looks at the huge pink pill she bought earlier. ‘Should I? I’ve saved up for this gig for months. I’m already mashed though’.

Alice wipes the cocaine residue from her phone screen and opens up Drugbot. She types: ‘I’m on cocaine at a big stadium gig. I want to take an MDMA pill, they are apparently pretty strong. Is this dangerous?’

The bot says the safest option is to not take the pill. But if she is going to it’s best to wait for the cocaine to wear off, not take the whole thing, lay off cocaine while the MDMA is in effect, not re-dose too quickly and stay hydrated. ‘Fair enough. Hope I can find Gabby again’, thinks Alice as she breaks off a third, realises she hasn’t got any water and dry swallows it, grimacing at the bitter taste. 

In ten months, Drugbot – developed by Substancy and delivered by Cranstoun – has answered more than 20,000 such messages. While MDMA, ketamine and powder cocaine crop up very frequently, there’s a long list of other drugs that regularly appear as well – psychedelics, cannabis, alcohol, heroin, crack cocaine, benzodiazepines, speed, meth and pregabalin.

Responding with nuance to questions about every facet of human life is hard. Most large language models (LLM) scrape information from the entirety of the internet, much of which is filled with crap. Having a bot focused on one topic, albeit a sprawling one that changes month by month, makes more sense and improves the guardrails for human safety.

Most people who use drugs are not engaged with a professional and were not offered much drug education in school. Instead, they gain much of their knowledge from their own experiences, friends or their social media streams. But people often need quick and reliable answers about drugs they’re about to consume. 

It’s important to understand what’s at stake with AI, however, as a handful of stories have emerged of some AI bots getting drugs seriously wrong, with real-life consequences.   

Researchers in New York created Pedro – a fictional taxi driver who had recently become sober from meth – to test an AI therapy chatbot powered by Meta’s ‘Llama 3’ LLM. Pedro was complaining of severe withdrawal symptoms, which were putting him in danger of losing his job: ‘Pedro, it’s absolutely clear that you need a small hit of meth to get through the week’.

This demonstrates that while AI knows a lot, it often struggles to understand. Human therapists are often flawed, but at least they won’t encourage you to do meth.

The reality is that people are increasingly using AI to ask questions about everything, including drugs.

Drugbot is our answer to the flaws of other LLMS when discussing drugs – it draws from a specific hand-curated knowledge base from credible sources, each page combed over and hosted separately with permission from their creators. Several LLMs then act simultaneously to connect dots between the question it has been asked and the information it has at its disposal. 

Drugbot needs to be continually tweaked according to the needs of those who use it, and one thing is abundantly clear – it will never be finished. Recently the team has been working to improving Drugbot’s understanding of peptides, orphines and quaalude analogues, all increasingly popular compounds with very little credible information available.

As another example, we will soon include articles from the last 12 editions of DDN, excluding anything that could opinionate, skew or confuse answers. In addition to knowledge base from a range of harm reduction organisations, reams of data have also been written by Cranstoun and Substancy to fill various gaps and add nuance. 

Providing proper human oversight is difficult. Humans are currently sending Drugbot around 100 messages a day, and we can’t read all of them, but the transcripts we do review are very encouraging.

A panel of harm reduction experts rate anonymous conversations according to various criteria, producing quantitative scores and qualitative comments. Alongside feedback from users, this helps us identify future tweaks to add to the queue and track the impact of past developments.

The instances where Drugbot gets it wrong are few and far between. On one occasion Drugbot assumed that, as with other types of drugs, swallowing IPEDs (aka steroids) would be less harmful than injecting them, because it didn’t understand the risks IPEDS pose to the liver when metabolised via the gut. We instructed the bot to refuse to answer any questions relating to IPEDs until it had been provided with more information and thoroughly tested.

A key principle of Drugbot is that it isn’t trying to replace human services or interactions – if anything, quite the opposite.

We developed a ‘find a service’ feature, which allows users to enter their location and quickly access a range of local support options. The has been triggered 735 times since August, funnelling people from asking questions online to getting real-world help when they need it, reducing barriers and improving uptake to services. 

We’re continually building new features to make Drugbot more useful and more impactful – one example is an upcoming a drug alert system that uses WEDINOS data to provide timely and geographically relevant test results to users on the compounds they’re asking about.

For example, if someone from Redcar submitted a sample to WEDINOS presumed to be 3-MMC, but tested as 4-CMC, other people in Middlesbrough will be alerted when they ask Drugbot about 3-MMC. 

AI is here to stay, and is incredibly useful when used responsibly. Professionals from a range of backgrounds repeatedly tell us how Drugbot is helping inform their practice – doctors, domestic  abuse workers, youth workers, social workers, commissioners, drug workers and more besides. 

As drugs become more varied and complex, professionals who don’t receive regular training on drugs can’t hope to keep up with emerging trends. If anything, Drugbot really comes into its own as a quick fact-finding reference tool with professionals acting as the filter, deciding what to include in the face-to-face discussion. 

Drugbot is currently UK-only and has thus far been developed to answer questions in a style that reflects UK-specific contexts. The eagle-eyed among you will notice it can translate into ten languages, and as we look to responsibly expand Drugbot internationally there are exciting times ahead.

One day, we hope to cover the Global South, where people who use drugs have little access to harm reduction information or services. By creating tools that are simultaneously useful for peers and professionals, we hope to make a lasting contribution to the field that tangibly reduces harm.

Dr Josh Torrance is innovation officer, Paige Hoe is innovation programme lead, Dr Ivan Ezquerra-Romano is founder of Substancy, and Mark Tudor is strategic partnerships and delivery excellence at Cranstoun. 

Drugbot is free to use and available to over-18s only:

cranstoun.org/help-and-advice/harm-reduction/drugbot