The promise of artificial intelligence has always been painted in broad, optimistic strokes, a shimmering future where algorithms solve our most complex problems. From the bustling streets of Colombo to the quiet villages nestled in the central highlands, the allure of instant answers and digital efficiency is undeniable. Yet, beneath this veneer of progress, a more insidious reality is taking root, one that threatens the very foundations of trust in our critical public services. My investigation reveals that AI hallucinations, particularly from widely accessible large language models like Google Gemini and OpenAI's GPT series, are not merely generating amusing falsehoods; they are actively causing tangible, sometimes irreversible, harm in Sri Lanka's healthcare and legal sectors.
I've been tracking this for months, observing a subtle but alarming trend. What began as anecdotal reports from junior doctors and paralegals has escalated into a concerning pattern. These professionals, seeking quick information or preliminary research, have increasingly turned to AI chatbots, only to be fed confidently asserted but utterly false information. The promises don't match the reality. Imagine a young medical intern, pressed for time, querying an AI about a rare tropical disease prevalent in our region, only to receive a detailed, yet incorrect, treatment protocol. Or a lawyer, navigating the complexities of Sri Lankan land law, being presented with non-existent case precedents or misquoted statutes. This is not hyperbole; these are scenarios playing out with increasing frequency.
Our journey into this digital quagmire began with a tip from a former colleague at the National Hospital of Sri Lanka, who recounted a harrowing incident. A junior doctor, relying on an AI-generated diagnosis and treatment plan for a patient presenting with unusual neurological symptoms, nearly administered an inappropriate medication. The AI, reportedly a version of Google Gemini, had confidently cited a non-existent medical journal article and recommended a drug contraindicated for the patient's actual condition. Fortunately, a senior consultant intervened, but the close call sparked a deeper inquiry.
Here's what the data actually shows. Through a series of controlled experiments, my team and I systematically queried various AI models with questions pertaining to Sri Lankan medical conditions, local pharmaceutical regulations, and specific legal statutes. We used a range of prompts, mimicking typical user queries. The results were stark. In approximately 30% of medical queries involving specific drug interactions or diagnostic criteria unique to our local context, the AI models generated information that was either partially or entirely incorrect, often citing fabricated studies or misinterpreting established guidelines from the Sri Lanka Medical Council. For legal queries, the situation was equally dire. Over 40% of responses contained non-existent legal precedents, incorrectly cited sections of the Penal Code, or misrepresentations of judicial interpretations from the Supreme Court of Sri Lanka.
One particularly egregious example involved a query about the Prescription Act of Sri Lanka. An OpenAI GPT model, when asked about the limitation period for a specific type of contract dispute, confidently cited a section that does not exist in the current legislation, and then proceeded to invent a Supreme Court ruling to support its claim. This isn't just a minor error; it's a fabrication that could derail a legal case and severely impact an individual's rights. The implications for justice, particularly for those who cannot afford extensive legal counsel and might rely on publicly available AI tools, are chilling.
Who is involved in this escalating crisis? Primarily, the developers of these powerful AI models, companies like Google, OpenAI, and even smaller players whose models are integrated into various local applications. While they issue disclaimers about AI output, the sheer confidence and authoritative tone of the hallucinations often lead users, particularly those less familiar with the nuances of AI, to trust the information implicitly. Locally, the problem is compounded by a lack of robust digital literacy initiatives and an absence of clear regulatory frameworks governing AI deployment in sensitive sectors.
When confronted with these findings, representatives from the Ministry of Health and the Ministry of Justice acknowledged isolated incidents but largely downplayed the systemic nature of the problem. A spokesperson from the Ministry of Health, speaking anonymously due to the sensitivity of the issue, stated,
