MANILA, July 13 — When you ask an artificial intelligence chatbot a question, the answer might sound confident and complete. But there is a chance that some of what it tells you is simply not true. This phenomenon, known in the AI world as hallucination, happens when large language models generate plausible but false information.
For communities across Asia who are increasingly turning to AI tools for everything from homework help to health advice, understanding this flaw is becoming essential. A 2023 study by OpenAI found that GPT-4, one of the most advanced models available, hallucinates about 15 to 20 percent of facts on specific tasks, particularly medical queries.
That is a significant margin of error for anyone relying on the technology for accurate information. The root causes of hallucination are baked into how these models are built. They train on enormous amounts of noisy web data, which contains plenty of misinformation and contradiction.
They lack real grounding in the world — the models have no lived experience to check facts against. And they generate text one word at a time through a process called autoregressive generation, which can lead them down a path of confident fabrication.
Teaching models to be truthful
Developers have been working on ways to steer these systems away from hallucination. One major technique is reinforcement learning from human feedback, or RLHF, which was used for GPT-3.5 in 2022 and for GPT-4. Human testers rate the model’s responses, and the model learns to prefer the answers that people find accurate and helpful.
Anthropic, another leading AI company, took a different approach in 2022 with what it calls Constitutional AI. Instead of relying solely on human raters, the model is given a set of principles — a constitution of sorts — to guide its own behavior.
The idea is that the model can check itself against these rules before generating an answer. Red-teaming has also played a role. Microsoft and OpenAI conducted adversarial testing in 2023, deliberately trying to provoke the models into producing harmful or false outputs.
These stress tests help identify weaknesses before the technology reaches a wider audience.
Benchmarks tell a mixed story
Standardized tests offer a clearer picture of where these models stand. On the Massive Multitask Language Understanding benchmark, or MMLU, which tests knowledge across dozens of subjects, GPT-4 scored 86.4 percent in 2023. GPT-3.5 scored 70 percent, and Google’s PaLM scored 69.4 percent.
But the TruthfulQA benchmark, which specifically measures how often a model tells the truth, tells a different story. On that test, GPT-4 scored 59 percent in 2023.
That means even the best models are still getting things wrong more than four times out of ten when it comes to factual accuracy. Closer to home, for families and students across the Philippines and the broader region who may use AI tools for research or learning, these numbers matter. A model that scores high on general knowledge can still produce falsehoods with unsettling regularity.
What comes next
One promising approach to reducing hallucination is retrieval-augmented generation, or RAG. Instead of relying entirely on what the model learned during training, RAG systems pull in information from external databases in real time. Google’s REALM, introduced in 2020, was an early example of this technique.
The idea is straightforward: if the model can check a trusted source while generating an answer, it has a better chance of getting the facts right. For communities that depend on accurate information — whether for medical guidance, legal advice, or everyday decisions — this kind of grounding could make a real difference.
As AI tools become more embedded in daily life across Asia, the challenge of hallucination is not going away. But the work being done — from human feedback to constitutional principles to real-time fact-checking — offers a path forward. The technology is still young, and the race to make it trustworthy is very much on.
Sources
{“@context”: “https://schema.org”, “@type”: “NewsArticle”, “headline”: “Hallucination Flaw Found In 15-20 Percent Of GPT-4 Medical Facts”, “description”: “MANILA, July 13 — When you ask an artificial intelligence chatbot a question, the answer might sound confident and complete. But there is a chance…”, “datePublished”: “2026-07-13T09:30:00+00:00”, “author”: {“@type”: “Person”, “name”: “Merry Gel Sigui”}, “publisher”: {“@type”: “Organization”, “name”: “InfoPulse Today”, “url”: “https://newsfeed.infopulsetoday.com/”}, “citation”: [{“@type”: “CreativeWork”, “url”: “https://en.wikipedia.org/wiki/Hallucination_%28artificial_intelligence%29”}]}
