SAN FRANCISCO, July 8 — When a large language model generates text that sounds confident but is completely wrong, the artificial intelligence industry calls it a hallucination. For anyone using AI tools in daily life — whether to draft an email, research a health question, or help a child with homework — the phenomenon raises a practical concern: how often does it happen, and what is being done about it? A 2023 study by the research firm Vectara put numbers to the problem.
The study found that GPT-3.5, an earlier version of the model behind ChatGPT, produced false or nonsensical information in 3 percent of its responses. GPT-4, the more advanced model, hallucinated in 2.5 percent of responses.
LLaMA 2, an open-source model from Meta, showed a 4.5 percent hallucination rate. These figures, while small in percentage terms, mean that users cannot assume any single output is reliable. The root causes lie in how these models work.
Large language models do not truly understand the information they generate. They operate by predicting the next most probable word based on statistical patterns learned from vast amounts of text scraped from the internet.
When the training data contains contradictions, errors, or gaps, the model can overgeneralize — producing statements that are grammatically fluent but factually empty.
Bias baked into the data
Hallucination is not the only risk. The same statistical patterns that lead to falsehoods also encode the biases present in the training data. A 2020 paper titled “On the Dangers of Stochastic Parrots,” authored by Emily Bender and colleagues, warned that large language models can amplify harmful stereotypes and that their apparent fluency masks a lack of genuine understanding.
The paper became a touchstone in debates about AI safety. The Stanford HAI 2023 AI Index provided concrete evidence of gender bias.
The index found that GPT-3 exhibited clear gender associations — linking the word “nurse” with female and “doctor” with male. For communities across Asia, where cultural stereotypes may already shape opportunities, an AI tool that reinforces such biases could have real-world consequences in hiring, education, and access to services.
Techniques to tame the models
Researchers and companies have developed several approaches to reduce hallucinations and bias. One prominent method is reinforcement learning from human feedback, or RLHF. OpenAI used RLHF in InstructGPT, which was released in January 2022.
The technique involves human evaluators rating model outputs, and the model is then trained to prefer responses that humans find accurate and helpful. Anthropic, another leading AI company, introduced a different approach in 2022 called constitutional AI.
Instead of relying solely on human raters, constitutional AI gives the model a set of principles — a constitution — that guides its behavior. The model is trained to self-correct when its outputs violate these rules. A third technique, retrieval-augmented generation, or RAG, was introduced by Lewis and colleagues in 2020.
RAG works by having the model check its responses against external sources of information rather than relying solely on its internal patterns. This allows the model to pull in up-to-date facts from databases or the web, reducing the chance that it will invent something out of thin air.
Benchmarks and consequences
To measure how well these techniques work, researchers have developed evaluation benchmarks. TruthfulQA, introduced in 2021, tests whether models can answer questions truthfully rather than repeating common misconceptions. HaluEval, released in 2023, is designed specifically to detect hallucinations in model outputs.
The legal system has already felt the impact of hallucination. In May 2023, a lawyer used ChatGPT to prepare a federal court filing in the case Mata v.
Avianca. The chatbot cited several court cases that did not exist — entirely fabricated rulings. The lawyer was sanctioned by the court.
The incident became a cautionary tale for professionals who might be tempted to treat AI-generated text as a reliable research tool.
Regulators step in
Governments have begun to respond. The European Union passed the AI Act in March 2024, which requires companies deploying large language models to be transparent about their capabilities and limitations and to manage the risks of harmful outputs. In the United States, an executive order issued in October 2023 mandates safety testing for the most powerful AI models before they are released to the public.
These regulatory moves reflect a growing recognition that hallucination is not a minor glitch — it is a structural feature of current AI systems. For anyone relying on these tools, the message from researchers is consistent: verify the output.
The technology is improving, but it has not yet learned to tell the truth on its own.
{“@context”: “https://schema.org”, “@type”: “NewsArticle”, “headline”: “**AI Hallucinations and Bias Remain Major Issues for Language Models**”, “description”: “SAN FRANCISCO, July 8 — When a large language model generates text that sounds confident but is completely wrong, the artificial intelligence industry calls it…”, “datePublished”: “2026-07-08T09:30:00+00:00”, “author”: {“@type”: “Person”, “name”: “Merry Gel Sigui”}, “publisher”: {“@type”: “Organization”, “name”: “InfoPulse Today”, “url”: “https://newsfeed.infopulsetoday.com/”}}
