Technological progress has always been measured by better, faster, and cheaper products, but newer artificial intelligence (AI) models are falling short in the “better” area; they are still plagued by hallucinations. Yes, AI still hallucinates, and human AI editors can control it. They are the ones who play the vital role in ensuring that these factual errors don’t turn up in court documents, medical information, or sensitive business data. If left unchecked and uncorrected, these errors can result in physical harm, lawsuits, and reputational damage.

AI hallucinations have been a problem that hasn’t been fixed even after more than two years of development. I wrote about the problem of AI hallucinations in my very first Writing For Humans blog and have continued to see this problem myself when working with the leading chatbots. I explored this problem later in another blog (“Generative AI Content Catastrophe Strikes Again!”) and have been writing about it ever since. 

AI Hallucinations Go Mainstream

Today, it’s a worsening problem that’s now drawing major mainstream media attention. The New York Timesrecently devoted a major article on how “A.I. Is Getting More Powerful, but Its Hallucinations Are Getting Worse.” The newest bots from companies like OpenAI, Google, and DeepSeek are better at math, but their ability to generate factually accurate text when creating content has gotten worse.

OpenAI’s new and most powerful system, o3, and its o4-mini models have higher hallucination rates than previous models. According to the Times article, o3 hallucinated 33% of the time when running the company’s benchmark test, which involves answering questions about public figures; this is more than 50% higher than the hallucination rate of o1. Even worse was o4-mini’s 48% hallucination rate.

Surprisingly, Anthropic’s chief executive officer recently claimed that “AI models probably hallucinate less than humans” (TechCrunch). However, the company recently was forced to apologize in court after their chatbot Claude hallucinated and generated an inaccurate title and authors in a court filing. In another recent episode, the AI bot that handles tech support for Cursor, a tool for computer programmers, announced a policy change that did not exist. The result? Customer complaints and cancelled contracts.

These latest incidents build on other notable AI hallucinations that have included ChatGPT citing a half dozen court cases that didn’t exist in a 10-page legal brief that an attorney submitted to a Manhattan federal judge. This lawyer then asked a judge for leniency after the discovery. I could go on and on.

What Exactly Are Hallucinations Anyway

Hallucinations refer to AI-generated text that is completely made up or is factually incorrect. It’s obviously a major problem that can have major consequences for companies if left unchecked. That’s why most AI tools now warn users to verify AI-generated answers, even though these disclaimers are usually in smaller print and far less prominent than the AI-generated text. This is a warning that should not be overlooked.

This problem is thought to arise because of gaps in the training data and the large language models that AI is trained on, according to TechCrunch. This appears especially difficult to resolve for the foundation models of AI for a simple reason. There is not enough data in existence to train AI models to comprehensively resolve all the questions we could possibly ask. 

“Despite our best efforts, they will always hallucinate,” said Amr Awadallah, chief executive of Vectara, an AI tools startup, and former Google executive (New York Times).

With this issue and others, like poor writing quality and a lack of personalization, using highly skilled human AI editors to review and correct AI-generated content is no longer an option for companies. Today, they are a mission-critical part of content production from start to finish – from content strategy through content creation to final publication.

The Role of Human AI Editors 

Expert human AI editors bring years of experience as journalists and writers at public relations, marketing, and advertising agencies to their role as AI content editors. It’s a role that demands experts who can comfortably wear a number of different hats to ensure that a company’s target audience sees the highest-quality, most relevant content. They bring more than their grammar skills to the table when reviewing, editing, and finalizing AI-generated text. They are:

• Expert writers that can rewrite hallucinations to turn them into factually correct statements or edit them out of the text entirely
• Fact-checkers that review and verify every source, claim, and statistic from authoritative sources
• Subject matter experts that can identifying subtle errors and glaring omissions in machine-generated text
• Guardians of brand voice to ensure consistency in tone, perspective, and key messaging
• Ethical editors that notice, react, and correct content that could mislead, offend, or misinform
• Empathetic writers that can humanize AI-generated content by injecting context, tone, and a personal point to view to make the content human-centered

The result of having highly skilled AI human editors drive the content review process is smart and better content. It is content that is factually correct, sounds more natural, reads better, and presents a personal point of view that properly reflects your brand. It is content that can be published confidently without the fear of lawsuits or reputational damage. 

Want to ensure your AI-generated content is accurate, trustworthy, and brand-safe? Let’s talk. I help brands and companies integrate expert editing into their AI workflows, turning fast content into smart content. Contact Randy Savicky at randy@writingforhumans.co or call (203) 571-8151.

###