BY NATHAN GUMA
AN analysis of generative AI models has shown that most models are still lagging in updating their databases, which experts say may worsen hallucinations, while hindering access to credible information.
This comes at a time when more people are becoming increasingly reliant on generative AI for daily tasks.
However, experts say the delayed updates worsen hallucinations, which see AI models making up responses, at times without warning.
AI hallucination refers to the phenomenon where a generative AI model produces outputs that are inaccurate, nonsensical, or entirely fabricated, despite appearing confident and coherent.
It also occurs when an AI ‘makes things up’, and these errors can range from minor factual mistakes to completely invented scenarios.
For instance, in July, a blunder in the legal fraternity underscored the need for ethical AI use, with prominent lawyer Professor Welshman Ncube apologising to Zimbabwe’s Supreme Court after citing a dozen AI-generated fictitious cases in a mining dispute.
As part of the analysis by The NewsHawks, simple prompts asking when the system was last updated were fed into five widely used AI models: Grok by xAI, Gemini by Google, DeepSeek by Chinese hedge fund High-Flyer, GPT-4o, and Meta by Meta AI.
DeepSeek
In response, DeepSeek said it was last updated in 2024.
“I was last updated in June 2024, so I don’t have information or data from July 2025 or beyond. My knowledge is current up until mid-2024, and I can’t predict future events or access real-time data,” says DeepSeek.
“If you’re looking for hypothetical scenarios, projections, or general advice about future planning, I’d be happy to help based on available information up to my knowledge cutoff! Let me know how I can assist.”
Meta
Meta AI’s Meta, running on WhatsApp, said that while its last update was in July 2024, it has been fine-tuned to provide information based on data available up to 10 July 2025.
“If you need information on a specific topic, feel free to ask, and I’ll do my best to provide the most up-to-date information available. Keep in mind that some sources may not be updated in real time, but I’ll strive to give you the most accurate and helpful response,” it replied.
GPT-4o Mini
GPT-4o by OpenAI responded that it was last updated in October 2021.
“However, I have access to information up to that date, and I can provide general insights or historical context. If you have specific questions, feel free to ask!” says GPT-4o Mini.
However, later versions of ChatGPT are regularly updated, and several models are available depending on whether you’re using the free or paid versions.
Gemini
Google’s Gemini said that its knowledge base is continuously updated.
“While the core training data I was initially trained on has a cutoff, I also have access to real-time information through Google Search and other tools, which means I can access and process information up to the current date,” responded Gemini.
“My models, like Gemini 2.5 Pro and Gemini 2.5 Flash, receive regular updates to improve their capabilities and performance. For instance, new stable versions of Gemini 2.5 Pro and Flash were released in June 2025, with further updates and new experimental models becoming available as recently as July 2025.”
Grok
Grok by xAI says it is constantly updated and therefore has the most accurate data.
“My knowledge is continuously updated, with the latest refresh in July 2025, so I’ve got a ton of fresh info to share. What’s on your mind?” it responded.
Dangers of outdated databases
According to AI expert Justice Mukaro, AI models are only as good as their databases, and failure to update them frequently makes them prone to misinformation.
“Top AI models have updated their databases… some last updated in late 2024. However, there is a misconception because older models like ChatGPT 3.5 are still accessible for free, and these were last updated in 2021,” said Mukaro, founder of Strateji, a start-up focused on building Afrocentric datasets for AI.
“Generally speaking, an AI is only as intelligent as the databases from which it gets information… not regularly updating them means the AIs end up hallucinating and creating serious misinformation, which might even be harmful in some instances.
“The delay in updating databases is largely influenced by the costs of gathering and cleaning new data before feeding it to the AIs. It takes a lot of time and money to get the data and to train the AI.”
Another expert, Munyaradzi Togarepi, says that AI models are usually trained on outdated public data, thereby often producing inaccurate or hallucinated responses.
“All AI models are generally trained using publicly available data. What this means is that all privately held data, which is usually accurate, is not used when training these models. For me, that’s the first critical chokepoint of any inaccuracies that may arise when using these models,” says Togarepi, chief executive of Procode Platforms, a company that builds AI platforms for enterprises in Africa.
“And adding on to the fact that these models have a cutoff date of two years ago just makes the problem even more pertinent. And then there are the hallucinations, the models are trained to return a perfect answer as much as possible, and in most cases, that means saying falsehoods that look credible.
These AI models can not cite anything that happened six months ago, and in most cases, the everyday user is interested in knowing recent information and knowledge. People are likely to get information that is not credible, and they might not even notice it.
Togarepi also says databases are rarely updated because it is expensive and time-consuming.
“Remember, all the raw information has to be converted into structured data, which can be fed into the models. It’s a process that requires time and resources. Another factor is capability; most AI companies are racing to create the best capable AI in terms of reasoning and other variables,” he says.
“This means they have to shift focus from the knowledge base itself and focus on the core algorithms for the reasoning. However, some AI platforms, such as OpenAI’s ChatGPT and Google’s Gemini, now allow their models to perform a real-time web search to look for any recent information available on the Internet. This enhances the data used in generating a result.”
