I remember being told that I couldn't go swimming or do hand-stands immediately after eating; that if I crossed my eyes too much that they would get stuck. As a child you tend to believe what your parents tell you, but it turns out these are myths. Your sphincter valve is very effective at keeping your dinner where it should be whatever your angle or activity, and it is impossible for your eyes to get stuck (or go square for that matter).
Parents are obviously concerned about their childrens' welfare, which is a good thing, but can also hold them back from the rewards of exploring, learning and having fun!
So it is with AI LLMs (Large Language Models) such as ChatGPT and Gemini. Leaders are concerned about adopting these tools for fear of exposing trade secrets and customer information. Being concerned about these things is of course, the right stance to take. Don't jump into the pond until you know what lies beneath.
But being fearful of things that are myths is not healthy. It will hold you back and you'll miss opportunities.
So, what are the key concerns and are they real?
There is a risk that information will be exposed or 'leak' into others conversations
In properly designed AI systems such as ChatGPT, individual conversations are isolated. The AI does not have access to other users’ chats when responding to you, and your information will not leak into other conversations. It is in the commerical interests of LLM providers to maintain these controls to build trust and adoption.
Furthermore, OpenAI and other major providers adhere to recognized security standards, including ISO 27001 and SOC 2, which involve regular independent audits of their security and operational practices. Most in-house environments will provide similar or in some cases stronger protection. However, as we've seen recently with some high-profile cases, in-house security can fall short. For additional control, users and enterprise clients can also configure strict data retention settings, including options to disable chat history or apply filtering to control what information is shared.
For highly sensitive data, where even brief exposure outside an internal network would present a material risk, LLM models can be deployed within your own firewalls ensuring that you have full control over data governance and access.
Everything typed into an AI chat is automatically used to train future versions of the AI model
Data usage policies vary by provider, but many AI companies allow users to opt out of having their data used for model training—particularly within enterprise and API offerings. Some providers do not use conversation data for training at all by default.
Large language models also handle information very differently from traditional systems. They do not store or recall conversations as discrete records. Instead, vast amounts of text data are processed to learn statistical relationships between words and concepts. The model’s responses are generated from these learned patterns rather than from stored memories of past interactions. As a result, tracing any specific output back to an original source is inherently complex and generally not possible.
When I upload a document to an AI assistant, that file is automatically added to the AI's training data and stored permanently
Documents uploaded to an LLM conversation are processed only to generate responses within that session and are retained temporarily in accordance with the provider’s data retention policy. They are not automatically added to the model’s training data. However, it is important to review each provider’s specific terms of use and to follow your organization’s data-sharing policies.
For enterprise environments, some offerings, such as Microsoft Copilot, can connect directly to internal document repositories (for example, SharePoint or OneDrive). In these cases, the data remains securely within the organisation’s own environment rather than being uploaded to the LLM provider’s infrastructure.
LLMs are prone to errors so I don't want them anywhere near my business process
While errors and 'hallucinations' are an inherent trait of generative AI, the commercial risk is low when systems are properly designed and governed. When AI tools are treated as assistants with a human in the loop rather than autonomous agents, the consequences can be easily avoided. In creative or exploratory contexts, the same generative behavior that produces occasional inaccuracies can also fuel innovation and uncover unexpected insights.
All in all, if a sensible measured approach is taken, AI can be used safely in many business contexts. Education is key. Which is why we focus on providing insights and information to our clients so that they can make informed choices. We help educate and inform across organisations, not just in the data science and innovation teams. If you'd like to know more about how we can help you get the best from your AI solutions, please get in touch.