By Anton Christodoulou, AI Specialist at allpoints
Are You Training the Model With Your Data?
While I always prefer to focus on the positive benefits of new technologies, especially when writing about them, I am also a passionate privacy and security advocate. I have been thinking a lot about the emerging patterns we’re seeing around AI platforms, and the challenges of navigating the platform “tiers.”
Most people I speak to now use platforms such as ChatGPT, Gemini, and Claude, to make their workload easier. But there’s a hidden detail many people miss, or underestimate. If you’re on a free or lower-tier plan, you’re not just using the AI model, you’re actually training it. It is why the free tier exists, as this is a great way to continue feeding and improving the models. In some cases this could prove to be particularly hazardous.
Why This Matters Right Now
Many organisations are already moving beyond AI in a chat window. It is fast becoming the orchestration layer, sometimes referred to as an “Agentic Ecosystem.” The benefits are immense, as the models are so powerful now, they can perform multiple tasks reliably and with limited oversight. While there are inherent risks around using “probability engines” to run your business (for a future discussion!), this also means they have unprecedented levels of access. There are now multiple new ways in which you can quickly and easily install AI directly on your machine – as an app, browser plugin, AI browser, even as a local command line tool. I use a combination of these in parallel (except the dedicated AI browser), and use a number of adjacent security and privacy tools to mitigate some of the risks.
The reality is that free-to-use tiers across all major LLMs (so called, Large Language Models) use your data and interactions to learn and improve. If you’re inputting sensitive budgets, client strategies, internal or personal data, you’re essentially feeding the model the information it needs to learn and train to get smarter. Your data becomes part of the model. This is why you can create images in the perfect style of a famous artist, and why you can ask them to build a go-to-market strategy and pricing model. On an enterprise tier, the benefits are obvious, as it can then use your treasure trove of internal data, without training the public model. You can even train your own internal models.
The models also use a version of the famous “summarise” feature to distill these interactions into a summary of you. When used ethically, this “AI memory” enables the models to provide much more personalised results. One positive step is that most models now, at last, enable you to at least access and delete your AI memory.
While enterprise models offer “walled gardens” and legal safeguards that keep your data private, not all platforms are created equal. We’re seeing a trend where Anthropic’s Claude is establishing itself as being more open and ethical, while all of the major players offer better protection on paid tiers, and the best protection on business and enterprise tiers – essential to building trust and meeting governance requirements.
I asked Claude how it differed from other companies in this regard. It started with “I have an obvious conflict of interest, so take it with skepticism.” Followed by stating its strengths, “Safety-focused research, relative transparency about methods, and designed to be honest rather than compliant.” In keeping with this, it was honest about its limitations, “Anthropic is still a commercial business, no AI company has solved alignment, Claude can still make mistakes, and privacy concerns apply industry-wide.” and finished with “You shouldn’t fully trust any AI system or company right now – including Anthropic. The technology is too new and the alignment problem isn’t solved. Healthy skepticism toward all of us is warranted.“ There are real solutions to this challenge, however, they are not yet easy to implement (For a much deeper insight into this, and the history of the web, I can highly recommend Sir Tim Berners Lee’s book “This Is For Everyone”.)
However, OpenAI has just taken the extraordinary step of introducing ad-based revenue on the free tiers. While they claim that this will not impact the responses you receive, OpenAI can now freely share deep knowledge of you and your interests to advertisers, which they already openly admit to harvesting, with limited controls. This move is following the old social media playbook of exploiting you and your interests in any way possible to make more money. That was in an era when most people were unaware of the risks, and the risks today are so much bigger and more complex. Hopefully the other large players do not follow OpenAI’s lead, and use that as a critical differentiator.
The Competitive Risk of Ignoring AI Privacy
During a time of unknowns, using AI without a privacy strategy isn’t just risky – it’s a competitive liability. However, analysis paralysis, not using AI at all or ineffectively due to fear of the risks, presents an even greater risk.
The Takeaway
If your business is using AI, you and your teams must be hyper-aware of the platform tier you’re on. Other businesses can and do use AI to find information on competitors, uncovering budget spends and strategic pivots that can “leak” into the models via lower-tier or free versions. While this “search and source” capability is a massive benefit for some, it’s a significant detriment to others who haven’t secured their data, and is the tip of the spear.
We started allpointsAI to specialise in navigating these complexities, and create and execute AI strategies that help companies reap the benefits, safely, and effectively. We are also passionate privacy advocates, always seeking to identify and mitigate the inherent risks when integrating this new and exciting technology.