Written by: Jason Crowley, Portfolio Manager, Veilant

Its uses and benefits are undeniable. Whether you’re trying to learn something new, develop an application, double-check your code, or confide in a new electronic friend, it’s hard to deny the usefulness of these tools. They aren’t going anywhere.

That makes it important to understand what happens to the information we share with them. As with any online activity which we provide data to, once that information is passed, it ceases to be ours. We essentially no longer own it.

When you enter information into an online AI service, you’re giving another organization access to it under that service’s terms, conditions, and settings. Your information becomes part of the larger data-sphere to be bought, sold, stored, and, in this case, used to further train the tool you’re working with. A single prompt may seem benign on the surface, but, when aggregated with other data you provide over months and years of use, the prognosis becomes darker. Months of prompts, files, and connected accounts can reveal considerably more about your work, interests, and life.

So how do we take advantage of AI without sharing more than we intend, and while still protecting ourselves from hacking, leakages, or, more simply, the AdTech economy?

First and foremost, if you’re using something that you don’t own or control, you’re already starting from a losing position. The best, most secure way to use AI is to self-host a model. There are several to choose from and some can be run on systems as simple as Raspberry Pi.

Other tools that can help you do that include:

There are trade-offs to this approach. Many tools that are run locally will be much less robust than online options like ChatGPT or Anthropic. These tools will likely be Language Learning Models (LLM) only and will not have some capabilities that you are familiar with, like image generation or graphic design, or it will take additional setup and computing power. Additionally, because these models are downloaded and used offline, their data will be confined to a snapshot in time and may not have all the answers you’re looking for. A local model won’t know about recent events unless you update it or provide current information.

The experience depends on your device and the model you choose. Check whether the tool is using a local model or an optional cloud service before you give it sensitive information. If the model and its supporting tools are running offline, your prompts can stay on your device.

For many people, setting up a local model will be more effort than they want to spend. You can still use familiar online tools with increased vigilance.

Some suggested steps you can take are: 1) Review the tool’s current privacy settings. Given that each AI tool may be different, and their ever-increasing improvements, one of the most effective methods is to ask the AI itself. You can ask an AI tool where to find its privacy settings, but verify the answer against the provider’s own documentation. Settings change. An AI tool’s explanation may be out of date.

2) Limit the information and data you share as much as possible. Instead of a person’s name, use [Employee]. Instead of a business name, use [Company]. Remove account numbers, customer details, private documents, and other identifiers that the task doesn’t require. Keep the details that matter to the answer: the timeline, constraints, or type of problem you’re trying to solve. If you’re using AI for work, follow your organization’s rules about what can be shared.

3) Check what happens to saved conversations, memories, uploaded files, and connected apps, and consider temporary chats. These controls vary by product and account, and a temporary chat does not necessarily mean immediate deletion.

And lastly, turn off model improvement or training where the service allows it. AI can help us think, build, and work faster. None of that requires giving every tool an unrestricted view into our lives.

Before you enter something into a prompt, ask two questions:

If the answer to either is no, change the prompt, choose a local model, or leave the information out.

Veilant delivers mission-critical Digital Force Protection, Signature Management, Secure Communications, and Managed Attribution for national security customers. We combine scalable technology with tech-enabled training to help organizations understand data exposure, control their digital signatures, and operate effectively in complex, contested information environments.