Typing is sharing
When you type a question into an AI tool or upload a file, it is easy to feel as if you are talking only to your own screen. In fact, your input is sent to servers run by the provider, and what happens next depends on how the service is built and what its terms say. Those terms differ a great deal from one service to another, and sometimes between plans of the same service, so no general statement can describe them all.
This article explains the questions worth asking and where to find answers. It is general information, not a review of any product, and it does not replace reading the actual terms or taking advice for sensitive data.
The journey of an input
A typical path looks something like this, though details vary.
- Transmission. Your text or file is sent over the internet to the provider, normally encrypted in transit.
- Processing. The system generates a response. Some providers use other companies’ infrastructure or models for parts of this, which means more than one party may handle the data.
- Storage. The input, the output and related information such as timestamps, account details and device data may be saved, for your conversation history, for troubleshooting or for other reasons.
- Review. Some providers allow staff or contractors to look at samples of conversations to check quality, detect misuse or improve the product, sometimes with safeguards.
- Reuse. Inputs may be used to improve or train models, unless you opt out or your plan excludes it.
- Sharing. Data may be disclosed to service providers, affiliates, or authorities when legally required, and in some cases in a business transfer.
- Deletion. Data may be deleted on request, after a set period, or kept longer for legal or security reasons.
The mention of training connects to how systems are built: the way training data shapes a model explains why providers want more examples, and why some inputs end up as part of that pool unless they are excluded.
Questions to ask
When you evaluate a tool, look for answers to these questions in the terms of service, privacy policy, any data processing agreement and the settings pages.
- Is my input stored, and for how long? Look for specific retention periods. “As long as necessary” is not an answer.
- Is it used to train or improve models? Is that the default, and can it be switched off? Does the setting apply to everyone in the organization?
- Who can see it? Are there human reviewers? Under what conditions?
- Who else receives it? Ask for a list or category of subprocessors and the countries involved.
- Can I delete it, and is the deletion real? Check whether deleting a conversation removes it from backups and logs, and how long that takes.
- What happens to temporary or “incognito” modes? They may still be retained for a period for safety reasons.
- What about uploaded files and connected apps? Integrations that access your email, drive or calendar may expose more than you realize.
- What security measures are in place? Ask about encryption, access controls and certifications, and request documentation instead of relying on slogans.
- What are the breach notification terms?
- Can the terms change? Many services reserve the right to update them, and you may be bound by default.
A prepared set of vendor questions can help you press for specifics.
Free, consumer and business plans
Differences between plan types are among the most important and most overlooked. Free or consumer versions often have broader rights for the provider to use inputs, while business or enterprise agreements may exclude training use, offer shorter retention and include contractual commitments. That is not guaranteed, so verify rather than assume. Using a personal account for work tasks can put work information under consumer terms that your organization never reviewed, which is how shadow AI creates risk.
Settings you control
Even within a plan, you may have options.
- Turning off chat history or model-improvement sharing where available
- Using organization-managed accounts with administrator controls
- Restricting integrations and plug-ins to approved ones
- Setting retention periods where the provider allows it
- Removing old conversations that contain sensitive material
Check the settings regularly, since defaults and menus change after updates.
Practical habits
The strongest protection is not to enter what you should not share. Leave out personal identifiers, confidential material and credentials unless the tool has been approved for that data class. The details on customer information are covered under protecting customer data, and the broader legal principles under data protection. A short note in your AI use policy listing approved tools and the kinds of data each may receive makes the rules usable.
Caveats
Reading terms takes effort and they can be hard to follow. If they are unclear, ask the vendor in writing and keep the answer. Be skeptical of marketing phrases such as “your data is safe” or “private by design”, which rarely settle the specific questions above. And remember that even well-run services can suffer incidents, so assume that anything entered could in some scenario be exposed, and decide with that in mind.