"Free AI" What to Check Before You Paste Data
Why AI chat can be free, what happens to what you type, and the short checklist worth running before anything from work goes into a chat box.
Why AI chat can be free, what happens to what you type, and the short checklist worth running before anything from work goes into a chat box.
When owning your hardware beats renting a virtual machine, when it does not, and the questions to ask a colocation facility before you commit to anything.
Object storage explained: how it differs from a server disk, what it is good and bad at, and the signs that tell you it is time to use a bucket.
Self-hosting an open-weight model or calling a hosted API? The drivers on each side, the break-even point, and the hidden costs.
A walkthrough of how to build semantic search with embeddings using RAG AI to index and chat with your docs, and when keyword search still wins.
How to turn scanned images and PDFs into typed fields your app can store, using local OCR and a language model, with the validation that keeps poor quality out.
Companies that offer free AI chat can be legitimate. Inference got cheap enough that giving away a chat window is a reasonable thing for a company to do, the same way free email was reasonable. The question worth asking is not whether free is a trick. It is what happens to the text you type, because that answer varies enormously between services, between plans of the same service, and between this year and last.
For public information, drafting, brainstorming, code you would post on a forum, and anything you would be comfortable reading aloud, free chat is fine. For customer records, credentials, unreleased plans, contract terms, health or payment data, and source code you are contractually obliged to protect, check three specific things first. They are listed below and they take about five minutes to look up.
Stanford's 2025 AI Index reports that "the inference cost for a system performing at the level of GPT-3.5 dropped over 280-fold between November 2022 and October 2024" (Stanford HAI). A conversation that cost meaningful money to serve in 2022 now costs a fraction of a cent. At that price, a free tier is a marketing expense, not a subsidy anyone has to justify quarterly.
There are four common reasons a chat is free, and they are not equally comfortable:
You can usually tell which one you are dealing with in a few minutes of reading, and the point of this article is that it is worth those minutes rather than assuming.
Retention: how long your conversations are stored, and whether you can delete them. Training: whether your text is used to improve models, and whether that is on by default. Review: whether humans at the company can read conversations, and under what circumstances.
Some detail on each, because the differences are not intuitive.
Retention is rarely zero. Most services keep conversations to provide history, and many keep them for a period after deletion for abuse investigation and legal obligations. A stated retention window of 30 days means what it says. "We may retain data as required" means you do not know.
Training is where free and paid tiers most often diverge. It is common for a consumer free tier to use conversations for training by default while the paid API does not, on the same service, with the same model. If you are used to one and start using the other, your assumptions may not travel with you. There is often an opt-out, and it is often not on the first settings page.
Human review exists at nearly every service, usually scoped to safety and abuse work or to conversations you flag. This is normal and it is also the thing people are most surprised by, because "the AI reads it" and "a person might read it" feel different even when the policy has always said both.
One structural point that outranks all three: where the inference runs determines whose laws apply to it. A conversation processed in another jurisdiction sits under that jurisdiction's rules regardless of what the interface says. If that matters for your work, data residency versus data sovereignty covers the distinction properly.
No, and the measurements are unflattering. LayerX's Enterprise AI and SaaS Data Security Report 2025 found that among enterprise employees using generative AI tools, "77 percent of these AI users have been copying and pasting data into their chatbot queries," with "82 percent of pastes coming from unmanaged personal accounts" and roughly 22 percent of paste operations involving personal or payment data (The Register).
The problem is not mainly that people use AI chat. It is that most of the pasting happens through personal accounts, outside whatever agreement the employer negotiated, under the consumer terms rather than the enterprise ones. Whatever data protections the company bought, they do not apply to a browser tab someone logged into with a personal address.
LayerX collected this through a browser extension, so it captures web-based chat and not API usage. That makes it a good measure of exactly the behavior discussed here and a poor measure of overall AI use.
Two questions worth adding if the work is regulated or contractual: does the provider offer a data processing agreement, and does the free tier fall under it? The answer to the second is frequently no, and that is the whole reason the paid tier exists.
Draw the line where you would draw it for a public forum post, then move it slightly stricter for anything belonging to someone else. The test that holds up: if this appeared in a search result with your employer's name attached, would that be a problem? If yes, it does not go in a free chat box.
Fine in nearly any free chat:
Check first, or do not paste at all:
The most common way people cross this line is not carelessness with documents, it is pasting a stack trace or a config file to get help with a bug. Those contain hostnames, internal paths, and sometimes credentials. Redact before pasting, or paste the error and describe the context in your own words.
A checklist does not make free AI chat appropriate for regulated data. If you handle health records, payment card data, or anything under a contractual data protection obligation, the answer is a service with a signed agreement covering exactly that use, not a free tier with good policies. Good policies are not the same as a contractual commitment you could enforce.
The other limitation is that policies change. The terms you read today can be revised, and services get acquired. If your process depends on a specific data handling promise, check it periodically rather than once. This is an argument for writing the answers down with a date on them.
And a note against overcorrecting. Banning AI chat outright reliably produces the personal-account pattern the LayerX numbers describe: people use it anyway, on their own logins, with no visibility for anyone. A sanctioned option with a clear line about what goes in it beats a prohibition that gets routed around.
Carpathian AI includes a chat that is free with no account, no message cap, no token metering, and no card on file. It runs on our own compute in the United States, the same infrastructure behind our model hosting, so prompts are not handed to a third-party API.
If you need AI built into a product with your data handling requirements designed in from the start, that is what our AI application development work covers, and the models can run on infrastructure we host so your data does not leave your control. When you outgrow the chat, the same platform exposes an OpenAI-compatible API, so the code you wrote against it keeps working.
