What a Free AI Tier Costs You

Running these models is expensive. A free tier is therefore either a loss leader, a data arrangement, or a degraded product. Which one determines what you should use it for.

The funding models

Loss leader. The provider absorbs the cost to acquire users who will convert. Usually the best free tiers, usually rate-limited, usually the previous generation of model.

Training on your inputs. Your conversations improve the product. Frequently disclosed in the terms and rarely read. This is the standard arrangement for consumer free tiers.

Advertising or lead generation. Less common in assistants, common in the smaller tools built on top.

Genuinely open. Local models you run yourself. The cost is your hardware.

Not funded at all. A product burning investment with no plan. These disappear, taking your history with them.

What "free" typically means in practice

A weaker model. Free tiers commonly serve a smaller or older model, sometimes without saying so.

Silent downgrade at the limit. Rather than blocking, some products switch to a cheaper model when you exceed an allowance. The interface does not change and the answers get worse. This is worth testing for.

Fewer features. No file upload, no code execution, shorter context, no search.

No data controls. The ability to opt out of training is frequently a paid feature.

Lower priority. Slower at busy times.

The data question, specifically

For most consumer free tiers, the default is that your conversations may be used to improve the service, which in practice means human review of samples and inclusion in future training.

What this means concretely: do not paste anything you would not be comfortable with a contractor at that company reading. Client material, personal medical details, confidential business information, credentials, unpublished work.

This is not a scandal. It is the stated arrangement. The problem is that almost nobody reads it, and people paste extremely sensitive material into free chat interfaces daily.

Check three things for any tool you use regularly: is training on inputs the default, can you turn it off, and does the setting apply retroactively or only going forward.

When free is entirely sufficient

Learning and experimenting.

Non-sensitive drafting — a blog post, a cover letter, a birthday message.

Explanation and study. Understanding a concept, working through a problem.

Occasional use below the rate limit.

Cross-checking a paid tool. A second opinion from a different provider, free tier, is a genuinely useful workflow.

Most people's use is entirely covered by a good free tier, and paying is a convenience decision rather than a capability one.

When to pay

Confidential material. The single strongest reason. Paid tiers generally come with an opt-out or a contractual commitment.

Volume. Hitting limits daily.

Specific features you need — long context, code execution, document handling, the strongest model.

Reliability. Priority access when you depend on it for work.

Integration into the tools you actually work in.

The tier that is worse than nothing

Watch for products where the free tier exists to harvest rather than to demonstrate.

Signs: a privacy policy permitting broad sharing with unnamed third parties, a requirement to connect accounts or grant permissions disproportionate to the function, aggressive permission requests on mobile, or an entity that is difficult to identify.

The specific case worth naming is browser extensions and mobile apps offering free AI features. Many request permission to read everything you do, and some are straightforwardly data collection operations with a chatbot attached. The permission request tells you more than the description.

The practical position

Use a good free tier from a company you can identify, for anything not confidential. Pay for one tool where you need the data controls, and use it for anything sensitive. Never paste confidential material into a free tier from a company you had not heard of a month ago.

That covers almost everyone, and it costs either nothing or the price of one subscription.

What to check at signup

Two minutes at the start, once, that prevents most of the problems described above.

Open the data controls immediately. Before the first real conversation. The setting for training on inputs is usually there and usually on.

Check whether disabling it is retroactive. It generally is not, which means anything you sent while exploring stays in scope.

Check the retention period for conversations, including deleted ones.

Check whether history sync is on and where it goes.

Look for a business or team tier if you will use it for work, and compare the data terms rather than the features.

Note the company. If you cannot find who operates the service, where they are and what their business is, that is the finding.

Then decide what this tool is for. Sensitive or not. Making that decision once, at signup, is far more reliable than deciding message by message under time pressure.