Welcome to The AI Brief.

Every week, you get clear, honest breakdown of what’s really happening in AI. No recycled press releases pretending to be insight. Just the developments that actually matter.

You close the tab. The conversation disappears from your screen. As far as the interface is concerned, it is over.

What actually happened is that a copy of that conversation is sitting on a server attached to your account, readable by company employees under certain conditions, subpoenable by governments, and in some cases being processed into signals that will shape how the next version of the model responds to someone else.

None of that happens dramatically. It happens quietly, in ways that are technically disclosed in documents designed not to be read.

This newsletter walks through what actually happens to your conversations from the moment you close the tab to the moment they are actually gone.

Let's get into it.

Most AI creative tools fall short for one simple reason. You can generate tons of ads, but they aren’t up to par.

Refining copy, adjusting layouts, or nudging a CTA into place shouldn’t require rewriting prompts over and over. It slows teams down and breaks the creative process.

With Hightouch Ad Studio, AI gets you 90% of the way there. For the final 10%, use a built-in editor to quickly refine copy and design, or export directly to Figma for seamless collaboration with your design team.

Move faster without losing control. Every ad, exactly how you want it.

See how Ad Studio works

When you close the tab, nothing is deleted. That is the first thing to understand. The conversation is sitting on a server attached to your account, in the same state it was in while you were typing, accessible to the company running the product. This is true for ChatGPT, Claude, and Gemini, and it is true whether or not you have opted out of model training.

The next stage varies by platform but follows a similar pattern. Conversations are processed after the fact: tagged, logged with metadata including timestamps and model version, and in some cases flagged by automated systems for safety review. If something in the conversation matches a policy trigger, it is more likely to be read by a human employee. If nothing triggers, it may still be reviewed as part of routine quality sampling. The selection criteria for that sampling are not publicly disclosed.

This is the part that surprises people most: human review is not reserved for flagged conversations. All three major platforms state in their documentation that conversations may be reviewed by employees for safety, quality, and research purposes. The practice is real and ongoing at all three companies. How frequently it happens to any given conversation is not information they publish.

If you have not opted out of model training, the conversation enters a pipeline where it contributes training signal for future model versions. This does not mean your exact words appear in the model. It means your conversation contributes to statistical patterns that influence how the model responds in future. Your specific sentences are not stored in the weights. Their influence on model behaviour is. That distinction is real and is almost never explained clearly by the companies that benefit from the ambiguity.

When you delete a conversation from your account, it disappears from what you can see. It does not disappear from backup systems on the same schedule. Companies maintain backups for disaster recovery and legal compliance, and a deleted conversation can persist in those backups for weeks or months after you request removal. The company has technically honored your request. The data is not technically gone.

Governments can request your conversations with valid legal process. A subpoena, a court order, or a national security letter can require a company to produce your conversation data regardless of your privacy settings. All three major AI platforms publish transparency reports showing how many such requests they receive and how many they comply with. The numbers are not zero. They are not published prominently.

The final stage is the one most people have not considered. After a conversation is deleted from active systems and eventually from backups, the training signal it generated persists in the model weights with no expiration. Your words are gone. The patterns they contributed to are not. This is how machine learning works, not a loophole. But it is worth understanding before you decide how much of your thinking to put into these tools.

Companies describe the full lifecycle of your conversation data using terms like "processed for safety," "retained for quality assurance," and "may be used to improve your experience." The same lifecycle described plainly is: stored, possibly read by an employee, trained on, backed up, kept after you delete it, and available to governments on request. Both descriptions are accurate. Only one of them fits on a settings page.

Every company running an AI product has legitimate reasons for the practices described in this issue. Safety monitoring requires access to conversations. Legal compliance requires retention. Model improvement requires training data. None of those are fabricated justifications.

The problem is not that these practices exist. It is that the gap between what the interface communicates and what is actually happening is large, consistent across every major platform, and not an accident. The delete button looks like it ends something. The privacy toggle looks like it covers more than it does. The temporary chat mode sounds like it creates real privacy. None of them work the way most users assume, and the companies designing these interfaces are aware of that gap.

People share things in AI conversations they would not share in email, in a document, or in a public post. The chat interface is intimate in a way that the data practices are not. Understanding what actually happens when you close the tab does not require you to stop using these tools. It just requires you to use them with an accurate picture of where your conversations actually go.

Two things worth doing this week. Search the name of whichever AI platform you use most alongside the words "transparency report" and spend five minutes reading the most recent one. The numbers are more interesting than you expect. Second, think about whether the things you regularly put into AI conversations are things you would be comfortable with a stranger at the company reading. Not because they will, but because the answer tells you something useful about your current habits.

Reply and tell me if something in this issue changed how you think about these tools.