A ChatGPT descendant that watches your screen, records every meeting and call, and holds perfect context of your entire life. Six months out, per Sam Altman.

Most of the reaction went in one direction: is that creepy? Fair question, already settled by the phone in your pocket.

Here is the one that lands on your desk instead. When that memory shows up inside your company, whose signature is on the retention policy?

Persistent memory is a system of record that arrived without an owner

Your company has rules for email. It has rules for the CRM. Someone once decided how long a signed contract lives and who can pull it from the archive.

An AI that records meetings and reads screens is another system of record. Timestamped business information, attributable to named people. The difference is that nobody approved it as one. It arrived as a feature, got switched on from a settings toggle, and skipped the conversation the CRM went through when it was purchased.

That gap is the story. Not whether the capability lands in six months or eighteen.

We made the general version of this argument in the model is a commodity, governance is the moat. This is the sharpest instance of it. Anyone can buy the model with a credit card. What separates you is having decided, in advance, what it does with what it sees.

For five months, deleting did not delete

This is not hypothetical, and it is still being decided.

On May 13, 2025, in the New York Times copyright litigation, a court ordered OpenAI to preserve all output log data that would otherwise be deleted, regardless of user deletion requests or privacy regulations. Judge Sidney Stein affirmed it on June 26 after OpenAI appealed on user-privacy grounds.

Throughout that window the delete button stayed in the interface and kept working. It just stopped deleting.

Judge Ona T. Wang lifted the going-forward obligation in an order filed October 9, 2025, covering data past September 26. Engadget covered the exceptions: logs already captured under the prior order remain accessible, and anything tied to accounts the Times flagged is still held. Otherwise OpenAI returned to removing deleted conversations and Temporary Chats within 30 days.

Then it moved again. Magistrate Judge Wang ruled in November 2025, and Judge Stein affirmed on January 5, 2026, that OpenAI must produce 20 million de-identified conversation logs to the plaintiffs. OpenAI objected on user-privacy grounds and lost. De-identification strips the personal identifiers, not what people typed.

The lesson is not that OpenAI is careless. It is this: your AI vendor’s retention policy can change through a decision you were not part of and will not be consulted on. Fifteen months of rulings so far, decided by neither the vendor nor the customer.

If the vendor’s promise is your only layer of protection, your data policy lives in a courtroom in another country.

What he actually said, in full

Worth quoting properly, because the line is circulating in fragments.

Altman’s words were that “in the next 6 months, we’re close to a world where a descendant of ChatGPT can watch your screen, record every meeting and call, and have perfect context of your whole life”, and that we are one model generation away from it being genuinely useful.

Two details in there deserve more attention than the headline got.

First, “perfect context.” Not better context. Not more context. Perfect, which in practice means complete, which in practice means unfiltered at capture time. A system that decides later what mattered had to record everything first.

Second, the clock. Six months is not enough time for a committee to draft a policy, route it through legal and communicate it to staff. Either you already have one, or you will be writing it with the product switched on.

What it sees inside a marketing team

This stops being an infrastructure topic the moment you run marketing.

Look at what is open on your team’s screens on an ordinary Tuesday. The full CRM, including notes salespeople typed assuming nobody else would ever read them. A recording of a call with an account that has not signed yet. Campaign creative shipping in three weeks that a competitor would pay real money to see today. A budget projection that never left the department. Data an agency handed you under an NDA you personally signed.

An always-on AI with screen access sees all of it. Not selectively. All of it.

Marketing also tends to adopt first. It carries volume pressure, sits under fewer internal controls than finance, and tries new tools without routing them through IT. That makes it the best lab in the company and the least defended entry point.

Then there is the layer that quietly voids most written policies: the personal account.

Business-tier and consumer-tier retention terms are not the same thing. The business tier comes with a contract, admin controls and, at best, an audit log. The consumer tier comes with terms of service nobody read and a default setting.

Real adoption rarely starts on the business tier. It starts when somebody on the team tries the tool on a Sunday from their own laptop, watches it save two hours, and keeps using it on Monday with work material. By the time that surfaces in a formal conversation, six months of recorded meetings are sitting in a workspace your company does not administer and cannot request deletion from, because it does not legally own it.

Banning it does not fix this. Issuing the business tier before people go looking for it does, which makes it a budget decision taken early rather than a memo sent late.

The third-party data point is the one that gets missed. When an agency or a client shares information under confidentiality, the commitment you signed does not distinguish between a human reading it and a memory-equipped system ingesting it. If that material lands in a store whose deletion policy you do not control, the problem is not technical. It is contractual, and your name is on it.

We wrote about the scale this can reach in the Mercor breach and biometric data you cannot rotate. A leaked password gets changed. A recording of your client’s call does not.

Four answers to get before you switch it on

When we work with a team that wants this capability, we do not start with the tool. We start with four answers that are almost always missing.

Who owns it. A name and a title, not a department. Someone who responds when a client asks what was kept from their call. If the owner is “IT” or “legal” in the abstract, there is no owner.

What is out of scope. Listing what the AI must never capture is faster and more honest than trying to classify everything it may. Compensation discussions, third-party material under NDA, anything with identifiable health or financial data.

How long, per record type. An internal meeting transcript and a recorded client call have no reason to share a retention window. Assigning one number to everything is the most common way to end up storing the sensitive material by default.

How deletion actually works. Ask the vendor in writing: what goes, what survives in backups, on what timeline, and under what legal conditions they may retain anyway. Keep the answer. It is the one you will need the day a client asks you the same question.

That work is the same shape as what we described in marketing does not need more AI, it needs less sediment: know what is actually in the process before automating it. Same discipline, pointed at what the system stores rather than what it produces.

None of the four answers depends on whether Altman is right about six months. They need one afternoon and one person with signing authority.

Persistent memory is coming to your team, and it will probably arrive through someone who just wanted better meeting notes. Whether it ends up an advantage or a liability gets decided before that, not after.

Name the owner before you switch it on