ChatGPT Ads Hit $1 Billion. Your Assignment Still Asks for the Answer.

Everyone quoted the revenue number. The line that changes your assignment is in the last section.

On August 31, OpenAI published a post titled “A milestone in expanding access to AI.” The headline figure is that ChatGPT Ads has reached one billion dollars in annualized revenue run rate in, as the company puts it, “less than 200 days after launch.” Self service buying opened across India, Europe, the Middle East and North Africa the same day. The platform is live in over 40 countries. Tens of thousands of advertisers are using it.

None of that is the part written about your classroom.

What OpenAI actually said about where ChatGPT ads sit

Read the section headed “Where decisions take shape.” OpenAI describes the moment it is selling: “They explain what they want to accomplish, what matters to them, and the constraints they face. ChatGPT can help them frame the decision, establish criteria, understand alternatives, and move closer to taking action. Discovery, consideration, and decision-making come together in a single experience.”

Frame the decision. Establish criteria. Understand alternatives. Read that list again with a student in mind, because that list is not a shopping description. It is a description of the cognitive work most of us grade. It is what a comparative essay asks for. It is what a case analysis asks for. It is what a lab writeup asks for when it asks why this method and not that one.

OpenAI is not hiding this. The company is telling advertisers, accurately, that ChatGPT is where people form criteria. That is the product. And the reason the sentence matters to you is that your students are inside that same moment for a different reason, and nobody sent them a memo about it.

The protections are real, and they are dated

Give the company its due, because the argument is weaker if you skip this. The announcement states plainly: “Ads in ChatGPT are always clearly labeled and separate from ChatGPT’s answers. Advertising does not influence the answers ChatGPT provides.” It says advertisers do not get access to private conversations and that people can control how their ad experience is personalized.

Take all of that at face value. Today the label does the work. A student who is paying attention can see where the answer ends and the ad begins.

Now read the closing section, headed “What’s next.” One sentence: “We will also explore new ways for businesses to interact with consumers in more native ways in ChatGPT.”

Native is the operative word, and it is a term of art in advertising. Native means the commercial content stops looking like a separate block and starts looking like the surrounding experience. That is not a rumor and it is not a hostile reading. It is the company’s own stated direction, published on its own site, in the section about growth.

So the honest statement of the situation is this. The separation between the answer and the ad is currently a design choice, it is holding, and the roadmap says the next phase makes it less separate. You are not being asked to panic. You are being told, in advance, that a boundary your students are currently relying on is scheduled to move.

Why banning ChatGPT ads is the wrong response

The instinct is to ban the tool, or to detect it. We already covered why detection fails in Syllabus AI Policy: MIT Says It Should Name Your Own AI Use Too, and nothing about an advertising layer changes that arithmetic. A prohibition you cannot enforce is a prohibition that punishes the students who read the syllabus.

There is a second reason, and it is the one worth sitting with. Banning the tool assumes the problem is the tool. The problem is narrower than that and it is closer to home. The problem is that a large share of the assignments we hand out can be satisfied by a single confident answer, and a single confident answer is precisely the artifact that a commercially shaped system is best at producing.

An assignment that asks “which of these three approaches is best” accepts a recommendation. A recommendation is a product. It can be sponsored, or nudged, or simply drawn from whichever alternative had the most written about it, and the student turning it in cannot tell the difference, and neither can you, and no detector closes that gap.

The assignment is the control surface

Here is the shift, and it costs you one editing pass rather than a policy fight.

Stop asking for the answer. Ask for the decision path that produced it.

A commercially shaped answer is very good at arriving at a destination. It is much weaker at three specific things, and each of them is gradeable:

• The rejected alternative, with the reason. Not “I considered X.” Which one you did not choose, and the specific criterion that ruled it out. Sponsorship pushes toward a winner. It has no incentive to make the loser legible.
• The criteria, stated before the conclusion. Written first, in the student’s own words, tied to the constraints of this course and this case. Criteria written after the answer are a justification. Criteria written before it are reasoning.
• The point where the student changed their mind. One sentence naming what they believed at the start and what moved them. A generated answer has no history, so it cannot supply one honestly.

None of those three requires new technology, a detector, or a proctoring contract. They require a rewritten prompt and one added line in your rubric. That is a Wednesday afternoon, once, per assignment.

What this looks like in a real course

Take an assignment that already exists in your Course Record. A business course asks students to recommend a project management tool for a small firm. As written, that is a recommendation, and it is now a category with an advertising market attached to it.

Rewritten, it asks for four things. The constraints of the firm, stated first. Three candidates with the source of each. The one the student ruled out and the exact constraint that ruled it out. And a closing line naming what they thought before they started.

Same subject. Same reading. Same hour of student effort. But the deliverable is now a decision path, and a decision path carries fingerprints. You will be able to see, in the stack, which students actually held a constraint in their head and which ones received a conclusion and dressed it.

That is what review first, always means at the assignment level. You are not reviewing the output after the fact. You are designing the assignment so the output has to show its work in the first place.

The part that does not scale, and what to do about it

You can do this by hand and you should do the first one by hand. It is twenty minutes with a pen and it will teach you more about your own assignment than any tool will.

Then count. Most instructors reading this have four or five assignments per course that ask for a recommendation, a comparison or a choice, and they teach more than one course, and the rubrics have to move with the prompts or the rewrite does nothing. The method is sound. The arithmetic is what stops people, and it stops them in about week three.

That is the job the Assessment Builder in the Faculty Toolkit does: assignments, rubrics and tagged question banks, built against your own Course Record so the criteria come out in your language and at your standard, and nothing reaches a student until you approve it. Do the first one by hand anyway. You need to know what good looks like before you ask anything to produce it at volume.

If you want to work through the rewrite alongside other instructors doing the same thing this term, the free Harvest Kernel community is open.

The line to remember

Ads in ChatGPT are labeled and separate today. The company that runs them has published its intention to make them more native. Between those two facts sits every assignment you are about to hand out this term.

Open the assignment you grade next. Ask who else would be satisfied with the answer it rewards.

Sources

• OpenAI, “A milestone in expanding access to AI,” August 31, 2026. openai.com/index/expanding-access-to-ai-with-chatgpt-ads. All quoted sentences in this piece are taken verbatim from that post.