OpenAI’s rules tell political campaigns to keep its consumer chatbot out of their advertising workflow, and the ChatGPT political ads ban is now being tested in the open: The Washington Post reports that campaigns are using ChatGPT to make ads anyway. No model changed this week and no product was launched. What surfaced is the distance between a written acceptable-use rule and what people actually do once they have an account, which is a problem every developer building on a hosted model shares with OpenAI.
Key takeaways
- The Washington Post reports that political campaigns are producing advertising with ChatGPT despite OpenAI’s policy prohibiting campaign use.
- The snippet available does not name campaigns, quantify how widespread the practice is, or include an OpenAI response, so treat the scale as unquantified.
- This is an enforcement story, not a capability story: a usage policy is a contractual condition checked after the fact, not a limit compiled into the model.
- Cost is no longer a gate. Per the Convly models database, Sora 2 starts at $0.05 per second of video, Sora 2 Pro at $0.15, Veo 3.1 at $0.05, Kling 2.5 Turbo Pro at $0.07 and Wan 2.5 at $0.05.
- Ad copy is cheaper still: Claude Sonnet 5 runs $2.00 in / $10.00 out per million tokens and Gemini 3.6 Flash $1.50 / $7.50, so a full script costs fractions of a cent.
- Open-weight models mean a single vendor’s restriction is a routing decision for the user, not a hard boundary.
What the ChatGPT political ads ban actually covers
The reported facts here are narrow and worth stating precisely. According to The Washington Post, OpenAI bars political campaigns from using ChatGPT to create advertising, and campaigns are using it for that purpose regardless. The material available does not identify which campaigns, how many ads are involved, what share of a given ad was machine-generated, or whether OpenAI has taken action against any account. Anyone quoting a number on this story is quoting something that has not been reported.
The policy itself is not secret. OpenAI publishes its rules for what customers may and may not do with its models in its usage policies, and political campaigning has long sat in the restricted category alongside other high-risk categories. That document is the primary source for the rule, and it is the page a compliance team should read rather than a summary of it. What the document cannot do is describe how the rule is detected or enforced, which is exactly where the reported behaviour sits.
Why usage policies break down at the prompt level
As context rather than reported fact, the structural reason a rule like this leaks is worth spelling out, because it generalises to every model in our AI models database. A usage policy is a contract term. Model safety training is a behavioural tendency. Neither is a filter that inspects who the end client is. When a request arrives asking for five variants of a thirty-second script about rising grocery prices, aimed at voters aged 35 to 54, there is very little to distinguish it from ordinary brand marketing work. The model has no way to verify the identity of the ultimate customer, and the operator has no reliable post-hoc signal either.
That leaves enforcement dependent on external reporting, self-disclosure, or account-level pattern analysis — all of which are slow relative to a campaign cycle. It also means the practical penalty for breaching an acceptable-use term is account suspension after the work has already shipped, not a blocked generation at the moment of the request. For developers, that is the operative lesson: the terms you agree to are enforced against your account, not against your prompt.
Ad-grade video generation now costs cents per second
The economics matter because they explain why a policy is being ignored rather than merely disliked. Producing a short spot used to require a crew, an edit suite and a budget large enough to be visible in filings. Generated video has collapsed that line item. The figures below are the starting per-second prices we publish in our own database, and they apply to any user of the hosted endpoints, campaign or otherwise.
| Model | Vendor | Starting price per second |
|---|---|---|
| Sora 2 | OpenAI | $0.05 |
| Sora 2 Pro | OpenAI | $0.15 |
| Veo 3.1 | $0.05 | |
| Kling 2.5 Turbo Pro | Kuaishou | $0.07 |
| Wan 2.5 | Alibaba | $0.05 |
At $0.05 per second, thirty seconds of footage is $1.50 of compute before iteration. Even at the Sora 2 Pro rate of $0.15 per second, the same thirty seconds is $4.50. Script and caption generation adds almost nothing on top: GPT-5.6 Sol is priced at $5.00 in / $30.00 out per million tokens, Claude Sonnet 5 at $2.00 / $10.00 and Gemini 3.6 Flash at $1.50 / $7.50, and an advertisement is a few hundred tokens. If you want to price a specific volume of drafts and variants, our AI API cost calculator and AI price-performance index cover the arithmetic across vendors.
Open weights make one vendor’s rules optional
There is a second reason a single company’s restriction has limited reach, and it is the same reason enterprise buyers keep asking about self-hosting. Restricted use of a hosted product is enforceable only while the user needs that product. Several of the strongest open-weight models can be run on hardware a small agency can rent or buy: Qwen3 32B needs roughly 20 GB of VRAM at 4-bit, Llama 3.3 70B roughly 40 GB, and both are capable copywriters. Once a model runs on your own machine, no acceptable-use policy is in the request path at all.
That is the uncomfortable structural fact behind the story: a policy of this kind functions as a reputational and contractual instrument, not a technical control. Readers weighing the trade-off can size the hardware with our free VRAM calculator, and the wider cost picture is covered in our open vs closed AI cost study. The point is not that campaigns are self-hosting — nothing reported says they are — but that the ceiling on enforcement is low even in principle.
What this means for developers building on model APIs
If you ship an application on top of a hosted model, this story is a preview of the risk you carry on behalf of your customers. Your provider’s restricted-use list applies to your account, and your end users are the ones generating the prompts. Three practical consequences follow. First, restricted categories need to be handled in your own product layer, because the upstream model will not reliably refuse work that is only contextually prohibited. Second, prompt and output logging is what lets you answer a provider’s questions if an account review lands, so retention policy is a compliance decision as much as a storage one. Third, contracts with clients should name the restricted categories explicitly, since a breach by a customer is still a breach on your account.
None of this is exotic. It is the same discipline that regulated industries already apply to model use, extended to a category — political advertising — where the reputational exposure is unusually high and the review cycle unusually short.
Provenance and the disclosure gap
The detection problem compounds the enforcement problem. Provenance metadata attached at generation time is fragile in ordinary use: files get re-encoded, cropped, re-uploaded and passed through editing tools, and embedded signals frequently do not survive that journey intact. Visible disclosure requirements for synthetic content in political advertising vary by jurisdiction and are still settling, which is context rather than anything reported in this story. The combined effect is that a finished ad rarely carries a dependable record of which model touched it — so the question of whether a specific spot breached the ChatGPT political ads ban is usually unanswerable from the artefact alone.
Frequently asked questions
What exactly does the ChatGPT political ads ban prohibit? OpenAI’s published usage policies restrict political campaigning uses of its models; The Washington Post reports that campaigns are nonetheless using ChatGPT to make advertising. The exact scope is defined in OpenAI’s policy document, which is the authoritative reference and is updated by the company over time.
Has OpenAI penalised any campaign? Not according to anything in the reporting available. The material describes the rule and the behaviour, not an enforcement action or a company response, so no conclusion about penalties should be drawn.
Why can’t OpenAI just block campaign ads at generation time? Because a request to write persuasive short-form copy or generate thirty seconds of video looks like commercial marketing work. The model cannot verify who the ultimate client is, so the restriction is enforced against accounts after the fact rather than against individual prompts.
How cheap is it to generate an ad now? Using the starting prices in our database, thirty seconds of Sora 2 output is about $1.50 of compute, or about $4.50 on Sora 2 Pro at $0.15 per second. Competing endpoints start at $0.05 per second for Veo 3.1 and Wan 2.5 and $0.07 for Kling 2.5 Turbo Pro. Script generation is a rounding error next to that.
Does this affect me if I only use the API for normal products? Indirectly, yes. Restricted-use categories are enforced at account level, so any application that lets third parties submit prompts inherits the compliance burden. Building your own category controls and keeping usable logs is the mitigation.
The bottom line
The reported story is small and specific: a rule exists, and per The Washington Post, campaigns are working around it. The significance is larger than the incident. Acceptable-use policies are being asked to do a job — governing the downstream purpose of a general-purpose tool — that they are structurally unable to do, at a moment when the cost of producing broadcast-length video has fallen to cents per second and capable open-weight alternatives run on a single workstation. Expect the pressure to shift from prohibition towards disclosure and provenance, because those are at least auditable. Until it does, anyone shipping on a hosted model should assume that policy compliance is their own engineering problem, not the vendor’s.
Sources: news.google.com. Reported September 05, 2026.