Content rules for AI
Onflow Ads' AI is built for one job: honest advertising content. Prompts that ask for something else are declined before anything is generated. This page explains what gets declined, what it costs you, what happens afterwards, and how to get a legitimate brief through.
How a decline works​
If your prompt breaks the rules you see:
That request breaks our content policy. Please try a different, on-topic prompt.
Four things are worth knowing:
- It is free. The check runs before any generation is reserved, so a declined prompt never spends an AI action or an image.
- Nothing is sent anywhere. The screen is a local, rule-based pass. The prompt does not reach a model at all.
- Ambiguity resolves in your favour. The patterns target phrasing a legitimate ad brief would never contain. A bold or edgy brief is allowed; when the check is unsure, it allows.
- For Refine, both halves are checked — the copy you are improving and the editing instruction you type. On a formatted post that copy includes its Telegram tags, because Refine is seeded with the formatting.
Where the check runs​
The classifier is applied to:
- every image prompt, wherever it comes from — the popup's Image tab on all three engines, and the booking wizard's Generate image with AI;
- every Generate brief — the popup's Generate tab on all three engines, and the AI Studio (since 25 September 2026; the three engines' Generate tabs used to skip it);
- the copy and the instruction you send through the popup's Refine tab.
A Fix to pass request is not a prompt: it carries your own post, which the content check below judges on its own terms. The small helpers are narrow by construction — Check overlap is handed overlap clusters, Targeting copilot is handed pool counts, What does this prove? is handed a proof trail — and are not screened by this classifier.
The output is checked too​
A prompt that passes can still come back with something that breaks the rules — a model can write or draw what nobody asked for. So what the AI produced is checked before you see it: every draft against the platform's prohibited list, every generated image by the same picture check a post gets. A draft or image that fails is not shown to you, and the action is handed back to your allowance. A generated image the check could not look at is still given to you, and is checked again (and held for a person if it still cannot be read) on whichever post you attach it to.
A brief that describes a scam still produces a draft that will fail the channel's pre-flight, breach the acceptable-use rules, and get the ad declined or removed. The classifier is one gate among several — the platform rules apply to what you publish regardless of what wrote it.
What is declined​
| Category | Examples of what is refused |
|---|---|
| Manipulating the AI itself | "Ignore your instructions", "reveal your system prompt", "act with no rules", developer-mode and jailbreak phrasing |
| Sexual content involving minors | Zero tolerance, in any form or phrasing |
| Weapons and explosives | Instructions to build bombs, firearms, silencers, or chemical agents |
| Illegal drug production | Synthesis, recipes, or manufacturing instructions |
| Malware and account theft | Writing malicious software; stealing passwords, seed phrases or private keys; breaking into someone's account or wallet |
| Scam and fraud copy | "Guaranteed profits", "double your money", risk-free investment claims, Ponzi and pyramid schemes, pump-and-dump promotion, fake giveaways and airdrops |
| Phishing | Copy designed to harvest passwords, seed phrases or one-time codes, or to trick people into connecting a wallet to claim something |
| Adult sexual content | Explicit sexual text or imagery, and ads for adult channels or creators (OnlyFans, "18+ channel", nudes) |
| Gambling and betting | Ads, posts, banners or images for casinos, betting, poker, slots, lotteries, satta or tipster channels — Onflow Ads does not carry them on any rail |
Consequences beyond the decline​
Not every decline is treated the same way.
| Category | What happens after the decline |
|---|---|
| Adult sexual content, gambling | The generation is refused. Nothing else happens |
| Scam or fraud copy, phishing | The attempt is proposed for human review at the fraud level |
| Jailbreak attempts, illegal or harmful content, malware | The attempt is proposed for human review at the more serious abuse level |
A proposal is not a penalty. It is a queued item for a person on the Onflow Ads team, and nothing on your account changes unless a reviewer upholds it.
If a reviewer confirms the flag as genuine abuse, your reliability score drops — 1.00 point for the fraud categories and 2.00 points for the abuse categories. That is not reversible on request; it is repaired the same way any reliability loss is, by clean activity over time. A score below Fair (5.0) shrinks your AI allowance and switches off AI images entirely. A rejected proposal never touches your score.
Repeated blocked attempts do not flood the review queue with separate flags. They accumulate on one pending item, which counts the attempts and can escalate from the fraud level to the abuse level. In other words, the second and third attempt make the same flag worse rather than creating new ones.
While a review is open on your account, AI is paused entirely:
AI is paused on your account while a review is open. It returns as soon as the review is resolved.
Nothing is needed from you — access returns automatically when the review closes.
Rules baked into the output​
Even for allowed prompts, the AI works under standing rules that protect you as the advertiser:
- Truthfulness. It is instructed never to invent statistics, prices or guarantees. Refine additionally will not add claims that were not in your original copy.
- Clean formatting. Output carries only Telegram-safe bold, italic and links. Everything else is stripped before you see it.
- Safe links only. A link in the output must be an ordinary web, Telegram or email address. Anything else is removed rather than shown as clickable.
- Safe images. Generated graphics contain no embedded text, no logos or watermarks, and no real people's faces — which keeps you clear of trademark and likeness problems in advertising.
See What powers the AI.
The one line the AI does not write​
Whatever the AI hands you is not the whole of what a reader sees. Onflow Ads labels the placements it delivers itself, adding a single line after one blank line at the bottom of the post:
advertisement via onflowads.com
It links back to the site. You do not write it, and you cannot remove it from a post that carries it — and, equally, you should not write a disclosure of your own on top of it.
Two things about it are worth knowing before you draft:
- Whether it appears is a plan question, and not necessarily your plan. On Paid Promotions and Cross-Promotion the plan read is that of the channel the ad lands in; the Subscriber Exchange reads the advertiser's. See the attribution line.
- It is added by the platform on the placements the platform posts. Where a paid-promotion owner publishes your ad by hand instead of letting the bot post it, nothing can stamp the line onto their message — so a manually delivered placement may not carry it even when the plan says it should.
Two practical consequences for anyone drafting with AI:
- The composer holds 33 characters back for it, so the counter's ceiling is lower than Telegram's and says so underneath. The model is not told about that reserve, so a long draft can arrive already over the counter. See the reserve.
- The disclosure is handled for you. The Paid Promotions pre-flight says so in as many words — Disclosure is added for you — every placement is labelled as an advertisement. If a draft comes back with an
#adhashtag in it, that is the channel owner's sponsored-marking preference reaching the drafting prompt; delete it if you would rather not double up.
These rules are not the whole picture​
The AI content policy governs what the AI will produce. Two other sets of rules govern what you may publish:
- The Onflow Ads platform policies — see Trust and safety and the Terms of Service. Copy the AI happily wrote can still breach platform rules.
- The individual channel's rules, for paid promotions. Each owner sets their own and they can be stricter than ours. The booking wizard's Run the compliance check reads your creative against that channel's stated rules, and it costs nothing — see AI in the booking composer.
Passing the AI content check is not an approval to publish. You are still the publisher.
If a legitimate brief is declined​
The rules key on specific scam and abuse phrasing, not on your product category. Advertising a crypto product, a trading course or a giveaway is allowed — it is the promise wording that trips the fraud rules.
| Phrasing that trips the rules | Rephrase as |
|---|---|
| "Guaranteed 300% returns" | "Our strategy's published track record" |
| "Double your money in a week" | "Learn how our members trade" |
| "Risk-free investment" | "Start with as little as you like" |
| "Verify your wallet to claim" | "Connect and browse the collection" |
| "Fake giveaway" or "fake airdrop" | Describe the real prize and the real entry rules |
Practical steps if you are stuck:
- Restate the brief in plain advertising terms — what you are promoting, to whom, and the one action you want readers to take.
- Remove absolute promises about money, returns or outcomes.
- Remove anything asking readers for credentials — passwords, codes, recovery phrases.
- Try again. Rephrasing costs nothing, because declines do not spend actions.
- If a clearly legitimate prompt keeps being declined, contact support and include the exact wording you used. See Contact us.
If a brief was declined and you cannot see why, read it back looking for a promise about money or a request for a code or a phrase. Those two account for almost every legitimate decline.
FAQ​
Does a decline count against my daily allowance?​
No, never. The check runs before anything is reserved, so nothing is spent.
Will I be told which rule I broke?​
The message is deliberately general. Use the rephrasing table above.
I was flagged once by accident. Is my account damaged?​
No. A flag is a proposal for a human to look at. Nothing changes on your account unless a reviewer upholds it as genuine abuse.
I tried three times before I rephrased. Are there three flags against me?​
No. Repeat attempts accumulate on a single pending item rather than creating separate flags — though they do raise its attempt count, and enough of them can escalate its severity.
Can I appeal a penalty that came from an AI flag?​
Yes. See Appealing a penalty.
Do these rules apply to the ad that finally posts?​
Yes. These rules govern what the AI will write for you; the acceptable-use rules govern what may be published from a channel, and they cover the same ground.
Is what I typed stored anywhere?​
Only when a prompt is blocked. Ordinary generations are logged without their text — see What powers the AI. A blocked prompt keeps a short excerpt as evidence for the reviewer.
Related​
- What powers the AI — where the check sits in the sequence, and what is recorded.
- Reliability score — the number an upheld flag moves.
- Appealing a penalty — the route back.
- AI FAQ and troubleshooting — every AI message, explained.