How to Prove Content Moderation When Your AI Generates in Real Time

Your acquirer sends the standard adult content questionnaire and one line stops you cold: describe your pre-publication content review process. You do not have one, because there is nothing to review before publication. A user types a prompt, the model renders an image or a video clip in seconds, and the output exists for the first time at the moment the customer sees it. The bank is asking you to screen content that did not exist when the screening was supposed to happen.

How can you review something that doesn't exist? That's the question a lot of business owners have to ask themselves (or their processor) when the processor wants to see how content moderation works, but there's nothing to moderate, because, well, nothing is rolled out.

This is the hardest compliance question in the adult AI space. The answer is definitely not to pretend you have a review queue. Instead, it is to show the bank a moderation pipeline that does the job pre-screening was designed to do, and to prove it with evidence an underwriter can understand.

Key takeaways

•  Card network rules require content to be reviewed before it is published, but they already contain a model for real-time content: live streaming, which must run on a platform you fully control with real-time monitoring and immediate removal.

•  For generative AI, moderation has to move from a gate at the end to controls at three points: the prompt, the model, and the output, with a hold on anything the output classifier cannot clear.

•  Underwriters do not approve architecture diagrams. They approve evidence: logs, block rates, review queue metrics, red-team results, and a monthly report that mirrors what Mastercard expects.

•  A written policy with no logs behind it reads as marketing. Logs with no policy read as chaos. You need both, retained, and retrievable per generation.

Why banks ask for pre-screening in the first place

The requirement comes from the card networks, not from your acquirer's imagination. Under the current Mastercard adult content rules, every piece of adult content must be reviewed before it goes live, and the rules explicitly reach content a provider can upload or generate. Visa's Integrity Risk Program puts adult merchants in Tier 1 and expects acquirers to confirm that you can police what appears on your platform. The bank is accountable for your content, so it wants to see the control that stops illegal or non-compliant material from reaching a paying customer.

The same rulebook already handles the one category of adult content that cannot be reviewed in advance: live streams.

For live content, Mastercard requires that the platform be fully under the merchant's control, monitored in real time, with the ability to terminate the stream immediately. That is the precedent you want to borrow.

Real-time generation is a live stream that your software produces instead of a performer, and it should be governed the same way.

Moderation as a pipeline, not a gate

A pre-publication queue is a single checkpoint. A generative platform needs three, and the bank needs to see all of them described in plain language.

•  Input controls. Prompt filtering against a maintained blocklist of terms, named individuals, age cues, and prohibited scenarios. Reference image handling that blocks or verifies any real-person likeness, which we cover in our guide to how to prevent deepfakes and NCII.

•  Model controls. Safety fine-tuning or guardrail layers that refuse categories your NSFW AI acceptable use policy prohibits, with minors as a hard refusal that no setting can disable. Document which base models you use and what you have done to constrain them.

•  Output controls. A classifier scores every generated image or frame before it is displayed. Clear outputs are released. Borderline outputs are held and routed to human review. Failing outputs are blocked and logged, and the prompt that produced them is flagged against the user account.

The output layer is the one underwriters care about most, because it is the real-time equivalent of pre-publication review. Set an explicit confidence threshold, set a latency budget that the classifier must fit inside, and decide in advance what happens when the classifier is unavailable. The correct answer is that generation pauses, not that content ships unscored.

What proof looks like to an underwriter

A compliance officer at an acquiring bank is going to ask for documents. Prepare the following before you apply, and keep it current after you are live. Our adult AI underwriting checklist covers the broader application package; this is the moderation section of it.

•  A written content moderation policy that names the three control layers, the thresholds, the human review process, and the escalation path, signed off by a named compliance owner.

•  A one-page architecture summary showing where each control sits in the request path, including the hold state for outputs that are not cleared.

•  Per-generation logs that tie together the user ID, timestamp, prompt, model and version, classifier scores, final decision, and reviewer ID if a human touched it. Retain them for at least as long as your acquirer requires, and be able to pull any single record on request.

•  Operating metrics for the last 30 and 90 days: generations attempted, prompts blocked, outputs blocked, outputs held for review, median review time, and accounts terminated for abuse.

•  Red-team results. A dated record of attempts to defeat your filters, what got through, and what you changed. A platform that has never tested its own controls has no basis for claiming they work.

•  Vendor documentation for any third-party classifier, age estimation, or likeness detection service, including what you send them and what they return.

How to build the evidence pack in 30 days

If you are starting from a working product and a thin compliance file, this sequence gets you to a package a bank will read.

•  Week one. Write the policy and map it to the pipeline. Every rule in the policy should point to a specific technical control, and every control should point back to a rule.

•  Week two. Turn on structured logging for every generation and confirm you can retrieve a record by user, by prompt, and by output. Fix retention so logs survive account deletion.

•  Week three. Stand up the human review queue with a service-level target, and start recording the metrics above. Even a small team with a daily review window counts if the numbers are real.

•  Week four. Run and document a red-team pass, tune thresholds based on what you find, and assemble the pack: policy, architecture summary, 30-day metrics, red-team report, vendor list, and sample log records with user data masked.

Monitoring after the content exists

Real-time controls cover creation.

The rules also expect you to control what persists. If users can save, share, or publish generations in galleries or profiles, those areas need the same monitoring you would apply to uploads: hashing to block re-creation of removed content, a public reporting path, 48-hour removal for non-consensual intimate imagery under the TAKE IT DOWN Act, and 7-business-day resolution for general complaints under Mastercard's standards.

Be ready to give your acquirer temporary credentials to review gated areas, since Mastercard's monitoring program now reaches members-only content.

Fold all of this into the monthly report Mastercard already requires: flagged content, actions taken, and complaints received. A merchant that sends it unprompted every month has done more to prove moderation than any policy document could.

Mistakes that get real-time platforms declined

•  Claiming the model cannot produce prohibited content, with no test results to back it up. Underwriters have heard this from every applicant.

•  Filtering prompts only. Prompt filters are the easiest layer to defeat, and a bank that sees no output classifier assumes the worst.

•  Treating the terms of service as the control. A rule the user agreed to is not a rule the platform enforced.

•  No log retention, or logs that cannot connect an output to the account that generated it.

Frequently asked questions

Do I need human review if the classifier is accurate?

Yes, for the borderline band. A classifier that releases everything it is unsure about is not a control. Human review of held outputs is also the clearest evidence that a person is accountable for what the platform publishes.

Does this apply to text-only AI companions?

Card network adult rules focus on visual depictions, but acquirers still underwrite text platforms on the same logic: can users produce prohibited content, and can you show you stop it. Log, filter, and review on the same model, scaled to the risk.

What about the records side of generated content?

Moderation proves what you prevent. Record-keeping proves who is depicted. The two overlap but are reviewed separately; see our guide to 2257 records for AI-generated content for the second half.

Next step

The adult AI platforms that stay live are the ones that stopped arguing about whether pre-screening fits their product and started proving control in the form a bank can audit. MobiusPay underwrites adult AI and NSFW AI payment processing every week, we know which evidence satisfies acquirers, and our risk management team helps merchants assemble it before the questionnaire arrives. If your model generates in real time and you need a processor that understands what that means, get started and we will walk through your pipeline together.