A claims-risk check is an automated review that scores every AI-generated shot for overstated or unsupported claims, the superlatives, guarantees, and hard numbers that promise more than you can actually back up, before that shot is allowed into the cut. In CoreReflex it runs as one dimension of the quality gate that grades each shot, so brand safety is enforced shot by shot instead of being discovered after a client has already watched the video. The result is AI video you can hand to a stakeholder and defend.
What a claims-risk check actually does
Generative models are fluent and confident, which is exactly the problem. Ask for an energetic product spot and a model will happily stamp "the #1 choice" across the screen or have a voiceover promise results it has no basis for. On a single hero shot you might catch it. Across a batch of fifty social cuts, you won't, and one overstated line in a paid ad is the kind of thing that draws a platform takedown, a legal note, or an awkward call from the client.
The claims-risk check looks at each shot's prompt, its narration, and its on-screen text and asks a narrow question: does anything here assert more than the brief can support? It flags absolute superlatives, unconditional guarantees, health or financial promises, and unsupported statistics, and it surfaces what it flagged so a human can make the call. It is a smoke detector for marketing language, wired into the same loop that already checks whether a shot is sharp and on-model.
Where the claims-risk check sits in the pipeline
CoreReflex runs an agentic Director loop: PLAN → PRODUCE → CRITIQUE → ASSEMBLE. The Director boards your shots, generates them on the owned Vertex AI stack, critiques each one, and only then assembles the cut. The claims-risk check lives in that CRITIQUE step, alongside the other gate dimensions every shot is scored on:
- prompt match, did the shot deliver the brief
- sharpness, is the image clean, not soft or artifacted
- motion coherence, does movement hold together
- on-screen text legibility, does any text render correctly and read
- on-brand, does it match your brand kit
- claims-risk, does the messaging overreach
Because claims-risk is scored at the same moment as everything else, a shot that looks gorgeous but says "guaranteed to double your revenue" does not get a pass just for being pretty, it is held at the gate with the rest. For how a flagged shot then gets fixed, see selective regeneration.
What counts as a risky claim
The exact threshold depends on your brand and category, but the recurring offenders are consistent:
- Absolute superlatives, "the best," "#1," "the only," "guaranteed", stated as fact without support.
- Performance promises, specific outcomes ("lose 10 pounds in a week," "triple your leads") presented as certainties.
- Regulated-category language, health, finance, and safety claims that carry real compliance weight.
- Unsupported statistics, a confident "93% of users" with nothing behind it.
- Comparative claims, "better than" a named competitor, which raises the bar on substantiation.
These show up in narration the voice engine reads, in the prompt that drives the visuals, and, most easily missed, in on-screen text baked into the frame. The gate inspects all three surfaces, which is why text legibility and claims-risk are scored together: a claim you can clearly read is a claim you can be held to.
What happens when a shot is flagged
A flag is not a wall. The Director treats a claims-risk failure like any other gate miss: it regenerates just that shot, carrying forward the plan, the camera move, and the continuity anchor, while leaving the shots that already passed untouched. In practice that means the offending line gets softened or re-grounded, "the #1 choice" becomes "a favorite," a hard promise becomes a benefit, and the shot is re-scored. You stay in control: every flag is visible, so you can accept the safer version, rewrite the line yourself, or decide the claim is in fact substantiated and keep it.
This selective approach is why the check is practical rather than annoying. You are not regenerating a whole film because one caption overreached; you are fixing one shot while the rest of the cut holds.
Provenance: a paper trail you can defend
Every generation in CoreReflex carries a portable trace, the model, the prompt, the parameters, and the scores it earned, including its claims-risk result. That provenance turns brand safety from a vibe into a record. If a client or a platform ever asks why a deliverable says what it says, you can replay exactly how each shot was produced and show that it passed the gate. Reproducibility plus a scored, auditable history is what separates "the AI made it" from a deliverable you can stand behind, the same argument we make for owning the whole AI stack on client work.
Why agencies and brand teams care
For anyone shipping video on behalf of someone else, the claims-risk check is risk insurance at production speed. An agency producing volume for multiple clients can't manually proof every line in every cut, but it absolutely owns the consequences when an overstated claim slips through. A per-shot gate moves that review from "hope the editor caught it" to "the system flagged it before assembly," and it does so consistently across every shot, every batch, every brand.
It also pairs naturally with the brand controls elsewhere in the platform. The claims-risk check guards what you promise; the brand-voice guard covered in keeping AI video on-brand guards how you sound. Run together, they keep volume output both legally cautious and tonally consistent, and because spend and safety live on the same surface, you get the unified view described in one billing and safety surface for AI media.
What it is, and what it isn't
Be clear-eyed about scope. The claims-risk check is an automated safeguard that catches the obvious and not-so-obvious overreaches before they ship. It is not legal advice, and it does not certify that a deliverable complies with any specific regulation or ad policy. Substantiating a claim is still your responsibility; the check's job is to make sure nothing reaches a client or a platform without a human at least being told "this line is making a promise." Used that way, as a fast, consistent first line of defense rather than a rubber stamp, it removes an enormous amount of avoidable risk.
Frequently asked questions
What is a claims-risk check in AI video?
It is an automated review that scores each generated shot for overstated or unsupported claims, superlatives, guarantees, performance promises, and unbacked statistics, in its narration, prompt, and on-screen text. In CoreReflex it runs inside the per-shot quality gate, so risky messaging is caught during production rather than after a client has seen the cut.
How does CoreReflex catch overstated claims?
The Director scores claims-risk in the CRITIQUE step of its PLAN → PRODUCE → CRITIQUE → ASSEMBLE loop, alongside checks for sharpness, text legibility, and brand fit. When a shot overreaches, it is flagged with the reason and selectively regenerated, the line is softened or re-grounded, while the shots that already passed are left alone.
Does the claims-risk check guarantee my video is legally compliant?
No. It is a safeguard that flags likely overstatements so a human can review them; it is not legal advice and does not certify compliance with any regulation or platform policy. You remain responsible for substantiating claims, the check's value is catching the obvious risks consistently, at production speed.
Can I override a flag if the claim is true?
Yes. Flags are visible, not blocking. If a claim is substantiated, you can keep the original shot; if not, you can accept the softened regeneration or rewrite the line yourself. The point is informed control, not automatic censorship.
Where is the result stored?
In the shot's provenance trace, along with the model, prompt, parameters, and every other gate score, that makes each claims-risk decision reproducible and auditable; see the developer docs for how to read a shot's full trace.
Ship video you can defend
Brand-safe AI video isn't about slowing down, it's about a gate that scores every shot for risk while you keep producing at volume. The claims-risk check turns "I hope nobody overpromised" into a record you can replay and defend, and it's one of several gate dimensions we break down across our how CoreReflex works explainers. Start free, no credit card, and put a quality gate on every shot you ship.