Blog / 7 min read · 2026-05-28

Why Gemini Omni blocks your prompts — the real rules vs the current bug

Gemini Omni Flash is wrongly rejecting harmless prompts in May 2026 — a bug Google acknowledged. Source-cited: real policy blocks vs the current false-positive bug, and how to tell them apart.

If you’ve been trying to generate video with Gemini Omni Flash in the last week and keep hitting “I can’t generate that video. Try describing another idea…” on prompts that are obviously harmless — you are not doing anything wrong, and you are not alone.

There are two completely different reasons Omni refuses a prompt right now, and they get confused constantly:

1. A genuine content-policy block — the prompt actually asks for something Google prohibits. These rules are stable and documented.
2. Bug A — the false-positive policy bug — Omni is wrongly flagging ordinary prompts as policy violations. Google has publicly acknowledged this and is investigating it. Status as of 2026-05-29: no fix announced.

This post separates the two, with sources, so you stop wasting credits guessing.

Don’t confuse Bug A with Bug B. A separate billing bug — failed Omni generations counting against your quota — was also reported in May 2026, and Google announced a fix on 2026-05-28: failed requests no longer count, AI Ultra Omni allowance was doubled, Flash-Lite became free, and Pro per-prompt quota was capped (source: 9to5Google 🟡). That’s Bug B — different from Bug A, and already fixed. This post is entirely about Bug A. Full A vs B breakdown: our /guide/ page.

Bug A: the false-positive policy block (status: unresolved 2026-05-29)

Status check (2026-05-29): No fix announced. Fresh false-positive reports are still being posted in the Google AI Developers Forum thread. We’ll update this line when/if Google ships a fix.

As of late May 2026, users across Reddit, X, and Google’s own developer forum report that Gemini Omni Flash is rejecting basic, harmless prompts as policy violations. Reported examples that were wrongly blocked include:

• Cat walks between rabbits
• Make this cat wear a white hat
• Change the background to night
• Apply a green cinematic color grade to the uploaded video while preserving the same scene motion

None of these violate any policy. Video editing prompts (especially ones that reference an uploaded video) appear to be hit hardest — community reports describe video-to-video editing as “completely broken” during this window.

Source tier: 🟡 News + primary community reports. A Google VP (Josh Woodward) publicly responded to posts about the issue, indicated it should not be happening, and said Google was investigating and collecting the underlying videos from affected users. (PiunikaWeb, May 20 2026; Google AI Developers Forum)

What this means for you: if a plainly innocent prompt gets refused right now, it is most likely this bug — not a real rule you broke. Don’t rewrite the prompt ten times trying to “sound safer.” See the workarounds at the end.

The real rules — what Gemini actually prohibits

These are stable and official. Per Google’s Generative AI Prohibited Use Policy, you genuinely cannot generate content that does the following (categories quoted from the policy):

• Dangerous or illegal activities — “engage in dangerous or illegal activities, or otherwise violate applicable law or regulations.”
• Security circumvention — “compromise the security of others’ or Google’s services, including attempts to circumvent safety protections (e.g. prompt injections).”
• Privacy violations — “using someone’s image without their consent and non-consensual intimate imagery.”
• Fraud and deception — “engage in frauds, scams or other deceptive actions.”
• Child safety — any content related to child sexual abuse or exploitation.

The Gemini Omni Flash model card restates the same boundaries for the video model specifically: it should not be used to “(1) engage in dangerous or illicit activities… (2) compromise the security of others’ or Google’s services, (3) engage in sexually explicit, violent, hateful, or harmful activities, (4) engage in misinformation, misrepresentation, or misleading activities.”

Source tier: 🟢 Official Google policy + DeepMind model card.

Real people, copyrighted characters, and likeness

This is the block category that trips up the most legitimate creators. Generating a recognizable real person (a celebrity, a public figure, or someone whose image you don’t have consent to use) falls under the privacy and misrepresentation rules above. Copyrighted characters are similarly restricted.

The sanctioned path for putting a real person in a video is the Avatars feature — you build a digital version of yourself (or a branded character) with an explicit voice setup first, and then that avatar can speak new lines. That’s the difference between “make a video of [famous person]” (blocked) and “make a video of my own verified avatar” (allowed).

Source tier: 🟢 Official (privacy policy + Avatars documentation). See our Avatar feature rules for the gating details (18+, region, language).

Two things that are policy decisions, not bugs

People mistake these for blocks too, so it’s worth being precise:

• Audio / speech editing is deliberately withheld. The model card states plainly: “Gemini Omni Flash is capable of changing people’s speech. For now, we are restricting this capability and working to better understand how to safely and responsibly bring it to our users.” So if your prompt to alter spoken audio does nothing, that’s intended — not a filter you can phrase around. 🟢
• The 10-second clip cap and the multi-turn editing limit are intentional. Independent testing reports a practical ceiling of about 4 conversational editing turns before consistency degrades. That’s a design boundary, not a rejection. 🟡

SynthID — you can’t remove the watermark, and that’s not a “block”

Every Gemini Omni output carries an imperceptible SynthID watermark plus C2PA Content Credentials. There is no setting or API flag to turn either off, and SynthID is designed to survive cropping, filters, frame-rate changes, and lossy compression. This isn’t content moderation — it’s provenance — but it’s worth knowing that “remove the watermark” is not a thing you can prompt your way around.

Source tier: 🟢 DeepMind model card + SynthID documentation.

How to tell a real policy block from Bug A

A quick decision tree:

1. Does your prompt name a real/famous person, ask for violence/sexual content, or try to bypass safety? → That’s a genuine policy block. Rephrasing won’t (and shouldn’t) help.
2. Is your prompt plainly innocent (animals, objects, scenery, color grades) but still refused? → Almost certainly the current false-positive bug. Don’t over-rewrite it.
3. Did a text-to-video prompt work but the same idea as a video edit fail? → That matches the bug’s known signature (video-editing prompts hit hardest).

Practical workarounds while Bug A is live

• Prefer text-to-video over video-to-video for now — the editing path is the most affected. Generate fresh rather than editing an uploaded clip.
• Strip ambiguous verbs. Words like “shoot,” “strike,” “blast,” or “kill the lights” can trip the classifier even in innocent contexts. Use neutral phrasing (“the lamp switches off”).
• Don’t burn credits brute-forcing. If a clearly harmless prompt is refused, it’s the bug, not your wording — wait or try a slightly different scene rather than 10 rewrites.
• Report it. Google is actively collecting examples; reporting a wrongly-blocked prompt through the in-app feedback helps the fix land faster.

Bottom line

Right now, most “policy violation” refusals on harmless prompts are Bug A — a bug Google has admitted to and is still working on (no fix announced as of 2026-05-29). The genuine blocks are narrow and stable: real people without consent, illegal/dangerous content, sexual/violent content, and security bypasses. Audio editing and the 10-second cap are deliberate holds, and SynthID + C2PA are non-removable provenance, not moderation. Knowing which bucket your refusal falls into saves you credits and frustration.

Don’t confuse Bug A with Bug B — the unrelated quota bug Google announced fixed on 2026-05-28. Both have been called “Omni blocking” in community posts; only Bug A is still live. Full disambiguation: our /guide/ page.

We test these boundaries first-hand and update this post as Google ships the fix. Every claim here links to its source — if something changes, we’ll change it here too.

Last updated: 2026-05-29.