Google Quietly Restricts Disney IP Prompts in Gemini: Legal Takeaways
Google's AI tools now refuse prompts that mention Disney-owned characters. Tests that once returned glossy images of Yoda, Iron Man, Elsa, and Winnie-the-Pooh are now blocked with a notice: "I can't generate the image you requested right now due to concerns from third-party content providers."
This shift follows a December cease-and-desist from Disney alleging Google's AI products - including Veo, Gemini, and "Nano Banana" - acted like a "virtual vending machine" for protected IP. The timing suggests Google implemented tighter filters in response to rights-holder pressure.
One wrinkle: image-plus-text prompts still appear to generate Disney-related output. For example, uploading a Buzz Lightyear photo and pairing it with a figurine-style prompt produced a new "virtual figurine" of the character.
Why This Matters for Counsel
Disney's 32-page letter reportedly accused Google of large-scale copyright infringement and demanded immediate changes, including halting training on Disney IP. It featured examples of outputs depicting Darth Vader, Iron Man, and other characters, arguing that simple prompts yielded high-fidelity, unauthorized works.
Google has emphasized ongoing engagement with Disney and said it trains on publicly available web data while offering controls like Google-Extended and YouTube's Content ID. That framing positions Google as providing opt-outs and rights-management tools rather than conceding infringement across the board.
Context matters: the warning arrived the same week Disney struck a reported $1B licensing deal with OpenAI tied to Sora. Rights-holders are signaling a two-track strategy - enforce where needed, license where strategic.
Key Legal Issues in Play
- Direct and secondary infringement risk: High-fidelity character depictions from text prompts can trigger claims. Even if prompts are filtered, image-plus-text paths may keep exposure alive.
- Training-data claims: Demands to stop training on Disney IP raise live questions about copying during dataset creation, model training, and model retention of protected expression.
- Trademark and trade dress: Character names, visual identities, and branded elements expand the risk beyond pure copyright.
- Filter efficacy: Blocking text prompts is only half the job. Upload pathways, style transfer, and "figurine" or "toy" framing can replicate distinctive character elements.
- Opt-outs and provenance: Policies such as Google-Extended matter, but counsel should assess whether they apply to images, video, and all relevant crawlers - and whether vendors actually comply.
- Regulatory and policy signals: Watch ongoing guidance from the U.S. Copyright Office on AI issues (copyright.gov/ai).
What To Do Now
- For rights-holders:
- Maintain active blocklists of character names, visual traits, and franchises; require vendors to enforce across text, image, and video inputs.
- Run structured tests and preserve evidence across prompt types (text-only, image+text, style prompts). Document bypass scenarios like "figurine" outputs.
- Press for training and retention commitments (no use of protected IP; deletion or isolation of known tainted data; model remediation plans).
- Escalate with targeted notices and renewal of C&Ds when bypasses surface.
- For AI vendors and enterprise adopters:
- Close the "upload loophole" by mirroring text-prompt filters on image/video inputs and blended prompts.
- Implement dynamic blocklists tied to active disputes and license terms; log denials and near-miss detections for auditability.
- Align user terms with your technical controls: prohibit attempts to recreate protected characters; rate-limit or sandbox style prompts that predictably hit IP.
- Offer opt-out honoring and dataset transparency; be clear on data sources and any licensed sets.
- Strengthen indemnities, caps, and remediation obligations in customer and supplier contracts.
Practical Signals to Monitor
- Does Google extend denials to all Disney-adjacent prompts, including ambiguous references and stylized descriptors?
- Will other studios issue similar demands, prompting broader, category-level blocklists?
- Do hybrid prompts (photo + text) continue to produce recognizable Disney characters, and how fast are those paths closed?
- Are new licenses announced that carve out specific use cases (figurines, toys, animation styles) and relax filters accordingly?
For ongoing policy updates, keep an eye on the U.S. Copyright Office's AI page: copyright.gov/ai. Google's opt-out mechanism is described here: Google-Extended.
If your legal team is building internal literacy on AI risk, you can browse role-focused resources here: Complete AI Training - Courses by Job.
Bottom line
Google is restricting prompts that call out Disney IP, but upload-based prompts still appear to produce recognizable characters. Expect tighter filters, more licensing, and continued pressure on training-data practices. Counsel should update testing protocols, refresh contracts, and close technical gaps before the next demand letter lands.
Your membership also unlocks:
