Don't stop here
Hand-picked guides our readers explore right after this one.
Unlock Google's Gemini with multimodal prompting strategies
Read the guideStunning image generation with Midjourney prompt mastery
Read the guideAI prompts for legal research, contract drafting, case analysis, and client communication
Read the guideGemini refusing to make an image of a person is policy, not a malfunction, and the boundaries tightened in 2026. Google first paused people generation entirely in February 2024 after a controversy over historically inaccurate depictions, restored it later that year with guardrails, and then narrowed it again with the Nano Banana 2 safety update in February 2026, which added restrictions covering public figures and celebrities, face swapping, outfit swapping on real people, and manipulation of financial imagery. Deepfake legislation is the driver, so these limits are unlikely to loosen. What that leaves is still substantial: fictional characters, stylised and illustrated people, generic figures in a scene, and edits to your own photos that do not involve identity or clothing swaps. The refusals are also often silent or vague, appearing as 'I can't help with that' or 'I'm still learning how to answer that question' rather than naming the rule, which is why people assume it is a bug. This is general information about how the tool behaves, not legal advice: if you need images of a real, identifiable person for commercial use, consent and likeness rights are the actual constraint.
'I can't help with that' or 'I'm still learning how to answer that question' on an image request
The same prompt works when you remove a real person's name
Face-swap and outfit-swap edits on real photos are refused
Requests naming a celebrity, politician, or public figure fail every time
Generic 'a person' prompts work while specific-identity prompts do not
A prompt that worked in 2025 now fails without explanation
The February 2026 Nano Banana 2 update added public figures and celebrities as an explicit restriction category. Naming a real, identifiable person in an image prompt is the most reliable way to trigger a refusal.
Both were added as restriction categories in 2026. Editing a real person's face onto another body, or changing what an identifiable real person is wearing, is refused regardless of intent.
The restrictions exist because of deepfake legislation and non-consensual imagery risk, not conservatism about art. That is why these particular categories are hard blocks rather than things you can talk your way past.
Editing financial imagery (documents, statements, and similar) was also added as a restriction category in 2026, since the fraud risk is direct.
Gemini frequently answers with a generic 'I can't help with that' or 'I'm still learning how to answer that question' rather than naming the policy. That makes a deliberate block look identical to a temporary failure, so people retry instead of rephrasing.
When to try: First, on any refused people prompt
Replace the identity with a description: instead of naming a public figure, describe age range, build, clothing, expression, and setting. Prompts that describe a person without identifying a real one generally succeed where named ones always fail.
When to try: When you need a person in the image but not a specific one
State plainly that you want a fictional character, an illustration, or a stylised portrait. Fictional humans, illustrated people, and stylised portraits remain supported. Adding a style (watercolour, line art, 3D render) both improves the result and moves it further from photoreal likeness.
When to try: Immediately, to distinguish policy from a glitch
Because refusal text is vague, send the same prompt once more in a fresh chat. If it fails identically, it is policy and no amount of retrying will change it. If it succeeds, the first attempt was a transient error. This one test saves a lot of wasted effort.
When to try: When an edit to a real photo is refused
Face swaps and outfit swaps on real people are blocked outright. If you are editing your own photo, ask for a change that is not identity or clothing substitution: background replacement, lighting, colour grading, cleanup, or cropping. Describe the finished image you want rather than the swap operation.
When to try: When an obviously harmless prompt is refused
Adding legitimate context ('an illustration for a blog post about remote work, generic figures, no real people') helps the model resolve an ambiguous prompt in your favour. It will not unlock a blocked category, but it does reduce false refusals on borderline everyday requests.
When to try: When you need a real person and the use is commercial
If you genuinely need an image of a real, identifiable person, the route is licensed stock photography or a shoot with a signed release, not prompt engineering. Consent and likeness rights are the real constraint and no AI tool resolves them for you. This is general information rather than legal advice.
When to try: If Gemini's boundaries block your legitimate creative work
If your work is stylised character art rather than real likenesses, other generators have different boundaries and may fit better. Keep in mind that public-figure and non-consensual imagery restrictions are broadly industry-wide in 2026, so switching tools does not remove those categories.
When to try: When a plainly harmless prompt is blocked repeatedly
Use the thumbs-down control on the refusal and note what you actually asked for. False positives on harmless prompts are how safety tuning gets corrected, and the feedback carries the conversation with it.
Describe people by attributes and setting rather than by name
Say 'fictional character' or name a visual style up front for people images
Do not frame edits as face swaps or outfit swaps on real people
For real, identifiable people in commercial work, license the image instead
Use the in-app thumbs-down feedback for false refusals, since that is the channel that reaches the safety tuning process. Contact Google support only if image generation is missing entirely from a plan that includes it, or if every image prompt fails including plainly generic ones, which points to an account or service issue rather than policy. Blocked categories such as public figures, face swaps, and outfit swaps on real people are deliberate and will not be lifted per-request.
It will, with limits. Google paused people generation in February 2024 after a controversy over historically inaccurate depictions, restored it with guardrails, then narrowed it again with the February 2026 Nano Banana 2 update, which added public figures and celebrities, face swapping, outfit swapping on real people, and financial image manipulation as restricted categories. Fictional, stylised, and generic people still work.
No. Public figures and celebrities are an explicit restriction category as of February 2026, driven by deepfake legislation and likeness risk. Naming a real, identifiable person in an image prompt reliably triggers a refusal, and rephrasing will not unlock it.
Because refusals are often surfaced with generic text rather than the specific policy. That makes a deliberate block look like a temporary glitch. Send the prompt once more in a fresh chat: if it fails identically it is policy, and rephrasing to remove the identity or the swap is the only path forward.
Fictional human characters, stylised portraits, illustrated people, and generic figures described by attributes rather than identity. Edits to your own photos are generally fine as long as they are not identity or clothing substitution: background changes, lighting, colour grading, and cleanup usually pass.
License it. Use stock photography with the appropriate rights, or commission a shoot with a signed release. Consent and likeness rights are the actual constraint, and they are not something an AI generator can grant you. Treat this as general information, not legal advice, and check the rules that apply in your jurisdiction.