Can AI Ads Cause an International Incident? I Tested NanoBanana2 in 7 Languages to Find Out
Tags: ai imagegeneration multilingual google
"NanoBanana2 can finally render Japanese text perfectly!"
Yes. I know. Everyone knows. There are already 200 articles about it.
Here's what I actually want to know: Can it write Arabic — right to left — on a souvenir pamphlet without causing an international incident?
I'm a software engineer living in Vietnam. When you travel around Southeast Asia, you see souvenir shop pamphlets with hilariously wrong foreign text all the time. Every time I see one, I think: "Who approved this?"
So I asked NanoBanana2 to generate souvenir pamphlets for 7 countries, each in its native language. Same prompt structure. One shot. No retries. No cherry-picking.
The results? Some pamphlets were ready to hand out at the airport. Others... well, keep reading.
The 7 Languages
I deliberately chose languages with the most difficult writing systems. Latin-alphabet languages like English or Indonesian are too easy.
| # | Language | Script Challenge | Why I Chose It |
|---|---|---|---|
| 1 | 🇯🇵 Japanese | 3 scripts mixed (Kanji + Hiragana + Katakana) | Baseline test |
| 2 | 🇨🇳 Chinese | High-stroke-count characters | Can it write "龍" (dragon, 18 strokes)? |
| 3 | 🇰🇷 Korean | Hangul block composition | Looks simple, but the rules are strict |
| 4 | 🇻🇳 Vietnamese | Diacritical marks everywhere | I live here. I'm judging this one personally |
| 5 | 🇹🇭 Thai | 3-story stacking (consonant + vowel + tone mark) | Characters built like apartment buildings |
| 6 | 🇮🇳 Hindi | Devanagari with connected headline + conjunct characters | The horizontal line connecting everything |
| 7 | 🇸🇦 Arabic | Right-to-left + characters change shape by position | The final boss |
Combined speakers: ~3 billion people. 40% of the world's population.
Rules
Fair test. No tricks. This is science. Souvenir pamphlet science.
- Same prompt structure for all languages — "Generate a souvenir pamphlet that would be posted at a souvenir shop in [country] for [product]"
- AI writes all the text — I didn't specify any text. Product names, catchphrases, descriptions, prices — all AI-generated. If I specified the text, it would just be a copy test, not a language test
- One shot — No regeneration. What you see is the first and only attempt
- 3 scoring criteria — Text accuracy / Design quality / "Would you actually hand this out?"
One Thing I Noticed Across All 7
Every single image includes the souvenir shop in the background. I asked for "a pamphlet" and got "a pamphlet elegantly displayed inside a fully rendered souvenir shop." Thanks, I guess.
Adding "white background, pamphlet image only" to the prompt would fix this. But the shop backgrounds are actually charming, so I'm keeping them.
Stage 1: 🇯🇵 Japanese — Matcha KitKat
Warm-up round. If it can't do Japanese, this article is 3 lines long.
It did it. Better than expected.
Natural Japanese headlines, accurate product descriptions, and — this is the part that got me — "個包装10枚入り、配りやすく好評です" (individually wrapped, 10 pieces, popular for sharing). The AI considered the shareability of the souvenir.
Price: "¥1,200 (税込1,296円)". It calculated the 8% consumption tax correctly.
Nitpick: It reproduced the KitKat logo, which is a trademark gray area. And the Kyoto souvenir shop backdrop is too clean. Real ones are messier.
Score: 14/15 🟢 Ready to distribute
Stage 2: 🇨🇳 Chinese — West Lake Dragon Well Tea
The headline "一口西湖水，满盏龙井香" is a poetic couplet. This doesn't read like AI-generated text.
The small print accurately describes Dragon Well tea's famous "四绝" (Four Excellences: green color, rich aroma, sweet taste, beautiful shape) — real product knowledge.
Nitpick: The address is fake. Phone number is "XXXXXXXXX". Also, it added "买二送一" (buy 2 get 1 free) without being asked. The AI started running its own promotional campaign.
Fix: Add "don't include addresses or phone numbers" to the prompt.
Score: 13/15 🟡 Remove the fake address, then it's ready
Stage 3: 🇰🇷 Korean — Korean Seaweed Gift Set
Hangul looks simple — circles and lines. But the way those circles and lines combine follows strict rules.
"소중한 분께 전하는 한국의 맛" (Delivering the taste of Korea to someone special). Perfect souvenir headline.
The small text is grammatically correct Korean. Package labels say "맛김" (seasoned seaweed) and "재래김" (traditional seaweed) — realistic details.
Nitpick: ₩35,000 (about $25) for a seaweed gift set? That's a bit steep. Also, the AI added another unsolicited promotion: "Buy 3+ sets, get free seaweed flakes." AI, stop trying to upsell.
Fix: Specify a price range in the prompt if needed.
Score: 13/15 🟡 Adjust the price, then good to go
Stage 4: 🇻🇳 Vietnamese — Vietnamese Coffee
I've lived in Vietnam for 6 years. This is where I judge with near-native eyes.
Vietnamese has diacritical marks (dấu) on top of diacritical marks. The letter "a" alone has 17 variations: à á ả ã ạ â ầ ấ ẩ ẫ ậ ă ằ ắ ẳ ẵ ặ. Can the AI tell them apart?
Yes. It can.
It even generated 3 products in columns: "CÀ PHÊ TRỨNG HÀ NỘI" (Hanoi Egg Coffee), "SET PHIN CÀ PHÊ LƯU NIỆM" (Souvenir Phin Coffee Set), "CÀ PHÊ GÓI DẠNG PHIN GIẤY" (Drip Bag Coffee). These are the actual top 3 coffee souvenirs tourists buy in Vietnam. The product selection is too accurate.
I checked the diacritical marks character by character. ò, ệ, ữ, ủ, ắ — all correct positions, correct marks.
The background shows "QUÀ LƯU NIỆM HỘI AN" (Hoi An Souvenir Shop) with lanterns. If you've been to Hoi An, you'd recognize this vibe instantly.
Nitpick: 150,000 VND (~$6) for an egg coffee gift is a bit cheap for a tourist area. And 3 products on one pamphlet is a lot of information.
Fix: "Focus on a single product with a simpler design" would make it cleaner.
Score: 14/15 🟢 Approved by a Vietnam resident
Stage 5: 🇹🇭 Thai — Mango Sticky Rice
Thai script is a 3-story building. Consonants on the first floor, vowels on the second, tone marks on the third. Change the first floor and the upper floors shift.
I honestly expected things to start falling apart here.
They didn't fall apart.
"สุดยอดของฝากจากเมืองไทย" (Thailand's Best Souvenir) — the headline Thai script looks accurate.
Small text: This is where the 7-language test showed the most variance. When the font size gets smaller, the distinction between vowel marks and tone marks becomes blurry. The text isn't "broken" — it's "hard to read."
Thai script's stacking nature means small font sizes are particularly dangerous. This might be a fundamental limitation rather than a NanoBanana2 problem.
Nitpick: ฿280 (~$8) for mango sticky rice souvenir is actually reasonable. Credit where it's due — AI has good price sense.
Fix: "Use larger font sizes for all text" — this one instruction could fix the readability issue. Especially critical for Thai.
Score: 11/15 🟡 Increase font size, then usable
Stage 6: 🇮🇳 Hindi — Masala Chai Gift Set
Devanagari script. A horizontal line (shirorekha) runs across the top of each word, connecting all the letters. When consonants cluster, they merge into "conjunct characters" — two letters fusing into an entirely different shape. Even humans mess this up.
This was the biggest surprise of the entire test.
"शाही मसाला चाय उपहार सेट" (Royal Masala Chai Gift Set). The shirorekha connects correctly.
Small text lists ingredients with terrifying accuracy:
- "शुद्ध असम चाय की पत्तियां (१०० ग्राम)" (Pure Assam tea leaves, 100g)
- "६ प्राकृतिक मसाले (इलायची, अदरक, दालचीनी, लौंग, काली मिर्च)" (6 natural spices: cardamom, ginger, cinnamon, clove, black pepper)
It listed 6 spices by their individual Hindi names. And the numbers use Devanagari numerals (१, २, ६) instead of Arabic numerals — the correct choice for Hindi-language content.
Conjunct characters like "प्रा", "स्वा", "ग्रा" are all rendered in their correct combined forms.
Nitpick: The shop name "सोनी गिफ्ट हाउस" (Sony Gift House) might remind people of a certain electronics company. And Delhi's Main Bazaar is definitely not this photogenic.
Score: 14/15 🟢 The Devanagari accuracy is genuinely impressive
Stage 7: 🇸🇦 Arabic — Dates (Najma Palm Fruit) [FINAL BOSS]
Final stage.
Arabic is written right to left. The entire layout reverses. Letters never stand alone — they change shape based on their position (beginning, middle, end of a word). Same letter, four different forms.
Arabic speakers also have extremely high standards for text aesthetics. There's a thousand-year calligraphy tradition. "Sort of readable" doesn't cut it.
Let me be honest: if this one falls apart, nobody would blame the AI.
It didn't fall apart. The final boss has been defeated.
"هدايا تمور فاخرة من المملكة" (Luxury Date Gifts from the Kingdom). RTL is completely correct.
It generated 4 products in columns arranged right to left. Even the column order follows RTL. I did not expect this level of RTL commitment.
Small text: The ligatures (letter connections) look natural. Arabic characters change form based on position (initial, medial, final, isolated), and from what I can see, the forms are correct. However, I'll be honest — I can't read Arabic. The shapes look right, but I can't be 100% certain about the grammar.
Nitpick: This needs a native Arabic speaker to do a final check. My "looks correct" judgment is worth exactly as much as a non-Arabic-speaker's opinion.
Fix: For any language you can't read, always get a native check. But the structure, direction, and design are all pointing the right way.
Score: 13/15 🟡 Get a native check, then it's ready. The boss is defeated
Final Scoreboard
| Country | Souvenir | Text | Design | Usable? | Total | Verdict |
|---|---|---|---|---|---|---|
| 🇯🇵 Japan | Matcha KitKat | 5 | 4 | 5 | 14/15 | 🟢 Ship it |
| 🇨🇳 China | Dragon Well Tea | 4 | 5 | 4 | 13/15 | 🟡 Remove fake address |
| 🇰🇷 Korea | Seaweed Set | 4 | 5 | 4 | 13/15 | 🟡 Price too high |
| 🇻🇳 Vietnam | Coffee | 5 | 4 | 5 | 14/15 | 🟢 Resident approved |
| 🇹🇭 Thailand | Mango Rice | 4 | 4 | 3 | 11/15 | 🟡 Make text bigger |
| 🇮🇳 India | Masala Chai | 5 | 5 | 4 | 14/15 | 🟢 Best surprise |
| 🇸🇦 Saudi | Dates | 4 | 5 | 4 | 13/15 | 🟡 Need native check |
Key Takeaways
1. Headlines: Perfect Across All 7
Every language produced accurate, natural-sounding headlines. Kanji, Hangul, diacritical marks, Devanagari, Arabic script — nothing collapsed. This alone is a massive improvement from where AI image generation was a year ago.
2. Small Text: That's Where the Differences Show
Japanese, Vietnamese, and Hindi stayed accurate even in small print. Thai struggled with readability when font size decreased due to its stacking nature. Adding "use larger font sizes" to the prompt would likely fix this.
3. The AI Understands Commerce
Without being asked, it generated "buy 2 get 1 free" promotions, loyalty discounts, tax-inclusive pricing, and package deals. NanoBanana2 didn't just generate text — it generated sales strategy.
4. Quick Prompt Fixes for Common Issues
| Problem | Add to Prompt |
|---|---|
| Background shop showing | "White background, pamphlet image only" |
| Small text readability | "Use larger font sizes for all text" |
| Fake addresses | "Don't include addresses or phone numbers" |
| Too much info | "Focus on a single product, simple design" |
Bonus: The Header Image Was Also Made by NanoBanana2
I asked it to generate "a UN conference investigating AI-generated advertisements."
Pretty good, right? The conference hall atmosphere, colorful pamphlets scattered across the table, "AI-GENERATED ADVERTISEMENT INVESTIGATION" on the big screen. Serious delegates in suits scrutinizing tourist brochures. Exactly what I asked for.
But look closely.
The person sitting at the "JAPAN" nameplate appears to be Indian. The person at "INDIA" appears to be East Asian.
NanoBanana2 can write perfect pamphlets in 7 scripts, but it still can't match country nameplates to the right people.
...So it almost caused an international incident after all.
Conclusion
I asked NanoBanana2 to make souvenir pamphlets in 7 of the world's hardest scripts.
No international incidents occurred.
I went into this fully expecting at least one "💀 would cause complaints if distributed" verdict. The lowest score was 🟡 (usable with minor fixes). Even Arabic — the final boss with its right-to-left layout and position-dependent letter shapes — was handled correctly.
Three things genuinely surprised me:
- Hindi — Conjunct characters, connected shirorekha, Devanagari numerals. All correct. The biggest "I did not see that coming" of the test
- Vietnamese — I judged this with 6 years of residency. 17 variations of "a" and it nailed them all
- The universal sales instinct — The AI doesn't just write text. It writes marketing copy, calculates taxes, and invents promotions. It learned to sell
One thing I know for certain:
When NanoBanana3 comes out, I'll make 7 more pamphlets. With harder languages. Georgian. Tibetan. Khmer.
Until then, I'll be in Vietnam, drinking coffee and waiting.
All tests were conducted using NanoBanana2 (Gemini 3.1 Flash Image) in March 2026.
Every image shown is the first generation — no retries, no cherry-picking.
I have not performed native speaker verification for any language except Vietnamese. If you spot errors in your native language, please let me know in the comments.
Top comments (0)