[yorwba]

Any even-length alphabetic ASCII string decodes to random Chinese characters in UTF-16LE. Digits and = unlock some Japanese hiragana, Korean hangeul and assorted punctuation, but those only make up a small fraction of the total.

For example, 'backdoor'.encode('ascii').decode('utf_16_le') == '慢正潤牯', which Google Translate turns into "Slow and positive", but it's just nonsense.

[thmstcls]

I'm naive to the translation tech space but is this sort of thing unique to languages like Chinese? I figured all this stuff was mostly solved. Like I wouldn't expect dflhglsdhfgalskjdf to have Google Translate output some grammatically valid Spanish output.

[cvhc]

There is one difference between gibberish Chinese and Latin character sequences. In Chinese, each character indeed carry some meanings (like a word). So I guess the model may hallucinate some output inspired by these meanings. In the case "慢正潤牯" -> "Slow and positive", it actually translated the first two characters literally (慢 -> slow, 正 -> correct/positive/upright).

So equivalent English gibberish would be like "hast prank bibble done anut me me ions." Google translates this one to "对我而言，恶作剧已经完成了。" (To me, the prank has been done.) in Chinese -- very valid sentence, and "¿Me has hecho una broma a mí, Bibble?" in Spanish -- also seems valid.

I guess the model is (over) optimized to generate valid outputs. This can be a feature, so it still translates grammatically invalid but to some degree understandable text (like with typos or non-standard Internet language).

[yorwba]

I think the Latin script might be somewhat protected because random jumbles of letters do appear as serial numbers and whatnot, but for other scripts, anything goes.

I say ҏӲҨЏ ҜъКѠ ЇЩіН гӞэѷ in "Russian", Google Translate says "Let's talk about it".