Why WeChat’s Translation Glitches Reveal Hidden AI Challenges

WeChat translation anomaly

When the name caixukun was entered into WeChat’s translation feature, the system produced unrelated Chinese words. The WeChat team explained that the error was not an intentional easter‑egg but a mis‑translation caused by the engine’s inability to handle informal or unseen English tokens.

Technical cause

Modern neural machine translation (NMT) models typically generate the most probable token for each position. If a token is out‑of‑vocabulary (OOV) or never seen during training, the model must choose a high‑probability substitute, which can be nonsensical. A copy mechanism —which allows the decoder to copy the source token directly into the target sentence—can prevent such forced substitutions.

Uncertainty in machine translation

Uncertainty can be divided into two categories (see Analyzing Uncertainty in Neural Machine Translation , arXiv:1803.00047):

Intrinsic uncertainty : a single source sentence may have multiple valid translations due to variations in voice, tense, function words, or stylistic choices.

Extrinsic uncertainty : noisy or low‑quality training data, OOV words, and imperfect preprocessing can lead to hallucinations and over‑confident probability distributions.

Even models equipped with a copy mechanism can suffer from overly broad probability distributions, which degrades calibration quality.

Similar phenomena in other systems

Google Translate has exhibited comparable hallucinations. For example, feeding a long sequence of the English word “dog” while translating from Māori to English produced a biblical‑style output. Such behavior is attributed to the same encoder‑decoder architecture and to training on noisy, heterogeneous corpora.

Background of NMT

The encoder‑decoder paradigm was introduced in 2013 by Kalchbrenner and Blunsom, who combined a convolutional neural network (CNN) encoder with a recurrent neural network (RNN) decoder. While this architecture sparked rapid progress, it also introduced challenges:

Slow training and decoding.

Inconsistent translation style for the same source token.

Out‑of‑vocabulary (OOV) handling.

Limited interpretability of the black‑box model.

Heavy reliance on large parallel corpora of varying quality.

Mitigation strategies

Improve the quality and domain coverage of parallel training data, especially for low‑resource languages.

Integrate a copy mechanism (or pointer‑generator) so that unknown or proper‑name tokens can be copied verbatim.

Apply uncertainty‑aware decoding techniques that account for intrinsic ambiguity (e.g., n‑best lists, diverse beam search).

Filter or clean noisy web‑crawled corpora to reduce extrinsic uncertainty.

Adopting these measures can reduce translation “car‑crashes” like the WeChat incident and produce more reliable outputs across diverse language pairs.