Myths about quality estimation and automatic post-editing
June 25, 2026
- At GenAI in Localization, ModelFront CEO Adam Bittlingmayer shared five myths about quality estimation and automatic post-editing, and the truth behind each one.
Myth #1: QE is bullshit.
Truth
QE has been successfully verifying machine translations behind the scenes for years. You've probably seen it out in the real world, in Fortune 500 content, from top travel platforms to biotech patents — you just didn't know it, because it keeps human quality.
Why does this myth exist?
QE has a bad reputation because the copycat QE features, which are easy to try out, are just random number generators and not even customized at all.
Myth #2: You can use raw QE scores directly.
Truth
Dumping raw scores into a workflow forces the localization team or CAT user to guess what the numbers mean and decide what to do next. A legitimate QE provider takes responsibility for the outcome — millions of words automated — by turning those underlying numbers into clear boolean decisions: approve the segment as-is, or send it to a human for review, under transparent monitoring.
Why does this myth exist?
For providers just adding a copycat QE feature, it's easier to generate a pseudoscientific score and provide it as-is than it is to build and maintain a system that actually works to safely automate a live production workflow. And scores seem scientific.
Myth #3: QE is for evaluation.
Truth
QE for offline evaluation is a low-value use case. The high-value use case for QE is approving or rejecting segments in live production workflows, while keeping human quality.
Why does this myth exist?
QE can be used to compare translation engines or to decide if the MT output is good enough, but evaluating manually isn't that slow, expensive, or frequent. Offline evaluation solves the builder's occasional problem — scoring thousands of segments. Online quality prediction in production solves the buyer's enterprise-scale problem — actually automating millions or tens of millions of words.
Myth #4: APE is for reducing edit effort.
Truth
The purpose of APE is actually to increase the number of segments that proper QE can automate. Reducing edit distance for segments doesn't significantly reduce cognitive effort for translators, so it doesn't help buyers. The key is reducing the number of segments that require manual checks at all.
Why does this myth exist?
Since the beginning of post-editing, the idea was that manually post-editing machine translations would be faster than translating from scratch. A decade or two ago, machine translation output was often so bad that it even caused more work. But beyond a certain point, lowering post-editing effort is like pushing on a string. The bottleneck is reading and checking every word, both the source and the target, whether the segment is edited or not.
Myth #5: QE + APE is replacing humans.
Truth
Making translation radically more efficient will let companies translate far more content into far more languages — Jevons paradox. Translating with today's legacy technology and processes is so painfully slow and expensive that companies can only afford human-quality translation for less than one percent of their content, into a handful of language markets.
Why does this myth exist?
The translation industry only sees a fixed pie — the limited demand that exists for today's inefficient offering: ten to twenty cents for every new word, turnaround times up to two months, opacity and lock-in. But the translation industry only grew after the rollout of previous technological efficiencies, like translation memory, translation management software and post-editing. Zero-sum thinking about efficiency is so common over the centuries and across industries, from agriculture to coding, that it has a name: the lump-of-labor fallacy.
Watch the full session on YouTube → (opens in a new tab)