I am incredibly honoured to be co-writing this series with Angela (Yang) Liu, Senior Localization Project Manager at Blizzard Entertainment. Her Part 1 named something the localization industry had been billing for years without a vocabulary. This is Part 2, from the other side of the brief.
In late March 2026, the English-language editor on Kingdom Come: Deliverance II opened LinkedIn and wrote a sentence the localization industry had been quietly waiting for without knowing it. His name is Max Hejtmánek. He had been let go from Warhorse Studios after joining in 2022. The reason given to him, by his own account, was that his role had been made “obsolete in favour of using AI for all translations going forward.” He had been vocally opposed to AI translation at the studio. Reddit moderators verified his identity. His LinkedIn confirmed the timeline. The studio’s official response did not directly deny the allegation.1
The sentence did not cost him the job. He had already been let go; what the sentence cost him was the silence. And without him planning it that way, it became the first publicly named, dated, AAA-anchored data point in a story the field had been telling itself in private for three years. Senior linguist. Flagship title. Replaced by a pipeline. Said so out loud.

Until that post, the industry could still describe what was happening as “trends” or “consolidation.” Once it has a person attached, you stop calling it a trend.
Angela’s Part 1 named the thing. Localization debt. The 1x cost at requirements, the 10x at design, the 100x at production, applied to the field that had been carrying that curve invisibly since the first multilingual game shipped. She did a job that needed doing.
I want to write the part she could not, and that her piece could not have included, because it sits underneath her framing rather than next to it.
Localization debt now has a lender. The lender is generative AI. And the bill being mailed for the credit drawn at 1x is being delivered to a different desk than the one that opened the account.
That changes the shape of the entire problem.
A pipeline that arrived faster than the industry’s vocabulary
The 2024 European Language Industry Survey put actual machine-translation use at 40 percent among language service companies and 33 percent among independent translators, up from 31 and 29 percent the year before. AI tools for language tasks made their first real appearance in the data at 11 percent actual use. ELIS’s own commentary on the curve was blunt: “at this pace, MT is expected to be used in a majority of professional translation work by next year.”2 Acolad’s 2025 translators survey, fielded later, already records 59 percent of freelance translators using neural MT, 79 percent reporting familiarity with AI tools, and 84 percent expecting demand for human translation to decrease as post-editing demand rises.3 Whatever the development side of the industry believed about gradual AI adoption, the loc side is living through the inflection right now.
What that looks like in production was made painfully visible at the end of June 2025. 11 bit Studios shipped The Alters, a narrative single-player game the team had cared deeply about. Players found two separate AI slips in the shipped build. On an in-game terminal that should have been silent in-fiction lore, the screen displayed a literal ChatGPT response: “Sure, here’s a revised version focusing purely on scientific and astronomical data:” followed by the Interstellar Exploration Log itself. In a Brazilian Portuguese subtitle on a licensed in-game movie, another ChatGPT prefix preceded its translation. The studio apologised publicly, calling the decision to use undisclosed AI “the wrong call.”4

The defect was not subtle. It was English text in a Portuguese build. Anyone on the team who opened that build could have spotted it. It still shipped. Which tells you exactly what the Portuguese build’s human review pass actually consisted of, which is to say nothing. And if the obvious English-language defect slipped through that gap, the question worth asking is what is happening to the subtle Korean register defaults, the Japanese honorific choices, and the Chinese name decisions, where the defect cannot be spotted by anyone who does not speak the target.
The tool vendor Gridly writes plainly on its own marketing site that “if there is any mistake made in the pivot translation, it will be reproduced in all the target languages.”5 That is the pipeline working as designed. The mistake is not a bug. It is a feature being executed correctly.
The math compounds along two axes, not one. Boehm’s 1x to 10x to 100x progression is temporal: the same defect, found at later stages, costs exponentially more to fix. The AI pipeline adds a spatial axis. The same defect ships in every supported target language at once. For a typical AAA loc footprint of twelve to fifteen languages, the cost at LQA is not Boehm’s 10x. It is Boehm’s 10x multiplied by your supported-language count, and it compounds with every fault the pipeline distributes. The cost-of-defect curve Angela borrowed from Boehm in Part 1 was always going to apply to localization.15 It just took a lender to make it real, and a pipeline to fan it out.
The part of all this that should keep procurement teams awake at night is that the same systems producing the defects also produce the quality scores used to grade them. A peer-reviewed evaluation of LLM-as-judge metrics on literary translation, published on arXiv in 2024, found that GEMBA-MQM, one of the most cited AI grading systems, preferred professional human translations over LLM output 9.6 percent of the time. Professional human evaluators, looking at the same pairs, preferred the human translations 94.4 percent of the time.6 The metric and the human are looking at different worlds. The metric is the one in the dashboard. The dashboard is the one being shown to the publisher.

The seat that used to be empty has been filled. The occupant does not speak the target languages.
Angela’s fourth recommendation, the coffee with the dev team, is really a quieter argument about a role. She writes that someone needs to be the person who can carry the entire localization picture at once: source changes, translation in flight, engine constraints, deadline pressure. She tells the loc team this person should be them. Angela and I gave that role a name. The Conductor. For most of the field’s history, it has either been distributed across three or four people who never sat in the same room, or it has been missing entirely.
It isn’t missing anymore. The Conductor function in 2026 has a desk and a job title. The job title is some variant of “Prompt Engineer (LLM Automation for Data Labeling and Localization),” and Innodata is actively hiring for that exact role.7 The role’s stated responsibility is to design and implement the prompt strategies that drive the localized translation pipeline. Required knowledge includes “localization best practices and cultural nuances for different languages and regions.” Native target-language proficiency in any of those languages is not on the requirements list.

That role has more authority over the localization output than any localization PM ever had. They configure the model. They write the prompts. They decide what gets human-reviewed and what does not. They also, by the job description, do not speak the languages they are conducting.
This is not a complaint about engineering. It is a structural observation about how the vacuum got filled. The Conductor seat used to be empty because no one was willing to take it. It is now occupied because the tooling forced someone to take it, and the someone who took it is the someone closest to the tooling. Angela’s call is for the loc team to step into the seat themselves. The trouble is that by 2026, someone else had already taken it. Just not anyone who speaks the languages they are conducting.
The voice side of the industry, at least, has put terms in writing about all of this. The SAG-AFTRA video game contract, ratified by 95 percent of voting members in July 2025 after an eleven-month strike, spent most of its duration negotiating AI consent, disclosure, and digital-replica provisions.8 Voice actors have AI in their contract. The translation side has nothing equivalent.
Angela’s four recommendations, one year later
This is where I would have agreed with Part 1 a year ago and where, looking at the field as it actually is, I have to disagree.
Angela recommended that internationalization come before localization, in the form of font selection, IME support, string metadata, an i18n handbook the dev team will actually read. She is right about the discipline. The AI pipeline, however, assumes you have already done it. If you have not, the pipeline ships defects at scale and signs them off as “high quality.” Coverage of Black Myth: Wukong’s localization documents what happens when the upstream is rushed: Game Science publicly admitted the team had “a lot to learn” after underestimating the time needed to localize into eleven languages, and early reviewers shared screenshots of in-game journal entries about the various yaoguai still appearing in untranslated Chinese in the supposedly multilingual build.9 That was a human pipeline missing its upstream. An AI pipeline missing its upstream produces the same defects, faster, and across more targets simultaneously.
She recommended data-backed arguments to move budgets. The trouble in 2026 is that the data is generated by the same system that produces the defects. If you build your dashboard on top of the model that translated the strings, you are reading the model’s opinion of its own work and calling it a metric. Almost no studio has a measurement layer that sits independently of the thing being measured. The data muscle Angela called for is harder to build now, not easier, because the room has been re-wired.
She recommended treating localization as an asset rather than a cost center. Larian’s Swen Vincke is one of the very few AAA studio heads who has publicly positioned human-language work as a differentiator. Larian is, in his words, “hiring writers, not having ChatGPT write their dialogues.”10 The fact that this counts as a noteworthy stance, rather than the default, is itself a measurement of where the industry has moved.
She recommended the coffee with the dev team. The relationship that makes the rest of the system work, the human connection across the source-language and target-language sides. The 2026 problem with the coffee recommendation is that in many studios, there is no longer a linguist at the table. Brian Merchant’s August 2025 investigation into translator economics documents earnings collapses of 60 to 80 percent compared with the 2020-2023 baseline among named freelancers, with one translator noting that machine-translation post-editing now “pays a quarter of what you’d get for translation work” and others reporting per-word rate collapses from $0.03 to $0.05 down to roughly $0.01.11 The 2023 CIOL / Inbox Translation Freelance Translator Survey found that 31 percent of translators flagged “the increasing use of machine translation” as a main challenge of the profession, up from 27 percent in 2020.12 The coffee Angela wanted the loc team to have with the dev team is now being had between the prompt engineer and the producer. The linguist is a downstream API consumer at one-quarter pay, and increasingly is not even invited.
That line obscures something important: the collapse is not uniform. The cross-sectional rate data shows a clear gap between members and non-members and between specialists and generalists, and the AI pressure documented above is asymmetric: it compresses the commodity end of the market while leaving creative and specialised work largely untouched. We do not yet have a longitudinal Walker series to prove the gap is widening over time, but the direction of pressure points one way. Callum Walker’s January 2024 ITI / Cardiff University rates study shows the existing divergence in the UK cleanly. Members of the Institute of Translation and Interpreting charge a median £0.075 per word for general texts via agencies, £0.100 for direct clients on general work, and £0.110 for specialised work to direct clients. Non-members sit at £0.062 (general, agency), £0.080 (general, direct), and £0.088 (specialised, direct). That is a 13 to 25 percent per-word premium at the median, and the non-member interquartile ranges run considerably wider, meaning the bottom of the non-member distribution reaches much further down than the bottom of the member distribution. Membership and specialisation pull rates up. Their absence exposes the translator to AI compression with no rate floor underneath.16 The trade press tells the same story across other markets: commoditised general translation is being squeezed by AI from one direction, while specialised, certified, and creative translation is holding or rising from the other.
I can speak to this from the senior end. The rates I charge in 2026 for CJK game and entertainment work are unchanged from the rates I charged before the AI wave: 0.10 EUR per English source word, 0.12 EUR per Japanese character. The gap between those rates and the one-cent-per-word MTPE floor Merchant documents is roughly a tenfold spread on like-for-like creative work, and that spread did not exist in 2019. AI is not killing translation. It is eating the commodity middle and pushing the profession into two camps: the linguists who can be replaced and the linguists who cannot. The steady mid-tier freelancer who carried a respectable career on general content is the part of the market that is disappearing into one end or the other.

None of this is an argument against AI in localization pipelines. AI in a localization pipeline with a competent human review layer is a tool that works; I use one every day in my own practice. The failure mode the field is heading into is not “AI exists.” It is “AI shipped without anyone left in the loop to catch what it gets wrong.” The Alters did not break because the pipeline used AI. It broke because no one on the team opened the Portuguese build before ship. The studios already running AI alongside an intact human QA pass do not surface the defects this piece documents. That is the version of the curve worth aiming for.
What I see from my own desk
The three languages I work in are also the three where this lands earliest.
In Japanese, every line of dialogue is a register decision. A linguist who does not know which character is speaking has to pick one, and most of the time they will pick correctly because they ask. An AI that does not know picks the most statistically frequent register and ships it across the whole game. Post-recording, the fix is not a re-translation. It is a re-cast. I have watched that conversation happen at a publisher. It is not a cheap conversation.
In Chinese, a character name is a world-building decision before it is a translation decision. Hero names, place names, faction names. The cost of getting these right at upstream is a conversation. The cost of getting them wrong at downstream is rebranding a character who has already shipped on a billboard in Shanghai.
In Korean, register is grammar, not style. There is no neutral default. An AI translator picking between banmal (the informal speech used with peers, juniors, and intimates) and jondaemal (the formal speech used with elders, superiors, and strangers) without speaker context is committing a defect that surfaces in voice direction, then in QA, then on X the week the game ships.
In all three languages, these defects are not subtle. A grandfather speaking like a koukousei, a queen addressing a servant in the form she would use for a kindred peer, a young apprentice using elder honorifics to a junior: these are not minor stylistic quirks. They are the localisation equivalent of dressing a samurai in tracksuit bottoms. The native speaker sees it immediately. The LLM judge does not. That is the actual gap the literary-translation metrics paper measured, and it is the gap that has trained Chinese-language Steam reviewers to organise.
The Chinese consumer base has already moved first.13 Chinese Game Loc maintains a public list of recently review-bombed titles where bad localization or translation errors were a cited reason, including Stardew Valley after its 2024 Chinese update changed the font and retouched the translation, Hollow Knight: Silksong after its translators imposed what the article calls an “overly literary, pretentious style,” and Kingdom Come: Deliverance over machine-translation errors and a clear lack of understanding of game world and story context.14 The Japanese and Korean equivalents are not coordinated yet. They will be.

Three predictions I expect to lose money on if I am wrong
The public dataset of named AAA AI-loc failures is still thin: one disclosed defect cluster (The Alters), one disclosed termination (Hejtmánek). The predictions below extrapolate from those two cases plus the structural pressure documented above. If I am wrong, it will be because the surfaced data is not representative of what is actually shipping.
The first major AAA title that publicly de-AIs its localization will use the decision as a marketing differentiator, the way “no microtransactions” became one. Larian has effectively pre-announced the move. The press release, when it lands, will be more triumphal than the localization itself warrants. It will still work, because the player audience in CJK markets is already reading AI output and knows what it tastes like.
Review-bomb cycles in CJK markets will start to correlate with AI-pipeline disclosure. The studios that disclose AI use first will become the test cases. The studios that do not disclose will be retrofitted to the curve by community OSINT. The 11 bit / Alters case is the template. A player spotted a literal ChatGPT response in the localised text, posted a screenshot, and the disclosure happened whether the publisher wanted it or not.
And the linguists currently being asked to “review” AI output at a quarter of their old rate will quietly walk away from the industry. The seniority gap will surface two or three years later as a quality cliff that no procurement spreadsheet predicted. Hejtmánek is the first public data point. There will be many more. Most will go unannounced. By the time the spreadsheet sees the cliff, the senior translators who would have caught the defects at 1x will already be doing something else.
Each of these is a 100x bill arriving for a 1x credit that someone drew cheerfully eighteen months earlier.
The bill is in the mail
Angela closed Part 1 on a line I want to borrow: localization is only a debt when it is underinvested. I will add the 2026 footnote. Localization is now being underinvested at industrial scale, on credit, with the linguist underwriting the loan.
The good news, if there is any, is that the people who will be holding the bag when the bill arrives are also the people who can read it first. The localization PMs. The freelancers. The in-context QA leads. The small studios that have not fully outsourced their pipeline yet. Anyone reading this is positioned to see the defect cluster forming before it surfaces on a review aggregator.
That visibility is the asset. Not the AI pass. Not the audit report. Not the dashboard. The pair of eyes that has watched the localization field long enough to recognise the shape of debt being industrialised in real time, and the willingness to say so before the screenshot lands on the wrong subreddit.
Hejtmánek said it. The Alters’ in-game terminals said it. The Chinese review aggregators are saying it. The bill is in the mail. Better to read it before it lands on the player’s desk.
Sources
- Joshua Wolens, “Kingdom Come: Deliverance 2 translator says he was fired from Warhorse Studios and replaced with AI,” PC Gamer, 28 March 2026. link
- European Language Industry Survey 2024 (ELIS), public results. link
- Acolad, “AI in Translation: Key Findings from Acolad’s 2025 Translators Survey,” 2025. link
- Bryant Francis, “11 bit apologizes for not disclosing gen AI use in The Alters,” Game Developer, 30 June 2025. link
- Gridly, “Pivot Language Localization.” link
- Ran Zhang, Wei Zhao, Steffen Eger, “How Good Are LLMs for Literary Translation, Really? Literary Translation Evaluation with Humans and LLMs,” arXiv, 2024. link
- Innodata, “Prompt Engineer (LLM Automation for Data Labeling and Localization)” job listing, LinkedIn, 2026. link
- Jennifer Maas, “Video Game Actors Strike Officially Over After SAG-AFTRA Ratifies Contract,” Variety, 9 July 2025. link
- “Lost in Translation: The Localization Challenge of Black Myth: Wukong,” The World of Chinese, October 2024. link
- Joshua Wolens on Larian’s Swen Vincke, PC Gamer, 2024. link
- Brian Merchant, “AI Killed My Job: Translators,” Blood in the Machine, 21 August 2025. link
- Chartered Institute of Linguists / Inbox Translation, Freelance Translator Survey 2023. link
- PC Gamer, “Steam review-bombing is working, and Chinese players are a powerful new voice.” link
- Chinese Game Loc, “Where Chinese review-bombing comes from: 10 games that got hit the hardest.” link
- Barry W. Boehm, Software Engineering Economics, Prentice-Hall, 1981. link
- Callum Walker, “Freelance Translator Rates: The Value of ITI Membership,” Institute of Translation and Interpreting / Cardiff University, January 2024. link


Leave a Reply