I'm usually the grumpy one when AI comes up in creative work. After 13 years of shipping software and leading teams, I've seen plenty of tools that were sold as "help" and turned out to mean "fewer people". So when I saw that CD Projekt had used AI on Geralt's voice, I expected to be annoyed.

I wasn't. Here's why, and also where I think this could still go wrong.

What happened

Jacek Rozenek has voiced Geralt of Rivia in the Polish Witcher games since the first one. In 2019 he had a stroke that left him unable to speak for months. During rehab he trained his voice for six or seven hours a day.

This week CD Projekt Red's localization producer Mikołaj Szwed explained to Game File how the studio handled Rozenek's lines for the upcoming Witcher 3 expansion, Songs of the Past, which is due in 2027. The reporting I read is from Dexerto, Heise and PC Gamer, whose headline calls it "a special case".

The short version:

  • Rozenek recorded every new line himself.
  • The team first checked how much his voice had changed.
  • With his approval, the AI voice company Respeecher trained a model on his original Witcher 3 lines.
  • The model adjusts the new recordings so they sound close to the Geralt players heard in the main game.

Rozenek told Game File the goal was for the voice to be "as close as possible to my voice from 12 years ago", and that he has "great trust" in the people making the game. He also admitted the session left him exhausted.

The studio's worry was simple. Players would hear old and new Geralt lines next to each other, and a noticeably different voice would break the spell.

Why I'm mostly fine with this

Think of it like a photographer touching up lighting on a portrait. The subject still sat for the picture. The tool only fixes what the conditions couldn't.

Three things make this case different from the AI stories that make me wince:

The person did the work. This isn't a synthetic Geralt reading a script. Rozenek acted every line. The AI adjusts the sound, not the performance.

Consent came first. He agreed, and he talks about it openly. Nobody found out from a leak.

The studio said it out loud. In a week when the Untitled Goose Game studio got grilled for using a chatbot on a tiny bit of code, CD Projekt chose to explain itself in public. That's how you earn trust: tell people before they find out on their own.

This also isn't new for them. In 2023, Cyberpunk 2077: Phantom Liberty used Respeecher to recreate the voice of the late Miłogost Reczek, with his family's blessing. It's a pattern, and so far a careful one.

Where it could still go wrong

I don't want to pretend the other side is silly. It isn't.

The risk isn't this case. The risk is what this case makes normal. Once "we adjust voices in post" is a working pipeline, it's tempting to use it when the reason is a tight deadline rather than a health problem. Or a cheaper actor. Or just "close enough".

I've seen the software version of this many times. A tool brought in for one hard case ends up as the default for every case, because it's there and it's cheaper. Nobody decides that on purpose. It happens over a dozen small meetings.

The fix is boring and it works:

  • Consent per project, not buried in a contract. "Yes, for this expansion, for this reason."
  • The human performance stays the source. AI adjusts, it doesn't generate.
  • Disclosure by default. If you'd be embarrassed for players to find out, don't do it.

The rule I use with my own team

When someone on my team proposes a new tool, I ask one question: does it help a person keep doing their job, or does it quietly remove the person?

CD Projekt's case is the first kind. A man fought his way back to speaking and still showed up to play a character he's voiced for most of his career. The tool let him finish the job. I'm fine with that.

I'd also like a version that makes me sound awake at Monday standups, but I suspect that's out of scope.

Would you be OK hearing this in a game you love, or does any AI voice work cross a line for you? Why?