Krafton began a closed test of generative dialogue for non-player characters in PUBG’s training mode, limiting the feature to offline lobbies while engineers tune latency and toxicity filters ahead of any ranked rollout.
Players invited through the PUBG Labs client can issue voice or text prompts to practice bots that respond with situational callouts rather than canned barks. Krafton said the models run on regional inference clusters in South Korea and Singapore, with prompts and outputs logged for moderation review. Ranked matchmaking and esports servers are excluded from the build.
Studio goals for the test
PUBG’s training island has long used scripted bots for recoil practice. Krafton wants bots that mimic squad communication—calling flanks, requesting meds, warning about third parties—without handing casual players an unfair information edge in live matches. Product director Kim Min-jae told Korean press the test measures whether large language models can stay within game lore and map boundaries when players try to jailbreak prompts.
Latency targets are stricter than for menu chatbots. Krafton is aiming for sub-300-millisecond round trips on fiber connections in Seoul, acknowledging that mobile tethering players may see delayed responses. When inference queues spike, bots fall back to legacy voice lines so sessions do not freeze.
Safety and rating considerations
The Korea Game Rating and Administration Committee received an updated technical document describing text filters for slurs and sexual content, plus a reporting button tied to Krafton’s existing griefing ticket system. Parents’ groups had questioned AI dialogue in teen-rated shooters after overseas titles experimented with unfiltered NPC chat; Krafton’s filing stresses human review sampling on 5 percent of sessions during the test window.
Data handling is domestic-first. Voice clips are converted to text on Korean servers and deleted within 30 days unless a player files a harassment report, according to the privacy addendum Krafton emailed to testers Sunday night.
What rivals are doing
Ubisoft and Electronic Arts have publicized generative NPC experiments, but few battle royale publishers have shipped voice-interactive bots at scale. Krafton’s advantage is a captive training population that opts in for faster match starts. If dialogue bots keep players in practice mode longer, Krafton could lift cosmetic sales tied to training challenges without touching core combat balance.
Investors will watch whether AI infrastructure spending shows up in margin guidance. Krafton capitalized cloud contracts earlier this year for anti-cheat machine learning; the NPC test reuses part of that capacity. Chief Executive Changhan Kim said on an August call that generative features must clear internal return hurdles before global release.
What players see this week
Testers receive a PUBG Labs banner explaining that bot lines may be nonsensical and should be reported. Korean-language prompts are prioritized in the first wave; English and Japanese packs follow if filters hold. Krafton has not announced a date to merge the feature into the main client, and league operators said custom esports builds will disable AI dialogue even if casual modes adopt it later.
For a studio whose fortunes still ride on a single franchise, the training-mode experiment is a low-risk petri dish. Success would modernize practice tools; failure would be invisible to ranked players still dropping into Erangel without a word from the bots shooting back.