Only 76 of the 2,502 demos that disclosed AI at Steam Next Fest (Valve’s online demo festival) in October 2026 reported generating content live while the game runs, against 941 that cited audio or voice (How To Market A Game, Next Fest Analysis 2026). That gap defines the topic: pre-rendered synthetic speech is spreading, while characters that truly improvise out loud remain rare. Among AI-disclosing Steam titles estimated above $1M, 27% cite voiceover or dialogue (Totally Human Media, GenAI Revenue Analysis 2025), yet 83% of surveyed core players were negative on AI-generated dialogue (Quantic Foundry, Gen AI Survey 2025). This roundup isolates the spoken line, while our wider AI in gaming statistics cover the rest of the development pipeline. We aggregated data from GDC, Google Cloud and The Harris Poll, Unity, Circana, YouGov, GameSoundCon, SAG-AFTRA, Sulka Haro’s Steam census, and peer-reviewed studies in PLOS One and Frontiers in Computer Science.
TL;DR
- 941 of 2,502 AI-disclosing Next Fest demos cited audio or voice in October 2026 (How To Market A Game, 2026)
- 3.0% of those disclosing demos reported runtime or live generation (How To Market A Game, 2026)
- 27% of AI-disclosing Steam titles estimated above $1M cite voiceover or dialogue (Totally Human Media, 2025)
- 24% of successful AI-disclosing games cite AI voice, against 8% of flops (Sulka Haro, 2026)
- 36% of game developers use AI for dialogue writing support (Google Cloud and The Harris Poll, 2025)
- 8% of game audio professionals use generative AI at work, and placeholder dialogue is the top use (GameSoundCon, 2025)
- 6,909 ms was the average speech-to-speech NPC cycle in one VR study (Korkiakoski et al., 2025)
- 83% of surveyed core players were negative on AI-generated dialogue (Quantic Foundry, 2025)
- 58% of voice clones were labelled human, against 62% of real recordings (Lavan et al., PLOS One 2025)
- 7.5x union scale is the 2025 contract minimum for a digital-replica-voiced game chatbot (SAG-AFTRA, 2025)
1. Developer Adoption: AI Writes the Dialogue Before It Speaks It
44% of Unity developers use AI for writing and narrative design (Unity 2026 Game Development Report, as reported by Digital Today), second only to coding support. That is a script-side number. As Wccftech notes in its report coverage, the data does not show whether AI-assisted work reaches the final game.
The Google Cloud and Harris Poll survey points the same way: dialogue help and NPC logic are mainstream experiments, and a smaller group still treats adaptive NPCs as a top opportunity. The question is no longer whether teams touch AI, it is whether any of that output is ever heard by a player.
| Metric | Value | Source |
|---|---|---|
| Unity developers using AI for writing and narrative design | 44% | Unity, 2026 Game Development Report |
| Unity developers using AI for coding support (for comparison with writing) | 62% | Unity, 2026 Game Development Report |
| Game developers using AI for dialogue writing support | 36% | Google Cloud / The Harris Poll, AI Meets the Games Industry 2025 |
| Game developers using AI agents for advanced NPC behavior | 34% | Google Cloud / The Harris Poll, AI Meets the Games Industry 2025 |
| Game developers naming NPCs that learn and adapt as a top future opportunity | 23% | Google Cloud / The Harris Poll, AI Meets the Games Industry 2025 |
| Developers reporting generative AI use, early 2026 vs early 2025 | 29% vs 36% | Game Developer Collective / Omdia, via Outlook Respawn 2026 |
| Generative AI use among staff at game studios vs staff at publishing, support, and marketing firms | 30% vs 58% | GDC, State of the Game Industry 2026 |
Context: Google Cloud surveyed 615 adults working in game development in the US, South Korea, Finland, Norway, and Sweden from June 20 to July 9, 2025, with a margin of plus or minus 3.9 percentage points at 95% confidence. The GDC 2026 summary drew on more than 2,300 professionals. The Omdia-backed tracker shows use falling, not rising, in its latest wave.
2. Steam Disclosures: Audio and Voice Tie Text as the Second-Largest AI Category
37.6% of AI-disclosing demos at the October 2026 Next Fest cited audio or voice (How To Market A Game, October 21, 2026), level with text and localization and behind only art. Dividing 941 by all 7,240 demos gives 13.0% of the event (our arithmetic).
Treat that as a ceiling for voice, not a count of voiced NPCs. The category is built by keyword matching on free text and combines audio with voice, so it is not limited to speech. Valve’s content survey groups “artwork, sound, narrative, localization, etc.” under the same disclosure, so nothing on a store page isolates dialogue.
| Metric | Value | Source |
|---|---|---|
| Next Fest demos with an AI disclosure, October 2026 | 2,502 of 7,240 (34.6%) | How To Market A Game, Next Fest Analysis 2026 |
| Disclosing demos citing audio or voice vs art vs text or localization | 37.6% (941) vs 75.7% (1,894) vs 37.7% (943) | How To Market A Game, Next Fest Analysis 2026 |
| Next Fest titles tagged for generative AI, June 2026 | 1,704 of 8,700 (19.5%) | SteamDB data via Eurogamer and Dataconomy, 2026 |
| New Steam games carrying an AI disclosure, 2024 / 2025 / 2026 to date | 10.9% / 19.9% / 30.8% | Sulka Haro census, via Games.gg 2026 |
| Steam titles disclosing generative AI, July 2025 vs November 2025 (share of the full catalog) | 7,818 (about 7%) vs 10,258 (about 8%) | Totally Human Media, 2025 |
Context: Haro’s census covers 53,597 releases from July 2023 to July 2026, and AI-flagged launches rose from about 13 a month before the January 2024 disclosure mandate to about 530 a month. June’s Next Fest figure comes from a Dataconomy report citing Eurogamer and SteamDB. Every Steam rate here is self-reported, so it is a floor.
3. Commercial Hits: Voice Is Over-Represented Among AI Games That Sell
27% of the 45 AI-disclosing Steam titles estimated above $1M cite voiceover or dialogue (Totally Human Media, November 2025). Haro’s separate sample lands close: 24% of successes disclose AI voice. Two independent methods, one signal.
Voice looks like a sign of scale, not a cause of sales. A game needs a large script and several markets before synthetic speech saves real money, and those are the games that already have an audience. Art runs the other way, appearing more often among flops, which fits its role as the cheapest shortcut for a small team.
| Metric | Value | Source |
|---|---|---|
| Among the 45 AI-disclosing Steam titles estimated above $1M gross: citing voiceover or dialogue vs in-game art vs translation | 27% vs 49% vs 9% | Totally Human Media, 2025 |
| AI voice in disclosures: successes (1,000 or more reviews, n=138) vs flops (under 50 reviews, n=400) | 24% vs 8% | Sulka Haro, Three Years of AI on Steam 2026 |
| Same successes vs same flops: localization, then AI art, images, or textures | 18% vs 6%, then 57% vs 72% | Sulka Haro, Three Years of AI on Steam 2026 |
| Demos with an AI disclosure, top 25 by peak concurrent players vs ranked 101 and below | 8% (2 games) vs 35% | How To Market A Game, Next Fest Analysis 2026 |
Context: the revenue brackets use the Boxleiter method (review count multiplied by price and a constant), which its author calls a rough estimate. The flops had been on sale for at least 9 months. Both samples are small and neither proves causation.
4. Inside the Audio Department: Placeholder Lines Come First
Only 8% of professional game audio respondents currently use generative AI in their audio work (GameSoundCon Game Audio Industry Survey 2025). Among those users, 43% cite placeholder dialogue, meaning temporary lines recorded during production and meant to be replaced. Multiplying the two gives about 3.4% of respondents (0.08 x 0.43, our arithmetic).
That is the realistic first job for synthetic speech: a scratch track, not a finished performance. Arc Raiders shows what happens when the line between the two blurs. CEO Patrick Soderlund admitted a “quality difference”, said “A real professional actor is better than AI; that’s just how it is”, and confirmed that some lines were re-recorded after launch (PCGamesN, March 14, 2026, citing a GamesIndustry.biz interview). The studio still licenses actors’ voices for text-to-speech on lines it considers less essential to immersion.
| Metric | Value | Source |
|---|---|---|
| Game audio professionals currently using generative AI in audio work | 8% | GameSoundCon, Game Audio Industry Survey 2025 |
| Game audio professionals exploring generative AI | 15% | GameSoundCon, Game Audio Industry Survey 2025 |
| Generative AI users citing placeholder dialogue, the most common use | 43% | GameSoundCon, Game Audio Industry Survey 2025 |
| Survey responses, and how many came from game audio professionals | 654, of which 631 | GameSoundCon, Game Audio Industry Survey 2025 |
| Line variants Embark’s CEO says text-to-speech can test before a recording session (his example) | 15 | PCGamesN, citing GamesIndustry.biz 2026 |
| Arc Raiders Steam user reviews at launch (the game discloses limited generative AI use) | Over 30,000, more than 90% positive | Totally Human Media, 2025 |
Context: the GameSoundCon survey ran from May 1 to June 15, 2025, was promoted by email and social media, and is not randomly sampled. Its respondents also named production voice-over, voice cloning, and voice changing among other uses, so shipped synthetic speech exists, just further down the list.
5. Live NPC Conversation: Rare in Games, Slow in the Lab
Runtime or live generation appears in just 3.0% of AI-disclosing Next Fest demos. Across the whole event that is 76 of 7,240, or about 1.0% (our arithmetic). Among $1M+ titles, runtime generative systems sit under 5%, and Haro calls genuinely AI-native games “a tiny minority” among about 9,400 flagged titles. Naavik’s review of the June event found none in the top 50 by wishlist increase.
Latency is a central reason. In one VR study, the full loop of listening, thinking, and speaking averaged nearly 7 seconds, and the 18 participants estimated it at 8.67 seconds. They still rated overall believability at 7.69 out of 10, which suggests patience in a novelty setting more than readiness for a shipped game. A 2026 Technical University of Munich study cut time to first audio to under 3 seconds, and our voice AI latency benchmarks show how far the wider industry has pushed the same metric.
| Metric | Value | Source |
|---|---|---|
| Disclosing Next Fest demos reporting runtime generation vs described as substantially AI-built | 3.0% (76) vs 0.3% (8) | How To Market A Game, Next Fest Analysis 2026 |
| AI-disclosing Steam titles estimated above $1M citing runtime generative systems | Under 5% | Totally Human Media, 2025 |
| Top 50 June 2026 Next Fest games by wishlist increase with AI-enabled gameplay | 0 | Naavik, Steam Next Fest’s AI Problems 2026 |
| Full NPC speech cycle in a VR study: average / median / maximum | 6,909 ms / 6,066 ms / 24,362 ms | Korkiakoski et al., arXiv 2025 |
| Average delay by stage: language model / text-to-speech / speech-to-text | 3,113 ms / 2,376 ms / 1,365 ms | Korkiakoski et al., arXiv 2025 |
| Usability of that VR system on the System Usability Scale (0 to 100) | 79.44 | Korkiakoski et al., arXiv 2025 |
| Time to first NPC audio in a VR study: short queries vs complex queries | 1.42 s (±0.2) vs 2.90 s (±0.6) | Berrezueta-Guzman and Wagner, Frontiers in Computer Science 2026 |
| Video memory needed by NVIDIA’s smallest ACE language model (the vendor’s own claim) | As little as 1.5 GB | NVIDIA, ACE Autonomous Game Characters 2025 |
Context: NVIDIA’s ACE announcement names PUBG, inZOI, NARAKA: BLADEPOINT, MIR5, AI People, and Dead Meat, but a named partner is not a usage figure. Valve adds friction too: live-generated content requires developers to explain the “guardrails” on their AI. The hobbyist route is visible in Mantella, a Skyrim and Fallout 4 mod linking speech-to-text, a language model, and text-to-speech, with 420 GitHub stars and 109 forks on October 31, 2026.
6. Player Acceptance: A Loud Rejection and a Neutral Majority
83% of surveyed core players were negative on AI-generated dialogue, up from 46% for the combined quests and dialogue option a year earlier (Quantic Foundry, Gen AI survey 2025). Circana’s broader US panel reads very differently: more than half neutral, about a quarter less likely to buy (Circana PlayerPulse, as reported by 3DJuegos). Both can be true, because they sample different people. Quantic Foundry’s 1,799 respondents were 74% core and 21% hardcore gamers, and players motivated by story were more negative (r=-.12).
Detection is a weak shield for either side. In Lavan et al. in PLOS One, listeners separated generic synthetic voices from people easily, but in the dedicated clone experiment could not separate clones of real speakers from those speakers (t(49) = 1.68, p = .100). Each clone used about 4 minutes of recordings, and reference audio benchmarks track how much shorter that input has become. The same class of voice cloning software now sits within reach of small teams, so consent and disclosure will carry more weight than listeners’ ears.
| Metric | Value | Source |
|---|---|---|
| Core players negative on AI-generated dialogue vs AI-generated quests, 2025 | 83% vs 77% | Quantic Foundry, Gen AI Survey 2025 |
| Core players negative on the combined ‘AI quests and dialogue’ option, 2024 | 46% | Quantic Foundry, Gen AI Survey 2025 |
| US players less likely to buy a game using generative AI for art, dialogue, text, music, or voices: December 2025 vs March 2024 | Just over 25% vs 22% | Circana PlayerPulse, via 3DJuegos 2026 |
| US players neutral on that same question | More than 50% | Circana PlayerPulse, via 3DJuegos 2026 |
| US gamers expecting AI performances to be worse than / as good as / better than human ones | 40% / 19% / 18% | YouGov, August 2024 survey |
| Listeners labelling voices as human: real recordings vs voice clones (Experiment 1b, n=50) | 62% vs 58% | Lavan et al., PLOS One 2025 |
| Listeners labelling voices as human: real recordings vs generic synthetic voices (Experiment 1a, n=50) | 81% vs 41% | Lavan et al., PLOS One 2025 |
| US gamers open to a price rise for advanced AI NPCs (vendor-commissioned survey) | 81% | Inworld AI, The Future of NPCs 2023 |
Context: Most recent available data on NPC-specific demand: Inworld AI, The Future of NPCs, 2023 (1,002 US gamers aged 16 to 50, surveyed in late 2022). Inworld sells NPC technology, so read it as an upper bound on enthusiasm. The same survey found 52% dislike repetitive NPC dialogue, the complaint AI voices promise to fix.
7. Performers: The Contract Already Prices a Market That Barely Exists
The 2025 SAG-AFTRA video game contract sets a minimum of 7.5x scale, the union’s base rate, for “Real Time Generation”, defined as a digital-replica-voiced chatbot inside a game (union board summary, via Graphic Policy, June 12, 2025). Set that against section 5, where live generation is a tiny share of disclosures: the union priced the use case before the market formed.
The people closest to game writing are also among the most opposed. Game design and narrative staff report some of the highest negativity in the GDC survey. Enforcement has already been tested: the union filed an unfair labor practice charge on May 19, 2025 over an AI Darth Vader voice in Fortnite (Associated Press, via ClickOnDetroit). For contract detail beyond games, see our voice actor AI cloning and union contract statistics.
| Metric | Value | Source |
|---|---|---|
| Contract minimum for a digital-replica-voiced chatbot in a game, as a multiple of union base rate | 7.5x scale | SAG-AFTRA, Interactive Media Agreement summary 2025 |
| Member vote to ratify the 2025 Interactive Media Agreement, closed July 9, 2025 | 95.04% yes, 4.96% no | SAG-AFTRA, ratification results 2025 |
| Length of the video game strike, July 26, 2024 to suspension on June 11, 2025 | 320 days (our count) | SAG-AFTRA dates, via Graphic Policy 2025 |
| Games signed to union interim and tiered-budget agreements by March 2025 | Over 180 | SAG-AFTRA, via Graphic Policy 2025 |
| Game design and narrative professionals vs all respondents saying generative AI is harming the industry | 63% vs 52% | GDC, State of the Game Industry 2026 |
| US gamers disagreeing vs agreeing that AI should reproduce performers’ voices and action without compensation | 65% (52% strongly, 13% somewhat) vs 26% | YouGov, August 2024 survey |
| US gamers opposing vs supporting AI replacing human roles for faster development | 56% vs 23% | YouGov, August 2024 survey |
Context: the ratification notice also lists consent and disclosure requirements for digital replicas, and lets performers suspend consent for new material during a strike. YouGov’s survey covered 397 US gamers.
Summary: AI Voice in Game NPC Dialogue by the Numbers
| Metric | Value | Source |
|---|---|---|
| Next Fest demos with an AI disclosure, October 2026 | 2,502 of 7,240 (34.6%) | How To Market A Game, 2026 |
| Disclosing demos citing audio or voice vs runtime generation | 941 vs 76 | How To Market A Game, 2026 |
| New Steam games carrying an AI disclosure, 2026 to date | 30.8% | Sulka Haro census, via Games.gg 2026 |
| Developers reporting generative AI use, early 2026 | 29% | Game Developer Collective / Omdia, via Outlook Respawn 2026 |
| US gamers expecting AI performances to be worse than human ones | 40% | YouGov, August 2024 survey |
| AI-disclosing titles above $1M citing voiceover or dialogue | 27% | Totally Human Media, 2025 |
| AI voice in disclosures: successes vs flops | 24% vs 8% | Sulka Haro, 2026 |
| Game developers using AI for dialogue writing support | 36% | Google Cloud / The Harris Poll, 2025 |
| Unity developers using AI for writing and narrative design | 44% | Unity, 2026 Game Development Report |
| Game audio professionals using generative AI at work | 8% | GameSoundCon, 2025 |
| Generative AI users citing placeholder dialogue | 43% | GameSoundCon, 2025 |
| Average NPC speech cycle in a VR study | 6,909 ms | Korkiakoski et al., arXiv 2025 |
| Core players negative on AI-generated dialogue | 83% | Quantic Foundry, 2025 |
| US players neutral on generative AI in a game they might buy | More than 50% | Circana PlayerPulse, via 3DJuegos 2026 |
| Listeners labelling voice clones vs real recordings as human | 58% vs 62% | Lavan et al., PLOS One 2025 |
| Contract minimum for a digital-replica-voiced game chatbot | 7.5x scale | SAG-AFTRA, 2025 |
Methodology and Sources
- How To Market A Game, How Sloppy Is the October 2026 Steam Next Fest, October 21, 2026
- Sulka Haro, Three Years of AI on Steam, July 13, 2026 and Games.gg coverage of the census
- Totally Human Media, Games With AI Disclosures Have Grossed an Estimated $660M on Steam, November 2025, and the July 2025 count as reported by 80.lv
- Valve, Steamworks Content Survey documentation
- Dataconomy, citing Eurogamer and SteamDB, June 17, 2026
- Naavik, Steam Next Fest’s AI Problems, 2026
- GDC, State of the Game Industry 2026 summary
- Google Cloud and The Harris Poll, AI Meets the Games Industry, 2025
- Unity 2026 Game Development Report, as reported by Digital Today and Wccftech
- Game Developer Collective and Omdia, as reported by Outlook Respawn, 2026
- GameSoundCon Game Audio Industry Survey 2025
- PCGamesN on Arc Raiders voice lines, March 14, 2026
- Korkiakoski et al., An Empirical Evaluation of AI-Powered Non-Player Characters’ Perceived Realism and Performance in Virtual Reality Environments, arXiv 2025
- Berrezueta-Guzman and Wagner, Next-Gen Orientation: Supporting International Students with Generative AI NPCs in VR, Frontiers in Computer Science 2026
- NVIDIA, ACE Autonomous Game Characters, January 6, 2025 (vendor source)
- Mantella repository on GitHub
- Quantic Foundry, Gen AI survey, December 18, 2025, with the quests and dialogue split as reported by GamesMarket
- Circana PlayerPulse, as reported by 3DJuegos, 2026, with the December 2025 date from Outlook Respawn
- YouGov, Gamers See a Role for AI in Gaming, September 11, 2024
- Lavan et al., Voice Clones Sound Realistic but Not (Yet) Hyperrealistic, PLOS One 2025
- Inworld AI survey, as reported by PocketGamer.biz, February 15, 2023 (vendor source)
- SAG-AFTRA board and strike notices, as republished by Graphic Policy, 2025, ratification notice, as republished by WNY Labor Today, and Associated Press report on the Fortnite charge, via ClickOnDetroit
- Data watch: No public dataset counts NPC lines voiced by AI, so every voice figure here is a proxy built on self-reported Steam disclosures or surveys. How To Market A Game’s ‘Audio / voice’ category is not limited to speech. Its rank buckets sum to 7,042 games and 2,445 disclosures rather than 7,240 and 2,502, and a separate SteamDB-based tally reported by PC Guide and quoted by Capital and Compute gave 2,593 of 7,145 (just over 36%). June 2026 Next Fest denominators also differ: 8,700 listed titles in SteamDB data against 4,382 featured demos in Naavik. One secondary article, Capital and Compute (October 2026), reverses Haro’s AI voice comparison, while Haro’s original text says successes were ‘3x more likely’ to disclose it (24% vs 8%). Generative AI use varies with the sample: 29% in Omdia’s early 2026 wave, 36% in GDC 2026, and 95% of Unity’s 300 surveyed developers adopting AI in some form. Quantic Foundry’s page lists quests and dialogue as ‘77% and 83% negative respectively’, and GamesMarket reads that order the same way, so the 83% dialogue figure depends on that reading. Circana’s exact question wording and sample size were not available from a primary Circana document. The Google Cloud 34% and 36% figures were cross-checked against Dataconomy, and the report’s phrasing leaves unclear whether 36% is shared with level design. Mantella’s Nexus Mods download count could not be read, so GitHub activity stands in for it. Newer data is expected from GDC’s 2027 survey, Unity’s 2027 report, and the next Next Fest counts.
Last updated: October 31, 2026. We update this roundup quarterly, and the next refresh is expected when GDC publishes its 2027 State of the Game Industry report.