AI callers in Scam With Your Friends
| How callers talk | Real-time AI dialogue |
|---|---|
| Internet | Required, including solo play |
| AI service | OpenRouter with Gemini 2.5 Flash Lite (developer, Sept 11); may change |
| Languages | 23 on Steam, all with full audio; 21 call languages in the Sept 15 beta |
| AI cost | Unlimited, included in the price; own AI keys removed in 0.1.95 |
| Playtest AI credits | Over; the $8 Unlimited AI page is removed |
| Speech recognition | Local: Whisper Small (default), Whisper Base or Moonshine (0.1.95) |
| Caller voices | Pocket TTS, MOSS TTS Nano, Qwen3 TTS, set per language (0.1.95) |
| Typing | Always possible instead of talking |
| Call saving | Newest 10 kept by default, or None (Stored Media History, 0.1.90) |
| Custom callers | Steam Workshop and Character Creator, on the main menu since 0.1.84 |
How conversations work
There are no dialogue trees. The Steam page describes real-time AI conversations, and its AI disclosure says the AI lets characters react dynamically to what players say and do. You talk, the caller answers in character, and both sides of the exchange appear as chat bubbles in the Phone window. While the caller is thinking, the window shows a "message sent, waiting" status.
The Steam page also lists the outside service behind this: the game "connects to 3rd-party service for AI content generation: OpenRouter". OpenRouter is a gateway to many AI models, so Steam alone does not tell you which one plays the callers. Developer Metater named it in the official Discord on September 11, 2026: Gemini 2.5 Flash Lite.
What runs on your PC and what runs in the cloud
Developer JakeHub explained the split in the Steam thread "The AI":
- Local: the speech-to-text that turns your voice into text, and the voice synthesis the callers speak with.
- Cloud: the caller's "brain", a language model on OpenRouter.
He added that the local language models they tried were huge and bad at conversation, so a fully offline mode is not planned for now, "maybe in a future update". That is also why the game needs an internet connection even in single-player.
Developer Metater gave the same split in the official Discord on September 11, 2026, with two extra details: the local voice synthesis and transcription prefer your GPU and fall back to the CPU, and the local language models they tried were "far too dumb" and would need something like a modern NVIDIA GPU. More on the hardware side is on the system requirements page.
The model may change
Gemini 2.5 Flash Lite is what the callers ran on in September, not a promise for launch. On October 1, 2026 Metater wrote in the official Discord that the model "just went down partially" and that "this model is going away anyway in 20 days, so we have already been looking into replacements." On October 5 he said they are working on "a deal with an AI provider", and on October 6 they put "more improvements to the AI" on the playtest's development branch. Which model the callers use since the October 15 launch has not been announced. On launch day Metater added that there is "a bring your own AI key option so you can try any models you want": hosts could plug in their own OpenRouter key. Update 0.1.95 (October 28) removed that option from standard builds, so every host now uses the Official Backend. The Steam page still names OpenRouter as the AI service.
How long the AI will be supported
No end date has been announced. In the same September 11 post Metater said they will "try our best to keep it up for as long as possible" and estimated that the full game will cost them "a few cents per player". On launch day he wrote that they are "committing to keep the game up as long as possible (years, not months)" and that "the game is not at risk of going down".
The phone has microphone and speaker buttons and Steam tags the game with Voice Control, so the intended way to play is out loud. You can always type instead.
Speech, voice and privacy settings
The launch patches, the October 18, 2026 Speech and Privacy Update and Update 0.1.95 (October 28) reworked how your voice and the callers' voices are handled:
- Speech settings tab for language, recognition, voice models, microphone leveling and caller filler words.
- Local recognition: since 0.1.95 a choice of Whisper Small (456 MB, the default), Whisper Base (141 MB) or Moonshine, which covers English, Arabic, German, Japanese, Mandarin and Spanish and also runs on older processors without AVX.
- No more online speech services: Deepgram (October 18) and Gladia (0.1.92) needed your own paid account; 0.1.84 and 0.1.95 removed them again.
- Caller voices: since 0.1.95 you turn the voice models Pocket TTS, MOSS TTS Nano and Qwen3 TTS on or off and order them per language. Qwen3 needs a compatible GPU (6 GB VRAM since 0.1.83), switches off if speech is too slow to start, and is not available on Windows ARM or Snapdragon PCs. Russian has a bundled Pocket TTS voice.
- Microphone: automatic mic leveling is on by default; turn it off to set the level manually from 50% to 200%.
- Typing: typed dialogue always works, and the main menu warns you if your PC cannot run speech recognition.
- Privacy: a first-launch EULA and Privacy Notice; Stored Media History (0.1.90), which keeps the newest 10 videos and pictures by default or none at all (temporary boss-review recordings remain); and Streamer Mode (0.1.90), which hides other players' screens. Update 0.1.92 removed the in-game web browser.
Metater wrote that players "were concerned about privacy so we hope this clarifies some of that." The patch notes also say Mac, Linux, CrossOver, Wine and Proton are unsupported.

Caller Trust
Above each caller's portrait is a Caller Trust bar with a percentage and a mood label. Official screenshots show the labels Trusting and Angry. The scam only works while the caller is still willing to cooperate, so how you speak is the main skill in the game.
The Steam description also warns about scambaiters: callers who are trying to waste your time or hack you back. Falling for one can give your in-game computer a virus.
Callers seen so far
Seven callers appear by name in official screenshots. The cards below show each one with the trust level and mood from the screenshot it was taken from. The full roster is not public.

Elderly and cooperative. Shown confirming a number back to the player just before a $250 Social Security scam completes.

Wears a witch's hat and reads out her Social Security number while fretting that it is a bad omen.

An astronaut who wants everything done by the book before he gives up his AnyViewer code.

Furious and nearly gone, while the player is left apologizing.

Sunglasses, insults and a pitch for the new token he is launching. Trust at zero.

Unimpressed by a player's very strange request. Trust at zero.
The seventh, Shanice "Dee" Bishop, appears only on the incoming call popup above.
AI credits
There are no AI credits in the released game. Metater confirmed in the official Discord on September 29, 2026 that Early Access and 1.0 "will both have unlimited AI credits inside the base price of the game. There will be no microtransactions." On launch day, October 15, 2026, he repeated that "Unlimited AI usage is priced into the game for everyone". Developer JakeHub asked for optional Ko-fi donations to help with AI server costs; Metater added there is "no obligation to donate anything".
Since Hotfix 2, AI failures and temporary usage restrictions show up as clear status messages instead of fake caller dialogue.
The credits were a playtest thing: free credits, shared round robin across a lobby, plus an optional $8 Unlimited AI purchase. The official Unlimited AI page now says it has been removed, and Update 0.1.73 restored VIP cosmetic benefits for players with legacy unlimited AI access. See the playtest page.
Languages
The developers' closed-beta changelog of September 15, 2026 lists AI calls in 21 languages: English, Spanish, French, German, Portuguese, Italian, Russian, Chinese, Japanese, Korean, Arabic, Hindi, Indonesian, Polish, Turkish, Czech, Hungarian, Swedish, Danish, Greek and Persian. You pick the call language from the phone or the pause menu. The game's interface was not localized yet at that point.
Steam lists 23 supported languages, all with full audio. Compared with the changelog, Steam's list also has Malay and Estonian; the other 21 match. The full Steam list is on the system requirements page.
On October 14, 2026 developer Metater wrote in the official Discord that they "switched to Qwen3 TTS for non-English TTS in 10 major languages", with a "big fix" for non-English voices coming soon. Qwen3 TTS is a text-to-speech model, so this changes how callers sound in those languages. Which ten languages was not said.
At launch the developers asked non-English players to let the Qwen3 TTS download finish on the main menu for higher-quality voices, and the first hotfix moved those downloads to a dedicated server. Qwen3 TTS now requires a compatible GPU. For your own voice, Moonshine recognition covers English, Arabic, German, Japanese, Mandarin and Spanish (October 18 update), and Whisper the rest. Update 0.1.92 (October 26) made the boss review speak and caption in the host's language, and Update 0.1.95 (October 28) added per-language voice models plus a Pocket TTS voice and better word stress for Russian.
PG mode and censored swearing
Developer JakeHub said in the official Discord on October 1, 2026 that they are "adding guards against this aswell as having a PG mode and a censored swear words mode". He was replying in the #bugs forum to a report about offensive language from the boss character. It shipped as Family Friendly mode: by Update 0.1.82 (October 19) it applied to caller openings, conversations and the boss review, and Update 0.1.92 (October 26) added cleaner review captions and filtering of sexual details; its notes also say the mode hides other players' screens. See settings.
Making your own callers
The game supports custom personalities built from JSON files and voice cloning WAV files. The closed-beta changelogs add two in-game tools: a Character Creator (draw a portrait, set the personality, record a voice; September 15) and community callers that players can upload, browse by language, vote on and add to their caller pool (September 16). Since launch, Steam lists Steam Workshop support and the patches fix Workshop caller uploads and WAV or MP3 voice imports in the Character Creator. Update 0.1.84 (October 24) put a Caller Creator shortcut on the main menu, and the developers are running a caller contest until November 1. See mods and custom personalities.
FAQ
- Does Scam With Your Friends use AI?
- Yes. The Steam AI disclosure states that AI powers real-time dialogue and responses from fictional in-game characters, allowing them to react dynamically to player conversations and actions. Since the October 15, 2026 launch, unlimited AI is included in the price. Speech recognition runs locally on your PC (Whisper Small, Whisper Base or Moonshine since Update 0.1.95), and you can always type to callers instead of talking.
- Are the callers real people?
- No. Every caller is a fictional character voiced and driven by AI. No real people are called.
- Do I have to pay for AI in the full game?
- No. Developer Metater wrote on launch day (October 15, 2026) that "Unlimited AI usage is priced into the game for everyone" and that donations are optional. The playtest's AI credits and the $8 Unlimited AI pass are over: the official Unlimited AI page now says it has been removed, and Update 0.1.73 restored VIP cosmetics for players with legacy unlimited AI access.
- Does the game work offline?
- No. A broadband internet connection is listed as a requirement, including for single-player, because the caller conversations run live.
- Which AI model does the game use?
- Developer Metater said in the official Discord on September 11, 2026 that the callers' dialogue runs on Gemini 2.5 Flash Lite, reached through OpenRouter (the service the Steam page lists). That may change: on October 1 he wrote that the model "is going away anyway in 20 days" and that they were already looking into replacements, and on October 5 he said they are working on a deal with an AI provider. Which model runs since launch has not been announced. On launch day Metater also mentioned a bring-your-own AI key option; Update 0.1.95 (October 28) removed that OpenRouter option from standard builds, so every host now uses the Official Backend.
- Which languages can the callers speak?
- Steam lists 23 languages, all with full audio. The developers' September 15, 2026 closed-beta changelog listed AI calls in 21 of them. Since Update 0.1.92 the boss review also speaks and captions in the host's language, and Update 0.1.95 added voice-model choices for each language.
- How long will the AI callers keep working?
- No end date has been given. Developer Metater said in the official Discord on September 11, 2026 that they will "try our best to keep it up for as long as possible", and on launch day, October 15, 2026, that they are committing to keep the game up "years, not months".
- Is there a family-friendly mode?
- Yes. Family Friendly mode was in the game by Update 0.1.82 (October 19, 2026), which applied it to caller openings, conversations and the boss review. Update 0.1.92 (October 26) added cleaner review captions and commentary; its notes also say the mode hides other players' screens. Developer JakeHub had promised "a PG mode and a censored swear words mode" on October 1.
- Can I type instead of talking?
- Yes. The launch patches say typed dialogue remains available, and if your PC cannot run speech recognition the main menu now warns you and reminds you that you can still type to callers.
- Can I stop the game from saving my calls?
- Yes. Set Stored Media History to None (Update 0.1.90, October 25, 2026) and the game saves no new videos, photos or artwork; by default it keeps the newest 10. The October 18 Speech and Privacy Update had added a "Disable Call Saving" option. Temporary recordings for the boss review remain. The game also shows an EULA and a Privacy Notice on first launch, with links in Settings.
- Does the game send my voice to the cloud?
- Not by default. According to developer JakeHub on Steam, speech-to-text (turning your voice into text) and the callers' synthesized voices both run locally on your PC; only the caller's 'brain', a language model reached through OpenRouter, runs in the cloud. The optional online speech services that sent your audio elsewhere are gone again: Deepgram (added October 18, removed in 0.1.84) and Gladia (added in 0.1.92, removed in 0.1.95).