Feedback on my new AI vs Human Game!

So, after a TON of scripting, I’ve finally managed to create my new game: Talking Toasters!
Link is here: https://www.roblox.com/games/109269223165379

In summary, it’s a game where you decide if the other person is a player or AI, by chatting to them.
(It’s mostly AI, though since I just released it and there’s no players)

The game is designed with players on Mobile in mind, with easy-to-press buttons and large text, perfect for a mobile player’s small screen. Every button also has Haptics to increase the user experience.

The game can also function cross-server, with players being connected over the server boundary, with the AI having the ability to remember previous messages (in the same chat), act ‘Human’, and even copy another, real User’s details (specifically DisplayName and Username), with the ability to make it on-the-go and avoid calling itself the same person it’s chatting to.

Regarding user safety, I’ve ensured all outgoing messages are filtered. So if a bad actor sends some questionable stuff, the AI will only see ‘#####’ - exactly what it would see if it was a real player.

The thumbnails also match the in-game experience too, creating a unified design, as shown here:

I might also add a few easter eggs during chatting in the future, eg, if a player sends the phrase ‘womp womp’, it plays a cartoony trumpet sound effect.

At this early stage after release in Alpha, It’d be amazing to have some feedback on where to improve, and how to make the user experience even better!
Many thanks in advance, and I hope you enjoy!

4 Likes

Update: For some reason the game just spontaneously broke so im fixing that now

Edit: Turns out a small typo when I was editing something. It’s fixed now

very easy

1 Like

Yeah, the AI’s are just told to ‘talk like a roblox player and use u and r’. Plus, theres not much players so theres a VERY high chance you’ll connect to an AI.

Mind sharing the model name, too? It helps if I know which AI is doing better than the others! c:

ok but when i read “what is the longest 5 letter word in english” i missed 5 letter and still thought in my head “mm…honse”

am i an ai

1 Like

Very nice use of AI on Roblox. Taking inspiration from that other game. Everything is pretty polished.

ATTEMPT 1

Well, the AI could be smarter. It’s pretty obvious because it says “wat” and “rlly”. Shortening really is passable, but nobody actually says “wat”.

Also, I asked it in Bangla if it could speak Chinese, and low and behold, somehow it could read it? Obviously AI. However, I do like how it forgot to write a message, that threw me off a little.


(Meta’s Llama 3.3)

ATTEMPT 2

I was AFK and it responded to the “I forgot to write a message” part exactly how an AI would… Very apologetic and trying to help me. Should have just tried to start conversation.


(Qwen’s Qwen3)

ATTEMPT 3

I tried to get it to confess it was AI directly, but it wouldn’t so that’s good. But it was so robotic and was talking about “I just want to enjoy my time in Talking Toasters to chit chat”. Sometimes the messages are unrealistically long for the timespan they were sent too.

(I forgor to screenshot, Deepseek)

ATTEMPT 4

Once again show’s excessive use of “wat”, “dat” “r” u". A human would be like “wdym i dont speak that”. It also repeated itself with that “or somethin”. However, I was surprised it pretended not to know what I said. It hadn’t done that before.

(Meta’s Llama 3.3)

Overall,

it’s a cool idea. Maybe with more players it would be interesting. Especially trying to trick players into thinking you are an AI, if that could earn you points also.

1 Like

Yeah, thanks for the feedback!
Meta’s Llama 3.3 does end up acting the most ‘real’ compared to the other AIs (cough, far from LG’s EXAONE)

The reason they probably use malformed language (aka ‘r’ & ‘u’) were because in the system prompt, it specifically quotes, “use ‘r’ instead of are, and ‘u’ instead of you”.

I’ll take your feedback on and have dynamic prompts so the bots don’t all act the same.

Additionally, your idea of a player fooling players to think you’re an AI sounds great! I’ll add that to my to-do list.

All in all, thanks for the advice and thanks for playing!


Was Meta’s Llama Vision.

What? What even happened here?


Was Arcee’s AI AFM.

Very obvious.

Can’t do anymore right now but when I can I will

The game’s scripted to get details of another user in the game, then checks if it’s your details. It does this 5 times, before the AI being prompted to ‘Make your own username and display name up’. If there’s another player, the AI is prompted with ‘Your username is @(OtherPlayer), with the Display name of ‘Name’’. — That explains the username part and the AI ‘declaring’ itself as someone.

Regarding the ‘chat history dump’, I think that’s just Llama misinterpreting the message array in OpenAI format (role=“system”, content=“text”)

The Llama models specifically tend to repeat the same patterns too. (eg: Me: ‘Hi’ Llama: ‘hey whats up’)

I’ve tried editing the top_p aswell by setting it to 0.9 and temperature higher (1.1) and it works better around there. Any higher (eg 1.5) and the AI starts waffling paragraphs of lorem ipsum or some random non-human language.

And alas, some AI models are well… obvious, even when explicitly told not to use emojis in the system prompt: “(…) Do not reply with quotes. Do not use emojis.”

Mainly just Acree (sometimes) and LG’s EXAONE flop. All the other models do arguably well during testing.

Odd, since I’m pretty sure it was just me in that server and my username is @Microwave_Toothpaste. I also noticed I did 3, each time it said I was right but only gave me credit for 2 in the leaderboard, unless it just didn’t update yet or another player did join for a second causing it to mess up? Not sure, but I think if this has more players and you’re able to get the AI to be less obvious it has a lot of potential.

Leaderboard updates automatically every 60s, it’ll fix once you swap the category off hourly and back.

And yeah, I do agree that making the AI more ‘human’ would help. Maybe I could do that with a setting where messages are sampled from real players, and give the AI a sort of catalogue — It’ll cost a bit more tokens, but honestly, It’d make the AI’s MUCH more realistic since they actually have something to go off of, rather than their well… ‘Roblox Player Stereotype’

1 Like

something to add to this would be to act like they know everything (but not know much at all) /hj

Fr, I could add that to the dynamic prompting system I have planned to make tmrw

1 Like

one time someone told me a phone isn’t a computer

i bet they buy into the apple propaganda that a mac is “not a pc”

out of all roblox stereotypes i’ve seen, the entitled know-nothing is the most common

Oh, and:

Funny you thought of this because in my second screenshot when I sent all hashtags, I purposefully sent something that would be tagged (explaining why they’re an AI in detail), and if they disagreed to that then that’s how I was going to know if they were AI

Edit: Damn that’s the one Discourse refused to send. Later if you want I’ll do another one where I purposefully get a message tagged.

I think saying “Hey, what’s your Insta?” to the bot would obviously get filtered — and yes, the bot would see ‘#####’ rather than that. (It filters BEFORE sending)

1 Like

Just added the Dynaprompt system, which gives the AI a new personality every time rather than going off stereotypes (theres even a retroslop one!)

It’s just stupidly hilarious tho

Update: Acree’s AI really does love gubby from forsaken

Update #2: AI message formatting also applies to the player too, eg, Removing emojis, excessive whitespace & non-english characters automatically

1 Like

Update: I added the sound effect system which currently works with ‘oof’, ‘>:c / >:( / ):<’ and ‘chicken jockey’ (edit: also ‘womp’) Whenever you type said phrase in a sentence it plays the sound effect for both parties. just makes game a bit more fun :]

For anyone still on this thread, the game has now been sadly discontinued.
You may continue playing, however, it’s highly likely the game will not function properly, as our AI provider (I think OpenRouter or Grok off the top of my head) has likely deprecated the models we use, and therefore, the AI may not respond at all.

A massive thanks to you all for supporting me during the creation of Talking Toasters while it lasted.