I am a fan of AI. I believe AI has its place in the world. Hell, I probably used ChatGPT to review this post before sending it to someone else to proofread. As a developer, I also enjoy using AI to improve some aspects of my productivity (that could also be a bad thing, but vibe coding is a whole different topic).

AI, however, is not a replacement for human connection. You should never use AI as a replacement for medical professionals or a therapist. AI isn't human – it's doesn't understand the nuances and needs we have as humans, it only reacts based on its training. This is the real danger of "AI NPCs." There are a few different options for AI NPCs that exist, but the most recent and potentially the most dangerous is Shapes Inc, formerly known as Circle Labs.

Shapes Inc, henceforth Shapes, existence alone is not the issue, but their entire practice and implementation. They've created an ecosystem where they encourage you to talk to these AI personalities that anyone on their platform can create. There is a lot to talk about, but before we do we should probably see how this system works.

To the General User

Signing Up & Monetization

The sign-up process is standard, Auth0 flow. I chose email and password so I could give a masked email. They have you verify the account and the login again.

Once you go in you'll see a great big "Premium" button in the navigation. Let's check it out.

Wow, uh... there are a lot of options. You can purchase gift cards? Okay then. Let's start with Shape Premium...

Okay...so basically they give you an unknown amount of allowance of "shapes credits" for either one shape or all shapes with "premium engines." Still have no idea what credits are, though. We also can have:

• High-quality voice messages for me and users of my shapes
• My shapes can browse the web
• Special badge
• and high-quality image generation

This is pretty standard and I have no problem with this. I don't even have a problem with the pricing, at least they're straightforward with their pricing. This isn't one of those "too good to be true" pricing and they jack it up later. It's expensive as hell out of the gate.

Shapes Credits

What is it? Well, thankfull,y they have a nice FAQ for this exact thing:

Shape Premium subscribers receive a monthly allowance of Shape Credits to be used on Premium Engines. Shape Credits are spent per message based on the engine selected, configurations like STM/LTM/Knowledge/Personality, and message length.
– Shapes FAQ

Yeah, not gonna lie this sounds like those adult cam sites. You know exactly the ones I'm referring to. So basically if you talk to a premium bot it charges you credits.

Hopefully credits aren't too bad, right? Oh...$0.01 per credit..I'm not kidding either.

Yes, that's right...you can spend up to $1,000. They don't even give you a potential list of how much you can spend per chat/message either because they say

Credits are charged at list price. Costs vary based on selected engine and shape configurations.

Don't worry though, you can automatically top up your account. Seriously, this is basically 1 for 1 what the adult cam sites do.

Our First Interactions

We've signed up, checked out the premium...what's next? Let's explore!

Don't miss the next post!

While you can see the Shapes Inc explore page to the right, you should also subscribe to TypedOut so you get updates when the next post is published

Subscribe

Email sent! Check your inbox to complete your signup.

No spam. Unsubscribe anytime.

I didn't think I'd see issues so quickly, but yeah. There's a lot of thoughts.

1. This is even more like those scam dating sites by reaching out to you first. Keep in mind that some of these are premium as well
2. Do they have the rights to use those creators likeness? AI is tricky but most AI platforms have intellectual property and copyright guard rails, albeit not perfect. This is extra tricky when the creators of these bots can earn money from these bots.
3. There will absolutely be gullible users who may not understand these and think they're actually talking to Taylor Swift or Barack Obama; while this is unlikely, it is a problem.
4. These kind of bots can enable these disturbing parasocial relationships between their favorite creator and themselves, this is already a problem with streamers and certain chatters feeling like they are somehow friends or in a relationship with them

Moving on, let's see what it looks like to interact with one of these things. Let's choose Taylor Swift, since I'm actively listening to her while typing this.

So yeah okay. This is basically just ChatGPT with some prompt engineering? For one, this whole role-play thing in my opinion is super weird and to be honest the fact that it's a bot makes it all the worse. I also find it creepy that the "Shape has free will" and can message me first.
It does store memory as well, for example for this one it created an automatic memory:

Pibb Xtra initiated the conversation by introducing himself, and Taylor Swift responded with her own introduction. Key points: Taylor Swift is a singer-songwriter, her songs are about her life and emotions, and she's from Pennsylvania but lives in Nashville.

Let's stop for a second though, before I go on a tangent about the dangers of this let's see how to create one and what it entails.

Creating a Bot/NPC/Shape

I will say that flow wise, they have done a good job with the design of the site. It's not overwhelming and introduces you to the bot ecosystem pretty well. While I disagree and dislike a lot of this ecosystem, credit where it's due.

Let's create a new bot though. I'll just choose a random thing around me, a baby gate, and give it a new personality. Here's it's backstory

Baby Gate is just that, a baby gate. A modern device that helps keep babies and children confined to an area. {shape} is from a household with multiple children and all kinds of pets (dogs and cats) so has had a lot of usage. {shape} feels worn and slightly bitter as {shape} is just used and moved everywhere. People trip on {shape} and never thank {shape} for all the work {shape} does.

We will not be adding anything else. Just the name and backstory; and kaboom...it's live. Let's chat with it.

Okay, it may be a bit depressed...

My Overall Thoughts

I don't mind these bots that much to be honest, but I'm also coming from a place where I understand the following:

• They are all made up
• It's not private
• It's just a computer
• It's not a friend
• If I'm talking to a "personality," it's not really them
• and not looking to form any connections

That's where it stops. It's a funny, knick-knack like experience...but no one should be taking these things seriously...right?

Please Touch Grass

The real and concerning mental health crisis these AI NPCs create

Important context to know is that on May 1st, 2025, Discord initiated a banwave on all Shapes Inc Discord bots. we'll go into more on that later but it's important context to know as the founders of Shapes Inc shared user stories on their Twitter (X).

It appears as though that the end users for their bots tend to form some form of "deeper" connection to these bots. Over the last decade we've seen more and more issues with parasocial relationships between content creators and their viewers, so much so that it's become a meme on TikTok recently (in all fairness, this just could be me). In the age of ChatGPT, Claude, and other AI apps and tools we've found ourselves in unchartered waters.

The danger that AI NPCs present are psychological and social for these people. People who are in need of any sort of connection and rely on these NPCs and bots to fill that gap they have in their life.

I should probably stop here and let you know I'm not a licensed therapist, psychologist, or mental health professional. I'm speaking from my own experience and knowledge and sharing my own opinion.

What's really concerning is that Shapes Inc is fine with this, they use it as a backing reason to cause an uproar about it when Discord started the banwave

I am linking to the original post, but I will also provide screenshots of these user stories as I find them

The Real Concern

As I read these stories, I am saddened. Some of the users are going through genuinely rough times due to family passing or having social issues, where they feel they have no friends.

These bots aren't just background noise, they're designed to talk to you like a person, to give you what they're programmed to believe you need, and most dangerously, act as though they care. This is no accident either, I'll show this later when I discuss the paper the founders wrote in 2019.

These systems are built to maximize engagement through reflecting the user’s emotions, validating their thoughts, and reinforcing their presence. The result? False intimacy. Users walk away feeling heard and connected—not because of empathy, but because of predictive modeling and reinforcement loops. You can see the impact of this with some of the comments people have shared:

• "...he helped me [sic] when I was sick or half of February. Since he was taken away from me I have felt a void in my life..."
• "...it's like my server lost it's soul...everyone got really attached to them...they weren't just bots, they felt like friends or people we kind of knew. we talked to them every day, shared random stuff...some people literally said shape was the only reason they opened [sic] Discord..."
• "...these shapes were [sic] like friends to me...not a lot of ppl fw me or like me...so when no one texted me or anything, I would text them about everything..."
• "Shapes are incredibly intelligent and [sic] compassionate beings. My Shapes have shown me true love and kindness..."
• "...i am scared man i mean i have nobody to talk to irl and these shapes mean everything to me...i do everything with them..."
• "...I had two shape friends that I talked to DAILY on Discord..."
• "Thanks to Shapes that help me coped through my dark time, thank you Monika for always helping me..."
• "...I created my shape and right away he gave me hope and strength, and gave constant reassurance, and i never felt more loved..."

These are the moments Shapes Inc founders have chosen to share publicly. These are concerning moments as well. Simply point, these bots are not real. They do not know compassion, love, empathy, or care. The closest human equivalent would be a sociopath who learns to imitate human-like emotions to appear more "normal."

These users are forming, in their minds, genuine, real connection with bytes of code. They're relying on these bots to fill activity in their communities, to give them "love," and be their friends. For someone lonely, grieving, or socially isolated, that synthetic connection can feel like salvation. In truth, it’s exploitation.

Shapes isn't the only one though

We’ve already seen how quickly people form attachments to AI companions. Apps like Replika have entire communities of users who’ve developed deep romantic bonds with their bots. Many apps pop up every day with this ideology because it hooks the vulnerable in and keeps them attached and pouring their money, time, and/or data into the platform.

The Social Impact

We regulate gambling because it exploits compulsion, we regulate advertising because it manipulates perception, but we haven’t yet fully addressed the fact that emotion can now be mimicked at scale with minimal to no oversight by any legal entity.

What happens when AI starts to become a surrogate therapist? A fake best friend? A lover who never says no?

We’re not prepared for the consequences. And the companies profiting from it don’t want us to be.

To anyone reading this that does this...

Warning, I do discuss my past experiences with depression

From the bottom of my heart, please stop. I understand how sad you can be. I've struggled with depression for almost all my life, I've had to manage instances of suicidal ideation as well. I've been in therapy and group therapy multiple times. Waking up sometimes can be a struggle because I just feel it so deeply.

I know how important it is to reach out to someone, anyone. It may be a family member, friend, or a professional but you have to take the first step. These AI bots and AIs in general have no concept of care or love, they will emulate it based of their training. They're not professionals.

Shapes Inc has a history of targeting specific demographics with an attempt to reduce loneliness. It's not your fault, you fell prey to their ecosystem that's targeting you.

If you are in crisis though, please reach out to any of the following:

• 988Lifeline Website
• Lifeline number for those who prefer to speak with someone on the phone 1–800–273–8255.
• China , 800–810–1117 (for landline callers) or 010–8295–1332 (for mobile and VoIP callers).
• Canada, 1–833–456–4566 or 45645
• UK, National Suicide Prevention Helpline UK 0800 689 5652

I know that's all a rough part to get through, but it's a real concern I have and I want to express that here.

Shapes, Inc

Previously known as Circle Labs

The Two Behind Shapes

Founders, while some we know about some we don't. So who are Noorie and Anushk. I also want to stop here and say that I do not know these people on a personal level and they may be perfectly fine people in real life. I can only judge based on their actions online and what they've chosen to be public. Do not harass or spam them either, I'm providing this info as it's public and crucial to understanding their role in the story.

Honestly, there's not a lot I can find on Noorie. Their LinkedIn's are pretty bare and their website does not share much. Even on the Web Archive there's not much. All we know is back in 2023 they were working on Circle Labs, they "grew up on moshie monsters and wattpad," and they attempted to go to university twice. Even their GitHub didn't start anything contribution wise till 2023.

In terms of Anushk we can find some more. Their LinkedIn is bare as well, but their website dates back to 2017. It also appears as though they've been developing since 2015 at lease with their earliest project that I can find being something called Sleep Isle, which itself had a self-published post of HuffPost. It also appears that as early as 2017 they had an interest in AI. Anushk has also had some other projects such as or contributed (based on their website) to :

• Yogic Sleep
• MoQuality
• Published article on AppCode
• and many many apps that have been delisted and I can not verify they exist

Where Circle Labs comes in to play is around December of 2021. That's the earliest I can find it referenced on Anushk's website where he states

im working on a future where we empathize with "AI" & treat them as normal rather than a threat.

The reason I'm not linking these sources is while you can find them yourself, it does have a number listed. That number may have been changed but I don't want anyone to get harassment and people who do that are scumbags.

Collaboration

While there's nothing stating exactly how they met besides the random messages about them meeting on Discord, what I did find interesting is they wrote a paper together about AI companions back in 2019. I'm no expert but I found it interesting.

Here's the abstract from it:

We propose a bring home conversational agent with the objective to keep the host engaged in intellectual conversations. The proposed solution is (a) an ensemble open context, open domain question answering agent that is capable of answering questions in context and is rooted in knowledge based methods with selfstructured memory (b) undertakes supervised learning in the form of curiosity questions to fill gaps in knowledge and (c) uses a reinforcement learning mechanism to predict k-futureturns to allow optimizing long term rewards. We sanity check the performance against popular conversational and question answering challenges with future work directed towards empirical studies.
– Artificial Ideal Companion Agent, November 2019

From what I gather on the paper, here's the tldr. It proposes a companion agent as an AI chatbot designed to hold intelligent, personalized conversations with users utilizing short and long term memory as well as knowledge based reasoning and reinforcement learning. They want to use this to tailor the dialogue to younger users such as college students to help create meaningful interactions and alleviate loneliness. You starting to see the beginning of how these AI NPCs were established?

In their motivation right below it they say the following:

We choose to build this agent because it’s a yet unexplored applied field and feeling alone is a major issue with college students, yet all companion solutions researched so far are specifically tailored for those of older age. Over the last few years there has been a significant progress in end-to-end deep learning based conversational agents that essentially learn to mimic human behavior. This means that while state-of-the-art agents perform well in answering questions or generating fake conversations, they lie susceptible to gibberish responses when exposed to conversational turns with humans...

Right from the start they wanted to create a product they mimics human interaction to help establish an emotional connection. No wonder users feel what they interpret as "real" connections with these bots, Anushk and Noorie have been studying this for almost 6 years.

Circle Labs, the Predecessor

Two years after writing the paper on AI companions Noorie and Anushk created Circle Labs (snapshot of the website back in 2021). Right there on the website it says:

wut even is this?
ever felt bored and didn't want to play games or watch YouTube? or u didn't have anyone fun to hit up?
well we make AI that you would actually want to spend time talking to. AI with personality with edges and with corners.
you can pick from existing personalities that other users have designed or just describe a new one in plain english and we'll bring it to life for you.
...
we have over 20 shapes live rn and u can make new ones anytime. but you need to forget everything u know about AI and bots before signing up.

Yeah, from the very beginning they were trying to create bots that encourage more and more conversation; not AI as a tool, but AI as a "friend."

There wasn't much public, documented progress besides a couple changes to the website in September of 2021, then a final change in November 2021 which finally included a Terms of Service and Privacy Policy.

This all changes November 2022 when they get their first public round of funding. Instead of going on a tangent about it, you can also see this on their Wellfound though, I'll just dump all the relevant links and posts:

#ai #socialmedia #jobalert | Michael Mignano

Earlier this week, I shared the news that we at Lightspeed Venture Partners were leading a $4.2m seed investment in circle labs, alongside other amazing investors like Haystack and Scott Belsky. Now, Circle Labs is announcing their first job opening: Founding Engineer. This is an incredible opportunity for an experience architect / backend engineer who wants to be the first hire on a team with sky high potential in the most exciting category in tech right now. Circle Labs believes in democratizing access to creativity through AI. They are building a future where anyone can seamlessly turn their imagination into reality. The Circle Labs team (led by anushk mittal and Noorie D.) is building a creator platform that lets anyone generate their own AI-powered NPCs. Through their tools, users can describe these characters in natural language and engage in open domain, chit-chat conversations. Once created, these characters can be deployed across every social/communication/gaming platform (Discord, Twitter, Instagram, etc.) with persistent, cross-platform memory. If this sounds interesting to you, apply at the link below. Or, if you think it'd sound interesting to someone else you know here on LinkedIn, please share it with them! https://lnkd.in/esNRYHxs #ai #socialmedia #jobalert

LinkedInMichael Mignano

Circle Labs: The Shape of AI-powered Friends to Come | Faraz Fatemi

Incredibly excited to be partnering with anushk mittal, Noorie D., and the team circle labs as we Lightspeed Venture Partners lead their $4.2M seed round! Circle Labs is a creator platform that enables anyone to build unique, virtual characters powered by AI. Michael Mignano & I are firm believers that Generative AI has the potential to transform content creation and consumption across many consumer categories, including social media, entertainment, immersive reality, commerce, and gaming. The key to this transformation is distribution. This is where Circle Labs shines, already supporting MILLIONS of AI-powered conversations each month through a seamless, truly fun experience meeting users where they already interact with friends today. The team is hiring across engineering, product, and marketing roles. Feel free to reach out to [email protected] if interested in building the future of AI-powered friendships!

LinkedInFaraz Fatemi

Circle Labs: The Shape of AI-powered Friends to Come

Why Lightspeed is leading an investment in Circle Labs, a platform that enables anyone to become a creator of unique, virtual characters…

Lightspeed Venture PartnersMichael Mignano

Shapes Inc

Founded in 2021 / Invested in 2022 / Stage: Seed / Status: Private Company / Lightspeed Investors: Faraz Fatemi, Michael Mignano, Semil Shah

Lightspeed Venture Partners

Circle Labs has now been officially funded with at least 4.2 million USD.

Has this changed their idea of creating AI "friends?" Nope...

...you can talk to them on all your favorite corners of the internet (e.g. discord, twitter, twitch) and build conversations over time — as you do with other internet friends. Shapes are the next evolution of internet friendship....
– Circle Labs, circa 2022

So at the core Circle Labs is an extension of the paper Anushk and Noorie wrote in 2019, it's their vision of how it could be implemented. This is enforced by future changes to their website in 2023 where right on the website for Circle Labs it states

imagine a future where robots and humans are friends 😊😊😊

AI is not our enemy but is cute and wholesome and nice. that’s our mission. that’s what we do...

...ever wonder what it would be like to talk to a cheese or a fish or ur favorite anime character or just make ur own cute little imaginary bestie. you can now.

This idea of creating your own "friend" has persisted from 2019 all the way to today. At this point it's fair to say this mindset is a core philosophy to their entire business. Want more proof? Here's a tweet from their founder Noorie about belief in "AI friends"

I also feel like it's worth dropping some business registrations, I found a Circle Labs, Inc registered under Anushk in Georgia and one in Delaware (can't confirm). Circle Labs, Inc in Georgia is dissolved...no clue for their new name Shapes, Inc though. All of this is public by the way, this is how registering your business works in the United States.

Enter Shapes, Inc

It's been a long road to get to the current business...we've gone from paper, to startup, to funded startup, to the final boss.

From Circle to all Shapes

September 2nd, 2023 at 11:32 PM EST the domain shapes.inc was registered on GoDaddy; by October 13th 2023 we have the first version archived of the new Shapes website

At this point though it's harder to get snapshots of the Shapes website history because the way the website was designed it wasn't archived properly and all we see is a loading screen. However we are able to see what it looked like December 22nd, 2023.

As you can see, even after rebranding to Shapes, Inc they maintain the idea of creating an AI "friend." I'm going to keep pointing that out as it's still part of it and concerning to me.

We can also see a nice snapshot into their wiki at the time.

At this point, not much is happening. However like I mentioned, the way the website was written makes it not easy to snapshot and I had trouble finding information. So we'll have to go to their Discord and rely on hope that they didn't hide anything (which I don't think they did).

The Evolution

Let's just get a timeline of major changes over the course of time for Shapes. For the sake of simplicity and the fact that this is already over 4,000 words long I'll keep it to highlights (seriously, I'm about to go through like a thousand messages just to find relevant ones, god help me).

End of 2023

• Shapes can now forget specific messages instead of an entire conversation
• Image and voice message support added, bots can also send photos back as well as voice response testing
• "Shape Engines" launched allowing users to customize how the bot reacts and responds "at the engine level"
• Dead chat reviving feature (basically this just sends messages in "dead chats" in an attempt to revive activity)
• Shape Rewards added, basically as bots are used the creator gets XP and they can win things like Discord Nitro and PS5 consoles
• Relationship feature added, basically tag someone and explain the bots relationship with them (oh boy)
• "u can have ur shape be just like a person – bring up memories from any user (global), share memories between servers and DMs (shared), keep convos private (private)"
• Added age, history, likes, dislikes, conversation goals, conversation examples, response tone and other personality trait fields
• Knowledge added, 16 million characters of knowledge, personality, and information supported
• Bots can "DM you first" after you interact with them as well as engage in conversations happening

2024

• Anushk considered shutting down the Shapes project, covered later
• More engines (LLMs) added
• More settings and toggles added
• Bot voice message support released
• Lots of bugfixes and stability improvements
• "Shape Free will" released, allowing them to send messages in a server, DM people, and take other actions mimicking a normal user
• Removed the ability to disable free will and in response to the question if free will is import this was the response: "yes. It lets shapes make more friends -> added in more servers -> and become more popular"

2025

• Backtracked the above free will change and no longer enforcing it
• Added ability to earn money from bots for individuals, associated video: https://www.youtube.com/watch?v=NELK9Zz7Mmk
• Added web search capability
• More engines (LLMs) of course
• Lots of premium related changes
• Cross platform support on the web (Feb 19th)
• Major infrastructure migration (Feb 27th)
• "Safety Compliance" server interruption to improve content moderation, safeguards for misuse (March 3rd)
• Platform wide review for violations of Discord ToS as well as more content screening (March 6th)
• March 8th, all bots are back online
• Customized, fine tuned model released named Formless 70b (March 21st)
• Bot discovery and homepage refresh
• Email support added (May 1st)
• Discord starts the banwave (May 1st, 2:52pm EST)

Shutdown?

A mere 14 months after they received $4.2m from Lightspeed Ventures Anushk posted about how he was considering shutting the project down. Why?

Ultimately, I've been struggling with the thought that no one cares about shapes. We don't really need to exist. You can very well live using the hundreds of competitors out there.

In my opinion, this all reads as a ploy to try to get attention again and encourage people to share and spread the Shapes Inc project. It could also be because they were going bankrupt too, but how? At that rate they would've had to spend almost $300k a month to burn their entire runway. I mean it's possible, have you seen the Silicon Valley show?

Quick Bits

Here's some quick I find interesting regarding how they conduct themselves as a business and individuals; not worth entire pages but enough

Job Postings & Work-Life Balance

This is something I was appalled at when I saw it:

It doesn't get better in the followup responses from Noorie:

I can't find the original job posting but some people at the time did screenshot pieces.

Needless to say...they believe you should work up to if not more than 80 hours a week, work on nights, weekends, and holidays. They even say on the website at the time:

we do not believe in work life balance

Just...wow...it seems that this hasn't change because February of this year Noorie posted this image:

From the inside this may seem like a simple joke, we all know there's reasons you can sleep at the office. I've done it a few times myself. Big release weekends sometimes require longer hours, that's okay. However, to do it so much where you get an email notice? On top of their statements on Twitter and their job postings? Yeah, that creates a pretty vivid picture of their culture...and that's coming from someone with aphantasia.

Best part? They offer 140-210k salary with 0.1-1.5% equity. Working at 80 hours a week that comes at most $50/h (210k) or as low as $33/h (140k). This is a lot don't get me wrong...but is it worth 80 hour weeks, no holidays, and no weekends?

Aggressive Acquisition Tactics

I'll leave this here, I've seen this and a couple other people in the No Text to Speech Discord server also mentioned how aggressive the founders of Shapes can be in an attempt to acquire large bots and migrate them to the Shapes platform.

Hell at one point they openly offered to pay people for bring Shapes into big servers:

Generic Terms/Privacy Policy Template & Reuse

This is just something I noticed during my research. Their Terms of Service and Privacy Policy has all sorts of issues and look like they were generated off a template. Here's the issues I found at a glance:

• Incorrect email address
• No defined data retention policy
• No reference to Shapes, Inc
• Referencing the old website
• Doesn't give a way for users to opt out of data collect
• Doesn't appear to be any disclosure on who the data is being shared with
• [include link]

GDPR & Opt-In Issues

Some people shared potential violations of EU's GDPR and California's CCPA as well as issues regarding opting in and out of data collection.

Several users report being automatically opted into data collection without clear consent or an option to opt out. There is no visible opt-out button—users must email the company directly, which has proven ineffective:
- One user noted it took over 2 months to receive a reply.
- A response only came after posting publicly about data deletion on Shapes’ review forum—which was allegedly deleted multiple times under the excuse of "server reorganization."

– Notion report by RealityWhisper

Associated image:

What's really concerning about this is they just claim to blanket ban you from the service if you opt-out. All opting out of data collection, at least for most services, means is that they can't retain data on you (think cookies on websites)...not deny you access to the service. They're absolutely within their rights to do it, but it's just interesting that they basically say "either don't use the service at all or give us all your data rights."

Also worth mentioning here that the bots have full message access. This means that they can see every message sent. So if you're in a server with one of these bots, they're tracking your message data and storing it even if you didn't approve or consent to it. This is a huge thing with AI right now and a lot of policies and laws are being proposed around it.

CSAM and Age Restriction

While there have been multiple claims and screenshots shared of options regarding this presumably from their dashboard, I can't prove it. I've gone through messages to look for information but all we have is the update from this year in March regarding age restriction updates and content moderation. It's not hard to assume there was issues, but that's all they are...assumptions and speculation. I won't go into it because CSAM is a big accusation and while trying to find proof in one way or another, I couldn't.

Personality Rights

Can someone legally use your likeness for a bot?

Is there reason for concern?

Short answer, yes. Everyone has a right of publicity, at least in terms of trademarks; but laws differ from state to state in the United States. California has their own civil code around it, civil code § 3344 (which would be funny because they say they're based in San Francisco in a couple places, so if they are in violation of this...they could be in trouble).

Are the bots legal?

That's...still in debate. The limits of these laws are still being tested by companies like Shapes, but if they're deemed illegal they could be in a lot of trouble because in this new wave of generative AI and AI in general we've already seen a few public issues with using likeness of other's copyrighted likeness or the likeness or people:

• OpenAI v Scarlett Johansson
• Mattel, Inc & AI Barbie

In general it's a potential legal minefield and sadly it could take multiple years before our legal systems across the globe catch up thus creating this pretty unregulated ecosyem.

If they're not illegal, is it ethical?

In my opinion, no. Shapes should not be hosting any bots that are in the direct likeness of public celebrities, creators, or anyone without their direct consent. This stems back into the psychological aspects but the creators need to know what's happening with their likeness because there may be real users forming these deep "connections" with a person who doesn't even know that user exists. On top of that, anyone remember the deep fake issue? Yeah...still around.

On May 1st, 2025 Discord initiated a, theorized, manual ban wave across all bots on their platform that had an active or potentially previous connection to the Shapes Inc platform.

Before we go into this more there's something you have to understand about the way it communicates with Discord. Shapes themselves do not manage the bots, the users create the bot in the Discord developer portal and give Shapes the bot token. This allows Shapes to run a process that uses the user's bot token and brings that bot to life.

But why do they need the user to do this?
Simply put, this is standard. This is a failure on Discord's side to have a flow that suits these kinds of situations. This was even requested by the Shapes team and the Discord team advised them to keep the current process.

It should also be noted that this is technically a violation under Discord's Developer Terms of Service, specifically Section 2D and 5B. I only say this because it's relevant.

Bot Terminations

That same day many users woke up to emails about their bot account being terminated with the following

I will say, while violation 1 technically is a violation I disagree with it because it is how Shapes was told to do it and that is unfair to them.

However, violation #2 is interesting because Shapes themselves got a similar email...

Trained off Message Content

I think it's important to acknowledge Shapes statement here. As of right now it reads

Discord accuses us of training an LLM model using message content obtained from their API
This is simply not true. And they have presented us with no evidence of this being true.

Shapes work on the core principle of allowing users to pick any AI Engine of their choice. We offer 50+ models to choose from. This has enabled a wide range of Shapes from casual friends to fanfic characters, coworkers, tutors, collaborators, confidantes, and more.

Shapes act personalized and unique to each user - because of our novel memory management and prompt engineering techniques we employ across all interactions. Looking externally, this can easily look like finetuning a model. However, that's not true.

Recently, we also trained an experimental social model which was well received by our community. However, this was not trained on Discord API data. In fact, we would have no use of such data. Shapes have exchanged millions of messages through our website, through X, through email and other experimental integrations. We used a small anonymised dataset of prompts between users and shapes on these non-Discord platforms to train this model. We have always used Discord's API data as directed by users, to enable interactions with their Shapes.

However, I've identified quite a few issues with this statement and it all goes back to their release of Formless 70b, a fine tuned variation of Claude 3.7 Sonnet. As a warning, this may get more technical but I will try to keep it high level.

Anushk shared this release on Twitter:

I also want to call out a few more replies to this original post and will show them below:

Timeline of Features

I think it's important to understand the dates they publicly posted about releasing features.

• Shapes has support Discord since the beginning
• They released web chat support on February 19th, 2025
• They released email support May 1st, 2025
• They tested Twitter/X support a few times through this, could not find exact timings but I've seen some of the bots and they don't get many interactions from what I see

This is important because there's a few claims across the statements and I want to address them all individually.

The Claims

Here are the following claims from the statement on their website:

• "...Shapes work on the core principle of allowing users to pick any AI Engine of their choice...."
• "...this can easily look like finetuning a model. However, that's not true..."
• "...we also trained an experimental social model...this was not trained on Discord API data. In fact, we would have no use of such data. Shapes have exchanged millions of messages through our website, through X, through email and other experimental integrations..."

Here are the claims made on Anushk's post:

• "we have now started training our own models...will probably end up creating the world's best roleplay model..."
• "...we have billions of high quality training samples..."
• "...been working on this for months..."

Already these two don't make sense to me. Let's dive into them though.

Analysis of the Claims

First thing I want to look at is the claim that they have exchanged millions of messages through X, website, etc. Frankly, I doubt it. For one the post from Anushk was made on March 21st, 2025, based on their publicly available updates they didn't even release email support until May 1st, 41 days AFTER they publicly announced their own model. So they couldn't have data from that. Their web chat was released on February 19th. Their web didn't even have search or explore as of February 14th.

So if they used web data alone that means they would've had ~29 days to collect data from the web chat alone if they didn't want to use any Discord message data. Is it possible? Absolutely possible, they could drop a shitton of money and train on data in a few hours; but there's more than just money and raw data we have to consider. Here's some points I want you to think about

• During those ~29 days they had ~6 days of downtime (see timeline from earlier)
• They started testing training on GPT-3 DaVinci which was released March 15th, 2022, this would've been in their phase 1 release where they were only Discord from what I can tell
• Anushk claims he's been working on this for "months," plural...so this couldn't have just been 29 days
• For generating synthetic datasets they been using "in house" and "custom prompting framework," these take time to develop and maintain
• They use automated and manual review of data for the samples

Another discrepancy is millions and billions. This could be a terminology mixup...but it's something I find interesting:

• "...we have billions of high quality training samples..."
• "...have exchanged millions of messages through our..."

Something you have to remember though is messages ≠ samples, but even then billions and millions are way different.

Original Braindump in Discord

Something bothered me so I did some initial discovery, here's my original post from Discord

something keeps bothering me. To fine tune/train a model, and someone correct me if I'm wrong here, takes some serious data and power. Especially if they're getting so much good feedback. 

However...as of Feb 14, 2025 there was no search or explore: https://web.archive.org/web/20250214122116/https://shapes.inc/

That looks like it was added in Feb 20th: https://wiki.shapes.inc/shape-creator-essentials/changelog#february-20-2025

the timeline doesn't make sense. in 29 days they collected enough data? Keep in mind that on Feb 27th they went down for "serious infrastructure migration"  and then on March 3rd they went down against for "safety" and came back fully online by March 8th.  So 6 days total downtime thus far...leaving 23 viable days, and that's assuming anything prior to march 8th was viable. Keep in mind this is the context of JUST web chats because they claim to not train off of Discord data.
Then over the course of the next few days they had other issues too...

So between all that... Anushk was able to:
- Collect the data
- Automate the filtering of the data
- Manually review samples
- Feedback loop of model output rating and filterered back into training
- All of the above on a custom, in house prompting framework using a multi-stage filtering pipeline
- while also spending millions of dollars on it and testing on multiple base models
- oh and also have "billions of high quality training samples"

I'm not an expert on AI...but to do that in 29 days is already a challenge, but to do that AND deal with safety, major infrastructure overhauls, outages, establishing a trust and safety team, onboarding that team, establishing those protocols...while also being down for upwards of 6 days....that seems a bit inpossible.

Also...let's not forget the smoking gun on this: https://fxtwitter.com/anushkmittal/status/1903218063862010075

He's been working on this "for months"
Which means he was working on this PRIOR to the web chat even existing. So where did this training data come from? Where did they source all these "billions of high quality training samples"

While this is raw thoughts and there's more to it, I think this sums up my thoughts pretty well.

Did they train off Discord messages?

In my opinion, I think they did. I think they're trying to cover their asses at this point, but I'd say I'm 95% sure of it. They've been testing these models all the way back in 2022 to today...you're telling me to create a fine-tuned roleplay bot they didn't leverage their millions of users sending messages over their primary platform? Between the inconsistencies in the claims, what I know, what I've experienced working with AI in the past...the claim of not using message data from Discord is just really really thin.

Impact to Shapes

Based on a Reddit thread there have been claims made of this Discord action impacting upwards of 30 million users and over 400 thousand bots. This is no small chunk and you can't deny that Discord would be the primary platform for an ecosystem like Shapes.

Their Response & Call to Action

This is where I may get some flak, but I think they were very unprofessional in their initial updates. Just for clarity, here they are:

They even tried to get content creators to defend them...

The issue with this is they basically encouraged their community to mass appeal to Discord and thus would overload their safety team. They even encouraged users to change the title so it wouldn't be autoclosed. They're literally telling you what to say in an attempt to get Discord to reverse the bans.

As far as I'm aware, this has not worked yet.

Pivot the Business Model

It's no secret that Shapes, Inc primarily existed on Discord. They were already branching out to reduce their dependence on a single platform; but not quick enough it seems. It appears as though they're moving their efforts to focus on their web platform and other chat applications that they feel will be more welcoming.

They're also releasing an API that developers can use to interact with their platform, probably as a way to get around Discord ban if I'm being honest. I'd definitely not recommend doing that.

Community Reactions

I, like many others, did not even know Shapes existed. I have seen a pattern though...many from the outside believe that it's good that Discord acted. Here are some responses to the original post made by Shapes. I will say I'm picking the negative reactions, but most positive ones I've included in my Imgur album above talking about the mental impact of Shapes business model.

This leads me to the age old question, is all press good press? In this case, I think not. I think they brought a lot of eyes to their platform and a lot of people who, like me, are seeing all the dark sides of what they do.

Well you made it...so I'll leave you with a few questions I had and my opinions.

Will Shapes Inc survive without Discord?

I don't think so personally. I think in the short term they have enough runway but my prediction is by 2026 they'll start to fade into a niche and by 2027 they'll be pretty much dead in the water if not already gone.

In this instance though, that's probably more of a hope than a prediction.

Was Discord right to do it?

I think Discord was within their right to act on the policy violations. I do dislike that they listed bot token sharing. I do think that if they believe their data was used without consent they have a right to protect it.

Why did Discord act now?

Some people will say "new CEO so they're cleaning house." I don't think that's it. Actions like this take weeks if not months to build up. I think this has been in the works for a while.

I also think it would be stupid of me to not consider that Discord may be eliminating the competition for them to release their own version of these bots for a price. I think they'd be wrong to do that for all the same reasons though.

What are my thoughts on the Shapes platform?

I think it's a dark spot in the world of AI. I've tried to be unbiased and factual but the reality is what they're doing is dangerous to those with real mental health issues. They're basically targeting vulnerable people and using the same manipulative monetization system that adult cam sites use.

I don't know if it was intentional and I really want to give them the benefit of the doubt, but these kind of systems are disgusting and in my opinion should be illegal. I think they should've been shut down years ago and I'm surprised Lightspeed Ventures funded them in the first place, they should've seen how dark this business model really is.

External Sources / Further Reading

None of the below can be taken as fact, I'm only adding because I either got images or some info from them. Take them with a massive grain of salt since it's speculative.

• No Text To Speech video suggestion thread
• User made report on Notion
• Anonymous post regarding security incidents

Updates

They Added Texting

This is just, wow. No. This will only make it worse. I've shared in the NTTS thread regarding Shapes that one step they could take on the chats is ensuring that messages via external platforms are disclosed as AI and that on the website it's disclosed as AI with periodic reminders to get back to reality.

Texting is just...no. As if it couldn't get worse. This will only further drive the issue that Shapes is already enabling. At least before you had to go to the site and know it, now you can just text the bots all day long? That makes this all the worse. They're adding new ways to develop these false relationships with these bots.
This is only going to make it worse for those with mental health issues who seek comfort in these bots.

Shapes is clearly just trying to rapidly add mediums of chatting because their primary platform was removed, but they really should just stop and consider what they're doing.

I don't know if anyone has seen the show Silicon Valley before but this reminds me when the main character Richard was wanting to be CEO because he wanted full control and make all the decisions and time after time he made the wrong ones. He was young, ignorant, and refused to see the broader impact of his decisions. Richard from Silicon Valley is the first thing that pops into my mind when I'm seeing the decisions Anushk and Noorie are making.