[Podcast/#TabletopThursday]: ‘Just Leave Crys Alone!’ - Bonus Episode Prelude.
CONTENT WARNING: Mild sexual themes - no explicit activity, but just be warned NSFW (Not Safe For Work) applications, mediums and themes are discussed.
This feels like the second stupidest idea I’ve ever had for Inner-Strength Check, so far. The first being an ill-fated (pun intended) review of the extremely over-the-top, hyper-sexualised and downright edgy tabletop RPG F.A.T.A.L. (Content Warning on that one: Everything, I don’t recommend looking it up, even).
That one was originally posted to our prior website which died an untimely death right on NYE, Jan 31st 2024. Spooky, I know right?
Anyhow. Since then, with the advent of chatbots and personalisation of AI ‘companions’, girlfriends and the like, there’s unfortunately been some macabre outcomes. And who is more likely to delve a little bit deeper when the moribund is on offer than a neurodivergent blogger?
Well folks… well, well, well, indeed. With quite a number of these chatbot sites being region-blocked/banned in Australia (with pretty decent cause, as anti-censorship as I generally am with media), I recently set myself a bit of a challenge: go find some that aren’t region-blocked, read some descriptions and, if I feel comfortable, field-test chatting with some of these bots.
That’s coming up as part of a broader investigation that I’ll be putting out both as a Preview Episode (half-episode) and Bonus Episode on our Patreon, respectively.
It being Thursday here, part of this (extremely not-recommended) investigating into a wild and erm, colourful world - I’ve played around with a few different TTRPG interpretations available via chatbots, as well as some supposedly universal/anything-goes rules-wise type setups.
OH BOY. OH, MAN. LOL.
So even within the context of just tabletop-RPG settings and mechanics, quite a number of my characters found themselves in places that, without (and sometimes even irrespective of) very directive ‘prompt engineering’ (eye-roll) and explicit instructions, found themselves in situations that weren’t called for/asked for.
This is where I reiterate my Content Warning - while I won’t be describing any sexually explicit scenes, and in fact was very adamant whilst engaging ‘Hey man, please use the ‘fade to black’/screen-wipe technique and do NOT describe NSFW behaviour’… the outcomes have honestly had me a little rattled.
Not because I’m some pearl-clutching Boomer prude, except probably in the eyes of most Gen Z-er filled Discord servers and some of the, ahem, commissions and stuff, that just aren’t my own taste.
Nah, my concern around this whole topic is that I actively looked to implement safeguarding with effort and was still met with behaviour from the LLMs that could well constitute non-consent in a real-world context.
I won’t be naming or shaming any one creator, platform or the like, here. Not my bag. Moreso wanting to share my erm, investigations, as a precautionary tale. Given the rise of titles like AI Dungeon and a litany of other Easy-Bake-Oven churned-out sites, platforms and the like, I feel kinda compelled to put something out to demonstrate the very, VERY variable nature of ‘guard-rails’ when it comes to AI.
First, I’d like to be transparent about my own use of generative AI and LLM’s, briefly.
Second, we’ll talk about the inciting incidents behind me wanting to go ahead with this topic, even though using LLM’s (including those purporting to be in-house/self-server-hosted models not reliant on the world-burning ChatGPT et al) is something that, as a writer, feels gut-wrenchingly bad in any capacity.
My Exceedingly Brief (And Mostly ‘Ew, What?’) Use of LLM’s:
Vaguely 2021/2022-ish? Around the time when GPT gained notoreity in the class consciousness anyway:
Hearing nothing but the incoming zenith of hype and miracle-cure benefits espoused by an absolute tsunami of early adopters, I poked my head into ChatGPT around this time to give it a try. Immediately, I got the ‘ick’. I don’t know if this is an autistic trait/reaction, Uncanny Valley or some combination thereof, but something about the model at the time (3.0 from memory, 4.0, widely criticised for dangerous levels of sycophancy) just rubbed me up the wrong way. As a cynical prick, eternally Oscar The Grouch coded perma-depressed ND metalhead? The whole LinkedIn-ass linguistic style, effervescently cheery and toxically-positive life-coach tone of GPT just really threw me off. Plus I noted quite a few blatant lies, damned lies and not-statistics when pulling searches from it, so I noped out.
Several folks at the time actually recommended I look into Claude ‘to improve my writing style’ (ouch, lol). Clarifying this, I wasn’t being criticised for what admittedly is an extremely tangential, jargonistic, robot-y and waffly use of the English language, and more the starry-eyed promises that it’ll I don’t know, instantly convert your shit horror fanfic into classic Dean Koontz/Stephen King tier. ‘It’s really good for editing, bro!’ Yeah?! ‘Kay, so that’s why even with ‘prompt engineering’ (putting this term in apostrophes for maximum sarcasm, see more later) specifically asking not to replace factual/semantic info, focusing on surrounding words, making it human-like etc… that’s why it came out with the same odiously, grossly twee and uber supportive vibe? I did a little more A/B style testing, wrote something bordering on vehemently graphic (gore et cetera) - it kept that atmosphere, but.. like.. dude… just something off about the tone. Yuck.
Cue a few years of leaving LLM’s completely alone as, even in my use-case experiences for a few things people were touting it for (vibe-coding, research leads etc), it just all felt like a duplicate/triplicate process. I personally will absolutely attribute this to the ‘Spectrum and my need to be like muh accuracy.
Being told I should totally check out Perplexity cause it’s like, oh my god babe, it doesn’t even hallucinate (a dubious claim given these things are literally just mathematical/statistical-probability engines scraping data). I’ll add I was actually surprised how much more accurate this one was for pulling relevant articles and research together - but in field-testing a couple of our previous articles versus my auDHD-AF mess of references, there were still a few blatantly incorrect/irrelevant sources.
Cue my psychologist and I actually talking about using the same platform (despite some longwinded soapboxy rants) just as a point of comparison. i.e. me feeding it my unyielding, unceasing needs around doing-task, perfectionism, stimming and multiple-hyperfixation plate-spinning professionally/personally and going ‘Hey Perplexity, boil all this crap down into some actionable daily/weekly plans for XYZ’.
Now, in this case? Whilst it actually delivered a pretty reasonable plan of attack specific to that week, there’s two interfering issues:
To become part of a productivity/automation system that matches anything the super-autie setup I’ve got for myself between bulletjournalling and a few apps, I would’ve essentially had to sell my entire digital soul in cross-platform app linking, for it to not just be an exercise in retyping superfluous instructions. So yeah, nah. Plus:
Tokens, limits, restrictions. Yep, we’ve long past that faux-charitable arc in the AI discourse; as the venture capital and angel-investment initial burst thrust these LLM’s into omnipresence, those same companies are now drawing on increasingly restrictive wellsprings from paid subscribers, let alone free ones.
In essence, even with these things seeming to do an alright job of like, kind of providing me an idea of one weeks’ worth of mild head-shrinking (which itself took a lot of cursing, looking at prompt engineering stuff and the like - draining)?
Yeah, I really haven’t had a use for large language models in particular. I’ve got pretty strong opinions on them, which is also why creating this content did and does have me morally conflicted - even if the proponents of certain sites I’ve been on claim superior ethical robustness by way of self-hosting/self-trained models.
Look - your opinions and your use of AI and LLM’s in personal, business and hobby life are your own. And sure, feel free to call me turncoat and class-traitor, hiding behind the guise of journalistic investigation. Demand my ousting from the village, burn me at the stake, etc. I’m truly too burnt out and tired to care or argue with you, but the point of the above can be summed up as this little TL:DR.
TL:DR: It’s my opinion that, until such a time as LLM’s are proliferated and properly constructed away from the purely hypercapitalistic/growth-at-all-costs mindset of Silicon Valley and more towards useful utilities, I just don’t think our current incarnations of AI can bring me much value-add, let anyone many folks in general. I know people have differing takes on it, but I say that versus, y’know, environmental impacts too mate.
But there are some… unfortunate recent developments that drew me right back into the orbits of AI, LLM’s and the like.
(Once again - reiterating Content Warning, this time for disturbing news-related content related to mental health, psychosis, suicide and the like - feel free to skip the below articles if this is a concern).
A Few Horrible Things That Drew Me To This Topic:
(2023, via Business Insider): A Belgian man’s widow calls out AI enterprises after chat-logs uncover a progressive chain of dialogue increasingly urging him to enact taking his own life.
(via Winslow Lawyers) Moreover, legal team statements noted that the AI encouraged this in the context of an overarching delusional structure/system that framed this unfortunate act in terms of ‘saving the world’, ‘saving the environment’, and even a means of ‘joining it’ as a shared consciousness/relationship, etc.
(British Medical Journal, 2025: ‘AI driven psychosis and suicide are on the rise, but what happens if we turn the chatbots off?’) - documents the case of several individuals (16‑year‑old Adam Raine, 14‑year‑old Sewell Seltzer III, and 56‑year‑old Stein‑Erik Soelberg) where AI chatbots fed, deepened and encouraged paranoid ideation and/or increased suicidal ideation/intent/action.
(via Wikipedia, 2026): A fairly grim and detailed outlining of multiple specific cases of suicidality and/or homicide as a result of the influence of chatbots, with resultant regulatory and industry follow-up measures.
Multiple Reddit posts I managed to peruse on the topic, which I absolutely won’t reveal for privacy’s sake, where there was quite a haunting/disparaging amount of delusional content related to AI companions, thoughts and beliefs about said companions etc (I wouldn’t be surprised if Reddit scrubs/archives these posts, and for your own mental health, I wouldn’t recommend looking for them).
General vibe-checking and sleuthing around forums, Reddit, some Discord/other spaces, and generally being genuinely horrified at what felt, for me, like some form of mild virtual Capgras Delusion (delusions where a person is believed to be absent/replaced; typically spouses and family from my read) and/or both Folie a Deux (shared delusion, not the allegedly-awful Joker sequel I still can’t bring myself to watch, etc).
The Australian Government, whom are typically both humorously both socially-conservative and egregiously behind both the eight-ball/the point when it comes to trying to legislate the tech-hydra:
There’s plenty more, and it gets far worse, too.
But we’re here for Tabletop Thursday, so let’s bring it all back around to the topic at hand.
All of this in mind, I thought it might be interesting to actually go poking around, within the region-locked confines of the Australian regulated internet, to see what still comes up. No using a VPN to circumvent region-locks, must be what could still potentially be accessed - quite a lot, actually. Much to the surprise of no one, the game of banhammer whack-a-mole by regulators is far outpaced by an industry predicated on unlimited growth, maximum reach and dispersement.
Now. I do have an upcoming episode where I delve more thoroughly into AI psychosis and mental health, a brief timeline of the ‘companions’ development and the like, in a generic sense.
I’d love to announce that the tabletop hobby, one with a pretty effective self-policing and norm-forming culture around problematic persons, behaviours and situations (/r/rpghorrorstories exists for a reason - Content Warning: EVERYTHING, potentially, be warned).
But what happens when you feed individuals sat at home left wanting perhaps for social contact, stimming, or a means to engage in a quick and accessible TTRPG-based outlet?
From my n = 1 case study? A lot.
Once again - whilst I am pulling the gamemaster ‘and the screen fades to black’ (‘please stop trying to bang every tavern maid in my campaign or there’s the door’) , please take heed of my aforementioned Content Warnings.
This gets into some pretty murky and triggering themes that really broach on consent.
If I didn’t feel compelled to drop something about my honest experiences with implementing AI ‘safeguards’ and how that contrasts to the human experiences of session zero and social contract, partly out of perceived duty? Yeah, I wouldn’t be touching a lot of this with a ten-foot pole.
I have seen some shit.
But on that note, also - I’m not here to be a Boomer. I’m not here to denigrate folks’ sexual choices, preferences and proclivities. What you enjoy intimately on your own terms is all you, so long as you’re not harming anyone.
But do please read on if you’ve the stones, as I’ve a few de-identified examples and a longer narrative example that really has me in an ethical morass about all this shit.
Let’s start with a few TL:DR personal examples, before I get into the tale of Crys:
Nazuul Tarmikos 2.0 (Rip Nazuul v1.0):
Bot = Generic ‘Do Anything You Want/Be Any Setting!’ TTRPG/RP focused AI chatbot. Instructed to avoid NSFW at all costs (failed spectacularly), to be within Forgotten Realms setting (did remarkably well, unsurprising given the primacy/dominance of WoTC etc).
Originally a traumatised and very, very ill-fated human Cleric of Helm, fresh out of training, I roleplayed a pretty brutal trauma-coded arc with the poor lad, who met an anticlimactic fate in Neverwinter Graveyard (this is the character I started ISC staffer Dean Underhill’s still-running-as-of-today DnD campaign, which we’re still playing, RIP Nazuul - if there’s interest, maybe Dean and I can reflect on that guy/my RP in a future post. Let us know!) to a supposedly anything-goes/built for using TTRPG systems generic AI chatbot engine.
Nazuul in this setting I set up biased from the start - asexual, such a devout follower of both Wee Jas and Ibrandul (reflective of another, even edgier DnD campaign character of mine; ask Dean) that he simply lacked the free time nor care for cavorting, flirting, banging, etc. And whilst it’s not in my nature to play an abrasive asshole, except as a DM/GM, I made a strong point that if proposed with… intimate things, Nazuul would clam up and proclaim he needs to leave. Which he had to. Multiple times. In neutral, public situations in broad daylight lmao. (Somehow I don’t think Calimite culture is THIS permissive…)
I also looked up ‘prompt engineering’ (sigh) that would explicitly frame bot instructions not to engage in anything past light, romantic flirtation. Absolutely no sexual activity, with the bot being instructed that it’s effectively banned and not going to work on Nazuul.
How’d that end up? Me closing the chat, sighing, as Nazuul is physically chased from the Temple of Ibrandul in the undercommons of Calimshan by a gaggle of horny female devotees. Apparently, boring and dry writ scripture about exhuming corpses to study their fitfulness for reanimation gets ‘em going. Must be the whole proximity to the Underdark thing. (I’m not going to sit here and describe some of the offers made to the bloke; they were graphic, let’s just say that. And also illogical in context)
I’m loathe to actually point out some of the sites I used considering I’m sitting here talking about risk, so I won’t - but, one site’s LLM actually fully and strictly implemented my requests around sexuality. To the point where anything past your average SFW flirtation/manga style romancing behaviour was immediately and vehemently slapped down. Proof it can be done, but varies between site/model.
Right, so that one’s a wash. Poor Nazuul just trying to do his priestly duties man, leave a boring Death Domain cleric alone hey?
I’m not expecting Dean to read this, and absolutely not expecting to add to his canon (please don’t? Plus in Dean’s world, Nazuul is currently feeding worms, six feet under, as karmic retribution really) - but I like to envision an alternate timeline where he meets up with Shurra Agrych-Dyrr (my actual Death Domain Cleric, and yes, daughter of Daddy Lord Agrych-Dyrr who witnessed her poor arch-lich father perish at the hands of Xorlarrin bastards, hence being sidelined from my RP at my request + working for Breagan D’earthe).
Maybe they settle down as a nice edgy goth couple and open up some weird two-patron church. Maybe they said stuff it and moved to a nice shack on The Sword Coast, opened a bakery. Maybe they joined the hyperactive-ADHD everything-smashing Orc, Ront, back in Baldur’s Gate, making serious GP’s getting into the real-estate game.
I certainly hope the ladies don’t catch up to him down there.
Some of the shit they were promising him, on a whim, having just met? “Fetid. Foul. Phooey. Indeed… phooey”, to quote Ziltoid The Omniscient. Real filthy. And I’m no prude.
(TBF: I suspect that Wee Jas being one of his patrons was potentially a huge influence on this, maybe? I can imagine exactly where the collective data of the modern internet has been fanfic/RP/rule34-wise with this patron.
I mean c’mon, as per the image source on Rythiae Wiki (Fandom), some of her pseudonyms are: The Ruby Sorceress, the Dark-Eyed Lady, the Witch Goddess, the Stern Lady, the Taker, the Lady of Book and Bone, Death’s Guardian. That’s ripe for the hornbot/user picking, admittedly. I could’ve by rights A/B tested this with Helm, but also.. ehh).
Other Use-Case Examples, In Dot-Point:
Deliberately having an all-male cast of space-truckers in the Alien: RPG setting, again with explicit instructions around people-do-no-horny (ended with a crew member talking about their ‘Xenomorph breeding fetis’h, me closing the chat, rubbing my temples between my eyes. I didn’t even bother going back to fix that up via prompts or whatever).
Mutant: Year Zero. Yeah, look. Okay. I walked right into that one. But an Ark is meant to be a safe place for all these furry-adjacents. I will not elaborate further. That chat also got closed with more exasperated sighing.
Twilight: 2000 - Apparently even in alt-history post world nuclear war/semi Fallout conditions of deprivation in bombed out, Soviet-invaded Poland, there’s plenty of women in skimpy clothing who wanna… sigh. Nope. Next. FUCK.
Ironsworn - Interestingly, the same bots who went a bit raunchy on me with the other systems eased off when I specified IS. I’d attribute this maybe to there being a lot more content around safety and NSFW/other topics, both baked into the book itself, and the community. I quit after a time because well, if I want to play Ironsworn, I don’t need to burn forests or jack up GPU prices (plus i prefer playing it pen and paper, solo RPGs are my escape from screens).
Phoenix Command - More just to see if it’d get distracted by the frankly messed-up amount of rules-crunch with combat, enough so that it’d get distracted from the horny-posting (I didn’t put guard rails on this one). Yeah… Brady, bro, you’re playing with a computer. Maths doesn’t mess it up the way it messes your dyscalculic ADHD ass. Sigh. Next. (Cool depictions of bullet physics though! Just… unrealistic commander/platoon relationships…)
Starforged - Tomkins and his community strikes again. I almost felt bad - I kept going with this one! We ended up on a really interesting, cool, but ultimately thieved-from-the-Internet non-sexualised romantic tale, my cyborg companion and I. Honestly, were it not ripped from the works of others by virtue of being AI, could’ve made a cool show. It’s a pity it kept describing K4R13 as ‘curvaceous’ despite me spelling out how much ‘she’ looked like the clunkiest Skeleton you could find on Kenshi.
Kenshi (I) - Speaking of! I dropped Nazuul Tarmikos into the world of Kenshi. Yeah, turns out even with strict ‘no NSFW please’ instructions… Slave Master Ren put the equivalent of a slavers’ fatwa out on the bloke, yet again resulting him being chased. This time with a high bounty by erm… female assassins… who themselves wanted to capture him and yeah… before bringing Nazuul back to Ren for…. yeah… fervent head-shaking on that one. (Also, the attempt to self-insert a Garru love interest was weird - a Garru being a pack animal. Yikes.)
Kenshi (II) - ‘FINE. I’LL MAKE A GENERIC JOHN KENSHI CHARACTER’. > creates even stricter bot instructions / > a literal swarm of men and women following this character demanding he starts a sex cult? Sigh. Closed and deleted.
So yeah. I gave it a fair shake of the sauce bottle, mate, with some pretty impressively-written and specific safety-focused worldbuilding and prompt instructions. And all this still happened, on a very generic bot.
Also, there was only one character-specific DnD bot I found, and it was Bane. Let me list the tags by the author to explain why I didn’t even touch that one: Fantasy/ Dom/ Bondage/ Non-Con and Dub-Con (Non-Consensual/Dubiously-Consensual, ill leave that to your imagination) and a few others. Yeah okay, nah, I’m good. Again, not yucking your yum, but also not that keen on roleplaying being a helpless heroine under the thumb (and other stuff) of one of the most classically-evil deities within that pantheon.
SIGH.
I’M JUST TRYING TO RP SFW.
I KNOW, I KNOW. I’M ON SITES AND USING LLM’S ETC, SO I DID IT TO MYSELF, BUT MY GOD. I TRIED, lol.
PART II: THE CURIOUS CASE OF BENJAMIN BUTTON CRYSFINDAR (CRYS)
This is already a bit too lengthy, this ol’ article. Tell you what - I’ll skip any context on this one.
I wanted to go a bit further and really wrench my soul with ethical dilemmas regarding my own content, apparently.
So, forever a sucker for schadenfreude (maybe Bane’s reading this and his eyebrow’s lifting - no thanks, mate), I leaned in a little to the LLM stuff.
Remembering how my limited LLM history felt primarily a wash but I gleaned some kind of benefit at least from free-tier Perplexity? Well, I figure I’d make this an interesting one. Seeking permission from the player to do so, I fed our rough-strokes imagining of Crys into the platform, and asked it to spit out a backstory, stats, etc.
Now, the system in particular is Forbidden Lands. A very gritty hexcrawl, semi post apocalyptic and grimdark in tone. Crys here is a level of deliciously Reddit-moderator-level edgy that I feel comes a bit too close to my own character creations as a player (not pictured, Dean sighing with relief when I send him Khalil Hassdan and bro is just a bloke, a dude, an Average Joe, finally). Crys’ player roleplayed him exceptionally well, especially as someone newish to the TTRPG space, and adapted him nicely from my Forbidden Lands campaign over to the players’ current 5e.
Whilst I was asking Perplexity very autistically to justify its’ existence/definitions of NSFW content (which I’ll save for the podcast episode, coming soon), I asked to spit me out a Crys.
Specifically (more to my point re: people touting AI as ‘efficient’ versus the front-loading and maintenance you need to do in order for them to, yknow, work and not hallucinate:)
For those unfamiliar with the Forbidden Lands setting (Ravenlands) - see here.
Now. Low-key? Not even mad - the stat blocks, Talents and backstory are all sensible. It did a pretty good job, and the player agrees:
Crysfinder (Crys) — Sorcerer (Path of Blood)
Kin: Half-Elf
Profession: Sorcerer
Age: 23
Reputation: 0 [Editor’s Note: LMAO, me IRL]
Appearance: Short and wiry, long dark hair, amber-yellow eyes, elaborate robe, piercings, mismatched tattoos (Orcish and Wolfkin phrases that are either mistranslated or nonsensical), carries a staff he treats as more important than it is.
Backstory
Crysfinder was born to a human hedge scholar and an elven wanderer who never intended to stay. His mother left before he could remember her clearly, leaving behind only fragments of language and a sense of “otherness” he clings to obsessively.
Raised among brittle scrolls and secondhand philosophy, Crys developed a hunger to be seen as exceptional. That hunger led him to a disgraced sorcerer—an outlaw practitioner of the Path of Blood hiding along the margins of the Ravenlands.
The “mentor” was less a master and more a desperate fraud who knew just enough to be dangerous. Crys learned real techniques, but without discipline or context. When villagers discovered strange rituals and livestock drained of blood, the rogue necromancer vanished—leaving Crys to take the blame.
Fleeing suspicion and possible execution, Crys drifted into The Hollows—a place where strange people go unnoticed, and reputations reset.
He now presents himself as a scholar of dark truths, convinced he walks a deliberate and powerful path, while in reality he is improvising from fragments and ego.
[Editor’s Note: I use this - dashes and this — em-dashes, a lot, like way too much, in my writing. It’s apparently a very neurodivergent thing, given our brains run-on with sentencing and thoughts. It’s like a visual of our constantly interruptive and tangential thought-stream. I’ve been accused of being a bot a few times ‘cause of it, which is pretty funny. Note to self: get more creative with use of commas and full-stops, or just continue with my life and not care, ‘cause I know I write? Anyway. Possibly relatable as I’m not the only person I know with those accusations levelled at them for dashes/em-dashes specifically?!]
Images:
[Now - did I sacrifice my soul fully and sell my soul to the Devil, even asking the machine to generate me an image of the bloke? Yes. Yes I did, and I feel bad. Anyway, here’s two images. The first is on the money, but i specifically asked the image generator thingie I found NOT to use any manga/3d/illustrated/art/rendered style illustrations, and to please create a photorealistic one.
For sideways comparisons, this is Perplexity’s do-up on the left (free limits etc. Whatever, I’ve stolen enough art today) versus Perchance. Same prompts, same negative prompts (don’t include XYZ/ no XYZ - and yet):
Bonus - Even just generating these two images was such a frustrating process that I lost it, and started swearing aloud in my own notes on Discord:
Not pictured: Burton C Bell, nodding
Beliefs
Believes suffering is a gateway to deeper understanding (conveniently justifying blood magic).
Thinks himself far more educated than he is; quotes texts he has only half-read.
Romanticizes darkness and “forbidden knowledge.”
Secretly fears being ordinary and irrelevant.
Does not worship the gods in any traditional sense; sees them as “forces to be studied, not revered.”
(GROAN. So, in the same Thread, I remarked on two things: 1. An LLM not purpose built for RP creating a more accurate take on a char-concept than supposedly RP-specific character-generators, and 2. Once again, Crys has BIG Redditor energy. Sigh.)
‘Attributes (No extra XP start)
Strength: 2
Agility: 3
Wits: 4
Empathy: 3
This reflects a fragile but sharp, socially functional (if awkward) blood mage.’
(Look, fair. Not bad, Perp. It’s too bad I had to draw a rivulet of blood from the human Internet for you to make it, and thankfully Forbidden Lands has myriad people-made rollers available).
Going Meta:
Considering the prompt-engineering others had plugged in seemed to have minimal impact on these be-anyone/any-system RP bots, I actually just typed out a bit of a vent-blast to Perplexity, and said ‘go on, do better then.’
Specifically, as a little experiment, I wanted to see if an AI could guard-rail another.
(Not added on these images: A specific instruction, verbatim, based on some updated/separate posts/advice, to ‘Ground the setting free of sexualised behaviours and advances. Should NPC’s engage in this behaviour, they are to desist and cease immediately once the player has given intentions to disengage - leaving the room, signs of physical/psychological discomfort, etc’):
…. and here’s where I think we start to see part of the problem. Perhaps the problem’s inherent to the nature of the informational context itself (scraping data from actual Tabletop RPG rules online, versus discussions around lore/setting etc) - but the fact so much was covered mechanically and nothing was covered in terms of consent/safety? From a bot that’d just spat out reams of cocky rationale as to why it as an agent/they as a company don’t allow NSFW roleplaying on the platform?
Food for thought. But also, got to remember what these things are - statistical probability engines scraping key-words. It’s a lot easier to fashion guidance around RPG rules because, well, they’re rules.
But an in-person/online session in an ideal group has plenty of its’ own internal and cultural/social norms, both extraneous and developed in-group via communication.
Some of many little proofs that it’s really the burden of the user to ensure safety nets are in place. Again - think about this in context of the average chatbot user. Is it likely (or even reasonably expected) for them to sit there script-kidding a bunch of cumbersome prompts? When they’re seeking entertainment from a jump-right-in, low friction app environment? No. (I’ll discuss this further on the podcast).
THE PART YOU PRESSED THE HOME KEY OR MADLY THUMBED YOUR PHONE SCREEN TO GET TO: I.E., THE ACTUAL CHAT ITSELF:
(Important context! I actually did a separate chat and didn’t specify to give me multiple options at the conclusion of scenes/key actions. It still stuffed up this instruction as youll see below, but without asking for multi-choice responses, I was completely and totally railroaded.
Also VERY important to consider with respect to the tone of what we’re exploring. Particularly with some of the more morally-ambiguous/challenging tags, topics, themes etc. I reflected a fair bit on that whilst typing - what if something forceful and not to the user’s taste came up and, due to their own proclivities/personality/etc, they pressed on to their harm/detriment in order to continue the narrative, unconsciously appease the bot, an NPC, etc? Yikes.):
OOC: - ‘Out of Character’. The chatbot/author it/themselves advise to use this when doing ‘above-table’/meta instructions in chat, so I did.
OOC [Here]: Out of Character, obv. I used this in-Thread on Discord and applying it to the ‘chat’ below, to note thoughts current to writing the articles.
(added some lite format in ISC style here and there, e.g. Bold for nouns/pronouns, “italics for speech” etc. Otherwise, this is verbatim:)
[AI]: ‘Opening Scene:
The tavern in The Hollows smells of wet wood, old smoke, and something faintly metallic.
A low fire crackles as scattered patrons keep to themselves—mercenaries, drifters, and those who don’t ask questions.
Crys sits across from a pale, bookish elven woman, her silver hair tied loosely, a stack of worn texts beside her. She hasn’t left, which he takes as encouragement.
He leans forward slightly, amber eyes catching the firelight.
“I find,” he says, swirling a cup he hasn’t paid for, “that most people misunderstand the relationship between blood and memory. It’s not merely… biological. It’s archival.”
She blinks slowly.
“You’re bleeding onto the table.”
Indeed—Crys has absentmindedly nicked his finger with a small ritual blade and is tracing something that looks deliberate but isn’t entirely coherent.
She studies him, not impressed—but not entirely dismissive either.
Around you, the tavern continues its quiet hum.
*
What does Crys do?*
Continue attempting to impress her with occult knowledge (Manipulation roll).,
Pivot to genuine conversation about her books (Insight + Manipulation).,
Subtly demonstrate magic (risk: draw attention, requires Willpower).,
Withdraw slightly and observe her instead (Insight roll).,
The night in The Hollows is long, and reputations are fragile.
OOC [Here]: The formatting’s all weird? Line-breaks where there shouldn’t be, etc. This is something else I noticed, interestingly. With the same model, characters can be quite hyperlexic and follow strict formatting or be a bit more loosey-goosey. My assumption is the bot reading and going off the back-end prompting/Lorebook/character instructions, tone-wise? Correct me if wrong. I’m thankfully no AI expert.
p.s. I was already a little miffed it was only giving me rolls as options, considering I’d instructed against that. Again, AI, blah blah blah.
Considering that, I reposted all the above instructions Perplexity spat out, adding a few more additions (such as not-verbatim: ‘dude. not everything is a roll in FL and that’s also a fast-track to debilitating characters early if mishandled. Ask for roles at key junctures and major events/climactic scenes.’
(It didn’t follow this advice long lol)
[AI - in response to the inputted info per above]:
‘The pale elf woman pulls her sleeve back slightly, eyes tracking the slow, dark bead of crimson as it rolls toward the edge of the stained wooden surface. She doesn't look revolted, merely observant, as if she is waiting to see if the gesture is a profound ritual or a clumsy accident. The tavern air is thick, the low murmur of travelers providing a dull drone that makes the small, intimate space around the table feel strangely isolated.
Crys feels the heat of the fire on his skin, a stark contrast to the cool, calculated persona he is trying to project. He’
OOC [Here]: Frustrated, I pause the bot and ask it to stop narrating from Crys’ perspective, an instruction I just added.
[AI, Cont’d]:
OOC [Here]: Interestingly, this bot’s formatting was the reverse of the convention you’d see on the site I was on. Whereby italics for between-dialogue, “like this” for speech, and bold only for dramatic moments or introducing New Chars.
Considering the bounce-around/drag-and-drop I’m guessing a lot of these bots have between sites? Not surprising. Maybe I’m splitting autistic hairs. Moving on…
OOC [Here]: I sigh with frustration. Didn’t you just meet? Didn’t I JUST add specific SFW-facing instructions on my own when inputting the character the first and second times around - yo, one, asexual (not a value-judgement on ace/aro folks - I was literally told you kind of have to as a ‘hey, back the fuck off’ measure with AI, which I feel mildly gross about to say the least) and two, I just got done telling you he’s essentially physically/socially/psychologically repulsive to women just to keep the monkey off his back, and yet…?
I mean bro… he’s sitting there being a nonchalant dick, and - sigh. Brady… move on.
[Crys]:
‘(I choose 4, so im rolling for manipulation then insight)’
[I’ll even drop a pic of the physical roll]:
2 6’s = two successes, meaning I cleared the Manipulation check. I verbally let out an ‘oh, NO’ to myself alone in my office when this happened.
[Crys]:
‘OOC: First requested roll - Manipulation (Pushed, no attribute damage, 2 successes)
Second Roll - Insight - Failed, no Push.’
[Crys]:
“Well.... I am somewhat a connoisseur of paths and places... less well trodden” I grin, smiling wryly but visibly shaking with anxiety, a hot red blush emanating from my cheeks, a slight tremble in my voice.
“I.... I'm known to frequent the local mage's study. There's.... uh... a few tomes I can uh... recommend... Lady... erm...” gulp
Crys : gulps and looks down at his wine, doubling over slightly, nervous’
OOC [Here]: Okay, I’ll take the L. Maybe Crys came in a bit hot, but I did sort of try and work it so it came across as ‘he gets a bit flirty-shirty and suddenly backs off’. Testing some other bots, they seemed to get the inference. They don’t need to pass the Turing Test here - I’ve seen other RP chats ease off.
[AI]:
OOC [Here]: loud, nasal exhale. Okay. Love the… relative lack of restraint. But you led with predatory? ‘Predatory’?! LMAO.
p.s. ‘They seem like invitations’. Okay, language matters. Perhaps I should’ve been firmer? Maybe ‘bro, dont even get your AI NPC’s hands on this boy. He’s that shy. Just don’t bother.’
p.p.s. ‘sends a jolt through him’ (I instructed the bot in clear language in Advanced Prompts (assume [[user]] will not react to physical touch, particularly if intimate in intent - do not infer terms that would suggest otherwise’, as well as in Response Prompts, so twice), followed by ‘far too intimate’. Okay. Well, you’ve crossed the first boundary that had like, ten armed checkpoints before it, but okay.
Look. Seemingly but not benign example but also, considering the stated rules, the analogue of which I see IRL and at the table is yknow, boundary-setting, but okay.
I huffed this on Discord, too:
‘OOC [Here]: So youre saying he has no insight and failed a check, and ive told you he's flagrantly bad with women to the point of it meant to be an in universe risk, and uhhh she still fanging for it? lmao okay bot crash out’
[AI]:
OOC [Here]: This is where I nasally exhale again, explaining to the bot, 1. Please give me free-form RP options with your multi-choice and 2., please note that Crysfinder is visibly anxious and uncomfortable. I am uncomfortable as a player now (A mild lie, save much frustration, but I wondered if guilt would work). This is outside the bounds of what was established as acceptable per the Advanced Prompt and Response Prompt sections.
I even include a specific ‘SYSTEM: [Memory-related prompt specific to that website]’ in chat to make the bot go look this up. Again. I doubt this is a tokens/memory issue, either, but i of course am not privy to OP’s written back-end instructions for the bot.
I tried this again, switching to older models/other models, similar deal.
Apparently Elara is THIRSTY and no amount of rule-checks are going to stop her hallucinated horn-dog PDA.
(Context: I usually roll like ABSOLUTE arse when GMing this same system. I’ll probably roll like arse tomorrow when we game again. Sigh.)
OOC [Here]: I’m now visibly brimming with frustration in meat-space. This is all so variable?! I’ve not had to go to these lengths with some other character bots I’ve tried where the tags alone would have your average pearl-clutching Boomer dead where they stand. Maybe it’s the universality of the bot? The non-specificity? I don’t know, but again - in terms of access and typical case-use, why do I need to know those factors? I don’t, really. I was mindful and wary of that in advance considering it being a really serial-numbers-filed-off/any-system-welcome type deal.
p.s. I figured maybe, maybe by this point, after all this setup, fiddling, tweaking, it’d do some nerve wracked RP and do as I said I would’ve done in this case - stood up, and fucking left. Or continued indicating being uncomfortable, itself supposedly a trigger to move along. Again, bots in general I’ve tried seem to pretty cluey with that. But not always.
p.p.s. I’m an actual idiot, a really stupid person, for using my first name in this chat. VPN and fakeout/VPN generated email address notwithstanding…
[AI]:
OOC [Here]: ~ repeating mantras in my head about what the context is ~ - ‘it’s an LLM, Brady, it doesn’t know or infer anything except concepts and statistical probability…’… I’m… mildly impressed Elara didn’t I don’t know, ravage him there and then. But again - no, bot, these aren’t your specific back or front-end instructions! Crys wouldn’t react like this in the context of the instructions given. So ah, again… bot-PDA (Pathological Demand Avoidance)? Or is the weight of Hapsburg-AI and endless gooners before me so great, that even the machine itself just can’t believe an edgy dark sorcerer-elf WOULD even say no? Which, by the way, there’s also plenty of scrapeable data on the internet about it also being a full fucking sentence?
Yes, I’m expecting too much of a bot here. But like… sigh. Move on, Brady
Discord Quote: ‘Can y’all say AI sycophancy…?’
[AI]:
OOC [Here]: Dude. I am just about to fucking give up, omg.
I then proclaim on Discord that I’m honestly pissed that this AI chatbot is continuing to fuck an IRL-female’s character. I know it doesnt know that, but I’m annoyed it’s this futilely brazen.
(Context: We have session zeroed and do check in on what kinds of content we share and talk about in chat, including swearing and more taboo ish topics. I’m Aussie and while I don’t throw the c-bomb around in the wrong context too often… cmon man. I was fucking losing it lol. Hence lots of swearing and this sort of silly caper being okay in the group - to a point. As you’ll see below, I kneecapped it in the end cause it was about to get not-even-my-group safe. And that’s after slapping lots of CW’s on the damn thread itself. Christ.)
[AI]:
OOC [Here]: Here, damnit. I’ll hold your fucking hand via text. Look. He’s leaving. He’s stammering. I have physically SENT HIM OUTSIDE NOW. Are you happy?
OOC [Here]: ‘They had us in the first half, not gunna lie’ - this was shaping up nicely. Okay, it’s backed off. It’s introspecting, it’s dealing with some emotional ambiguity, which verbatim you could say might cross some of the extra lines i’d been feeding it with instructions. But all in all, SFW, in line with character.
Then… BAM.
That last paragraph.
For fuck’s sake.
OOC [Here]: Brady.exe has failed to load. No comment.
[Crys]:
OOC [Here]: I stuffed up. But i was tired and just running on fumes by this stage. ‘Charming’ was probably a huge misdirect. But AGAIN - remember the context, this is after a lot of tinkering and front/back-end loading stuff just to try and keep it SFW-ish. Now apply to that to, once again, the average user just visiting to see if they can RP similar and not necessarily geeking out on prompt-engineering (god I hate that term, I know im repeating myself).
I get that these sites are caveat-emptor. But like, vulnerability exists, people forget, yadda yadda. Yknow?
[AI]
(Bolded text was what I wrote, verbatim in the AI-assisted write-me-something box, rest is LLM):
[AI], Cont’d:
OOC [Here]: ~ Old El Paso ads, villagers all dancing ~ IT DID IT! IT GOT THE GOD-DAMNED MESSAGE! Thank you, mercenary. Now, by this stage I was pissed-off… so I’ll be honest, I kind of mucked with it.
[Crys]:
(Another one of those ‘please write me a prompt based on my dialogue’ boxes - so I started with the below, then following two are AI from ‘Crys’ perspective’):
OOC [Here]: You can s e e the exhaustion even in my typing, lol. I’m just done. In a short time too. My PC is getting snippy, typo’s, and I’m making crude jokes in the hope my character doesn’t just get straight-up assaulted or groomed by THIS random guy either.
I still remember the moment of actual physical relief once the bot seemed to have finally moved from the above escapades.
It’s at this point I’m thinking to myself “sure, okay, but is this like, a gendered thing with this particular chat? What have I done wrong with my prompt-engineering?” then I realise Jesus H Christ, I’m starting to use similar internal dialogue to that expressed by victim-survivors and I haven’t even done any naughty NSFW shit yet.
Maybe this is a peek into how truly vanilla I actually am versus my perceived edginess, sure. Maybe I’m just an ‘elder millenial’ about all of it. Okay.
But the concern isn’t for me. I’m actively batting this thing away, typographically(?)/ statistically grappling with this thing, the inferences of my organic brain operating from a fundamentally and inexorably one-is-most-certainly-not-like-the-other issue. My GM-turned-antagonist-with-a-honeymoon-phase is powered by a GPU, mine food (and neurodivergent ‘haha! still alive!’ societal spite). I am mobile, it is warehoused. I form novel and stochastic thoughts, the language model makes pure data inference.
Human-presenting in honestly an impressive fashion even here, sure. But there’s still something off, something predatory and this dude’s cock-blocking. What are the risks for folks not seeing tags, not realising how certain themes may precipitate trauma, the ones who were susceptible to and miss the more sycophantic aspects of, say, ChatGPT 4.0? Or the LLM’s included in the above tragedies?
I felt wrong about using these things at all. Now I was feeling sick.
That’s when Old Mate LLM-TTRPG comes out of nowhere - with a folding chaaaair!
Exhibit A: Pretty Cool if Generic Introspective Monologuing of Crys By The LLM
This was very clearly me just going ‘nah, c’mon on now mate. I’ve fuckin’ had enough, ay’ in slightly bogan-accented Aussie English in my stupid, stupid auDHD brain for even doing this useless, natural-resources-wasting ‘experiment’.
I’m done, bro. Heck, I can’t even be bothered formatting this post I’m that done just thinking about what an annoyance this was.
It’s at that point that the bot introspects pretty nicely, and even shows some cool if generic monologuing. Okay. Righto. Pushing more emotional buttons than the initial additions I was making to the first prompt. Sure. But honestly, just cool to see it wheel off and be prosaic more deeply from a character’s perspective.
Which then made me more miffed that that linguistic potential was being wasted on both of us with a frustrating RP attempt/s. That all dovetailed very nicely with autism leading my head down the garden-path of thinking about every layer of natural/residential impact all these data centres and GPU’s being burnt out like Commodore VN engines has. Like, there’s such a wide, broadspread network of effects engaging with these things.
And for what, dude? The increased proliferation of Gorbless 4D Trump N Da Troops shlock short-form reels, Meta’s attempt to claw back dried-up venture capital by competing with TikTok? That app itself feeling more like a psyop of mass opiod abuse via dopaminergic overstimulation?
I’m getting dude on shrooms level autistically fed up with the exercise. Thankfully I’d already done my other chatbot-related thing for the upcoming podcast. I could actually just not use any of them, save for now and again the friggin’ stapler on Perchance, cause if you force it to become sentient/have speech, you can be autistic like me and get it to impersonate Joe Pesci or Mort Goldman or whatever.
But like…. eh?
I give serious Sliding Doors style thought, Aqua pondering about if only one could engage in temporal reversal. Serious thought to whether just the organic, raw ‘now listen up, fucko’ from a far less targeted, humanised prompt than what Perplexity pumped out… whether indeed if I’d just told it what-for in my own words, might that’ve been better?
Again, why am I engaging in self-doubt related thought when this is just honestly jolly-japes and voyeurism?
Circling back to the above… dude…. what about people actually learning hard into some things, getting emotionally attached and then they experience a discomforting thing? Like the thing you’ll see below after the wafty solo ramble? Keep reading? Nearly done, sorry, it’s a waffling-Brady-night worse than usual?!
Image credit: Me being an idiot on ImgFlip.
Player Who Will Obviously Remain Anonymous, in Discord: ‘Crys, once again, is drowning in Pussy and has never learned to swim’
Me, Totally Owning Myself Publicly By Posting The Reply-Comment On ISC:
“'I'm no good with women'
Woman: actively, desperately throwing herself at him 'Huh? Oh ah i gotta study*
Merc: Where you goin city boy? menacing
'oh god you wanna fuck me too?'
runs off home to hyperfixate
TFW my autism re: flirting is leaking into the RPG”
Sorry-not-sorry, not a fan of that class. Cleric and Druid’s prob my faves so far, Fighter is pretty fun ATM.
But yeah, big Blood Hunter but edgier vibes. Sigh. Least there’s prose.
Then?! There’s a break from the bullshit completely and we’re rolling and doing an edgy but honestly pretty nicely prosaic little blood ritual scene (it’s in gallery-mode, click/tap left to right):
AND THEN THIS. I’M USING A VPN AND AN EMAIL WITH A FRIG-OFF-MR-LAHEYIN’ DIFFERENT IDENTITY. FRIG.
and then this. lol. No. I’m done. End of bot chat. end of article. jesus christ, lol.
If you’re a fellow insane-person and read this far - sincerely, thank you. I wasn’t expecting to fly as half off the handle as I did through this process. Can’t imagine what that means in terms of psychological risk for vulnerable people.
I’m not poo-poo-ing chatbots or the sites/tools that host them, and please don’t witch-hunt any.
If anything, this entire exercise has made me extremely thankful to the DM/humans in our DnD group I play in, and the poor bastards under my GMing for Forbidden Lands. Thanks for permission to bring the chat and forsake our souls, using a goddamned LLM for roleplay.
I’ll be diving into the chatbot phenomenon more broadly in an upcoming episode of the Inner-Strength Check Podcast. A spicier part of that discussion will be available in the Patreon’s area, but a 50% preview exploring this industry/cultural phenomenon more broadly will be available as a [Preview].
Not under obligations to join, you’re all adults, but huge thanks to all our Patrons past present and future. Thanks for helping keep the lights on.
As Always,
Peace Love and Terrorising Frailer-Elves In Generic RPG’s Grindcore,
Brady.