AI Disclosure
Bottom Line Up Front
Currently, we only use AI voice-changing tools to create distinct fictional voices from human speech performances, in order to represent our characters in social media posts. That's it. Period.
At Loot Cave Toys, we take pride in providing high-quality, ethically-produced products, and this extends beyond those products to our marketing, design, and miscellaneous activities. We generally do not utilize Generative AI technologies, except in certain specific circumstances where we are confident in establishing a climate of ethical use. We will never use Generative AI technologies simply because they are convenient or save us time. We only use Generative AI technologies to deliver capabilities that cannot be achieved without it. We always perform a human review of any AI-generated content before it ever leaves the shop.
In rare circumstances where we use AI technologies, we have thoroughly reviewed the tools and systems we use to ensure that datasets were ethically sourced, taken steps to maximize inclusivity and mitigate bias, and will always perform a human review of generated content prior to release. We will always disclose this content within and/or alongside such media, and will always provide a detailed explanation of our use cases for any Generative AI use on this page.
Quick Note: Why do you call it "Generative AI"?
We use the term Generative AI here because we felt it would be disingenuous to use a less well-known term like "Transformative AI", however our current use does transform human input like-to-like (input audio sample is processed through AI voice changer to receive output audio sample). When we use the term Generative AI, we mean it in the broadest sense; a system which generates new data from a set of inputs. We don't think terms like "Transformative AI" add meaningful clarity; we would prefer to let our practices speak for themselves than using terminology that isn't industry-standard or well understood.
Since we're confident that our practices can satisfy the highest level of scrutiny, we don't think it would be appropriate to deceptively modify how we refer to the tools we use. We would rather bear the burden of some customers misunderstanding our use, as opposed to manipulating the discourse to make the optics more favorable to us; put differently, if we're not confident with our business practices in broad daylight, we should not be confident with them at all.
Generative AI tools we DO NOT use:
- Any form of Text Generative AI (e.g. ChatGPT, LLaMA, etc.)
- Any form of Image Generative AI (e.g. DALL-E, Stable Diffusion, etc.)
- Realtime Voice Cloning (RVC) tools, especially where voices are synthesized from mixed datasets or sourced from parties who may not have been compensated or informed, or who may not have provided consent to the creation of a model of their voice.
- Text-to-Speech (TTS) tools, whether AI-powered or not
- Anything else not mentioned above or in the next section.
Generative AI tools we may use in specific circumstances:
- Non-RVC AI-powered speech transformation ("voice changer") tools trained on data from paid, informed, consenting voice actors (e.g. Voicemod AI-powered voices)
The Long Version
You likely ended up here because you saw the following line in one of our social media posts:
Video uses AI-powered voice changer software. Read more: lct.gg/AI
If you did, we commend you for taking the initiative to understand our use of Artificial Intelligence tools, specifically Generative AI. We believe that ethical use and consumption of AI-generated content starts with extensive, voluntary disclosure. We're not afraid to eat our own cooking in that regard, and we will always provide detailed information on our use of any AI technologies.
However, our decision to use an AI tool does not start or end with disclosure. We thoroughly evaluate any and all technologies we are considering long before we would ever use them in our work. This starts with an in-depth technical analysis of the tools; we will never use a tool we don't understand. Once we have a technical understanding, we investigate dataset sourcing and quality; as a baseline we require that any model we use must be trained on data which contains no harmful or illegal content, does not represent disproportionate dataset bias or skew, and was sourced from individuals who were compensated for their work, informed of its use in AI, and who consented to such use. Finally, we perform an analysis of necessary safeguards for our use, including mechanisms to mitigate algorithmic bias, controls for safety and legality of generated content, and tight scoping of potential use cases.
So far, we've gotten to the point where we identify a use case and have assessed an AI-powered option for use. Let's assume we've decided that a particular tool meets our bar, and we're ready to start experimenting. Before we dig in, the next part of the process is all about due diligence. We fact-check all of our research and assumptions, codify any constraints or mechanisms for safe & ethical use into our Standard Operating Procedures, and document the use case on this page. It is only after all of this is complete that we even begin to experiment hands-on with the AI tool (except for any analysis necessary to perform the above research and detail).
Now that everything is written out and ready to proceed, we'll start experimenting to validate our findings, tweak our use case to any new behaviors we encounter, and prepare samples to benchmark. That means you might see a use case appear on this page before we've even confirmed that the outputs even meet our business requirements; we're so serious about disclosure that we want you to know if we even considered using an AI-powered tool. This experimentation phase also helps us fine-tune our human review process, identify additional edge cases where we can improve our mechanisms or constraints, and validate business utility.
Finally, if we're still happy with things, we start moving towards use of an AI-powered tool for a particular use case. First, we will prepare a human-written, human-performed input to the AI; for example, in the case of our character voices this includes preparing a script, and then performing a reading of that script. Then, after the content is generated, we perform a human review to quadruple-check the output against the constraints and requirements, as well as to determine business fit. Then, when we use any generated content, we will assess the available disclosure mechanisms for all platforms where the content will appear and identify the maximum-impact, maximum-visibility avenues for disclosure. We aim, where possible, to disclose AI use both inside the content (e.g. disclaimer slide in a video) and alongside the content (e.g. accompanying text, declarative mechanisms). We will always disclose AI use in one of these two ways, if both options cannot be executed simultaneously.
Use Case Breakdown
For any areas where we utilize AI technologies, we have prepared information on how the technology is used, what research we conducted prior to using it, and how we ensure safe and ethical use.
AI-Powered Speech Synthesis
We utilize Voicemod voice-changing software to create some of the synthetic voices which represent our characters in our social media posts. In all cases, we use voice-changing effects to modulate a human performance (not-TTS). Some of these voices utilize traditional digital signal processing techniques (no AI), while others use custom in-house AI voice models developed by Voicemod. Voicemod has confirmed that they train their models with data from paid voice actors using a synthetic process which does not result in an identical clone of the actor's voice.
In our decision to utilize these voices, we established the following constraints for our use:
- All voices start with a human-made recording. The Voicemod AI-powered voices modulate speech as an input, and preserve performance characteristics of the input voice samples. We do not, and will not ever, use Text-to-Speech because it does not meet our bar for ethical use and generally produces poor quality output compared to human performances.
- We only use voices available specifically through the Voicemod application, and do not rely on "Community voices".
- We use these voices to generate specific, obviously fictitious voices for our characters only. We never use these voices to portray any real individual, or to portray any fictitious character other than our own characters.
- We always disclose the use of AI-powered voices through declarative mechanisms (e.g. TikTok's AI-Generated Content Label) where available, alongside the content on all platforms (e.g. in accompanying text), and inside the content (e.g. AI content disclosure slide).
- When selecting voices, we do not include perceived race, national origin, group or cultural stereotype, or vocal trait attributed to a medical condition or disability as factors in our selection criteria. We also do not utilize voices we believe unfairly or irresponsibly represent caricatures of these factors. We will never disqualify a voice because of these criteria if we believe it is an authentic, respectful use. Our characters represent a diverse (even ignoring the alien part) group across a spectrum of genders and other personality traits; we do target for fit to these characteristics when selecting voices. However, our controls in this regard generally rely on non-AI portions of the voice generation, such as pitch-bend, modulating filters, or audio effects.
- We comply fully with the Voicemod Terms of Use, which permits commercial use of generated audio.
Factors we considered when deciding to utilize these voices include:
- Models are "first-party", created in-house by Voicemod
- Voicemod AI "Personas" are not directly cloned from an individual's voice
- Models are trained on voice samples from paid professional voice actors
- Voicemod's ToS permits commercial use and does not prohibit our use case.
- Usage in this manner complies with applicable ToS/AUP/similar documents for all platforms where content is delivered.
Last Updated: 2024-03-22:
- Updated disclaimer text to match what will be used on posts: "Video uses AI-powered voice changer software. Read more: lct.gg/AI"
- Small edits throughout for clarity and ease of understanding.
Previous edits:
2024-03-20:
- Added clarification about our use of the term "Generative AI" to refer to our current use case.
- Updated our "do not use" section and voice changer usage constraints section to indicate that we do not use Text-to-Speech (TTS)
- Updated our description of model selection criteria to indicate that we utilize diverse voices but do not specifically select voices based on such criteria.
This page, like all pages on our website, was entirely written by a human. If you want to get in touch with a human about any questions or concerns you may have, you can reach us at:
Space Magic Manufacturing, LLC
221 Palmer Country Ln
Palmyra, VA 22963
sales@lootcave.toys