The Ubergroup

To “search” means to use Google or any comparable tool regardless of brand to find information to study with your own human brain. For example, typing "which libraries contain the most useful manuscript and map collections for 10th century medieval history" into a Google search bar produces the following results:

Even though this is technically provided by Google Gemini, reading such a summary does not constitute “using generative AI.” Artists are further permitted to refine their searches with natural-speech prompts, such as follows:

The refined search provides results such as follows:

Although this is using an AI interface, it is not considered AI generation. The difference between searching and generation is whether the artist is seeking information to process with their own human brain. Other examples of search prompts include:

• “How do the emergency brakes on a tram work?”
• “Please give me maps of 19th century Charleston.”
• “Besides the stage and the lobby, what other rooms exist in a theatre, and what floors would they normally be on? Please show me a typical floorplan of a mid-sized theatre for about 500 people.”
• “Most popular baby girl names in 1980s Japan”
• “What are the most common occupations given for characters in modern thriller films?”

The artist must verify the quality of their search results and make all final executive decisions about how to create the work themselves. The online search must mimic work they could have done in an analogue fashion, such as visiting a tram museum and asking the question of a tour guide, visiting a historical society and referencing maps in person, or opening a book of baby names. The results may not be copied or passed off as one’s own, eg, the “typical floorplan of a mid-sized theatre” may not be published at the front of a novel or shown in a video, only studied by the artist for their own comprehension. Extremely short, literal, generic phrases which do not violate copyright may be quoted with only light modifications. For example: “the magnets under the tram car clamp onto the track” would be acceptably different for an educational youtube script or a scene in a novel, modified from the following answer found on Wikipedia: “Heavy electromagnets suspended on the bogie (underframe) just above the steel rails drop down and clamp directly onto the track via magnetic force.”

When search results lead to a licensable product, such as music or images, the artist may elect to license it for use in the standard legally compliant manner for their own final product.

The artist is responsible for fact-checking their results. Google and other AI-assisted search engines are notoriously inaccurate. If factual accuracy matters to the final product in any way, the artist must treat Google results with the same skepticism that a historian or scientific researcher would treat a crowdsourced Wikipedia summary, or a late-night comedian going around with a microphone asking Americans about geography. Most information on the internet is written by a mixture of fallible humans with no screening whatsoever and Russian disinformation bots.

To analyze means the artist writes, draws, composes, designs, or crafts the art themselves, then uses tech-assisted tools like spellcheck, find and replace, pattern recognition, auto-correct, or software like Photoshop, Premiere, Reaper, Wilcom Embroidery Studio, AutoCAD, etc. as a tool to modify their own original source material (or legally licensed images, music, etc) according to the parameters decided in their own human brain.

AI features in software may not be used to generate anything from unspecified sources, e.g. to generate an entire background on a video, due to the risk of plagiarism. All source material must be original or properly licensed.

For example, the pre-AI “find and replace” function was able to quickly change all versions of a character name. An artist working with an entire TV season may have a to-do list of several characters to cut, change names for, or otherwise streamline across multiple episodes.

Typing “Please help me cross-check all episodes against my intended notes for continuity and create a list with headings separating the to-fix issues by episode and page number, and making a note when a change in one place will also need matching changes made in other episodes, because I like to flip back and forth between all relevant scenes for continuity, but struggle to hold the big picture of where to find all relevant scenes in my head simultaneously alongside the more in-detail focus of actually editing a scene” into Claude provides the following checklist:

This constitutes “analyzing” and not “generating” because the Large Language Model is being used as an improved find-and-replace tool, to provide a checklist of all instances of a character name for the artist to do with what they will. All source material belongs exclusively to the human artist, all editorial parameters were decided upon by human artists, and all actual changes will be executed by the human artist.

For another example, a visual artist can check a piece against existing principles of graphic design by asking the AI to point out compositional impressions such as focal point, hierarchy, shapes and flow. This type of analysis fundamentally existed long before AI, such as placing a grid over an image to check how elements align with the rule of three, or using a swatch booklet to analyze colors in a piece.

This constitutes “analyzing” and not “generating” because the tool is comparing the original work to other works in a discussion format for the artist to think about, not changing their work to include elements copied from other works. It is the same fundamental difference between studying other paintings and forging them.

An artist may analyze comic page layouts in a similar manner, comparing their own work against the established principles of composition and flow.

What makes this “analysis” and not “generation” is that the original work was completely human made, and the software did not change the work in any way. After the analysis, the artist makes all final decisions about composition, color choices, etc, and alters their work (or not) by hand.

It is important to remember that even the analysis contains naturalistic language, eg “What I’d look at,” AI is not an art critic or a human teacher; it is a measuring device. It can compare your art with other art which exists. What you choose to do with that information requires your own human opinion. It can cite textbooks written by human art teachers, but every art form has multiple schools of thought and multiple famous teachers, so deciding who’s advice to follow is always an individual choice. If you are unsure what to do with this plurality of information, bring those concerns to human teachers and peers in the Ubergroup and discuss. Never blindly accept the first suggestion offered by AI analysis without a holistic understanding of all major schools of thought within your art form and a period of deep introspection regarding which one you’d like to follow.

Another example of analysis is compiling a list of publishers or producers who have a history of producing work similar to your own, to determine whose style you may be a fit for. An artist may upload a sample of their novel or graphic novel and ask a large language model to provide a list of submission opportunities for which it might be a fit.

The analog way to accomplish this is taught in one of our grad classes. In essence, one compiles a list of all publishers or producers to whom one might pitch, and reads or watches samples from the most recent dozen or so things they have put out, to get a sense if one’s own work fits their “house style”. It is highly recommended that any artist considering this use learn the analog method first to prevent cognitive atrophy, and continues to use the analog method at least once a year to keep themselves familiar with the current state of their industry. AI-assisted submission analysis is best treated as a supplementary element which leaves no stone unturned, in case the human artist accidentally overlooks something when searching the analog way.

To “generate” means to prompt any web-based, app-based, or software-based tool to create work which you then pass off as your own, or to modify your work in a way that uses components which are not your own (or legally licensed.)

All “generative” AI uses unconsenting human-made works as source material and is wildly destructive to human creativity. Famous examples include Anthropic destroying rare antique books they bought to train Claude, and visual artists having their signature style illegally copied and monetized. Adobe, Meta, and Zoom famously all changed their TOS to make users agree that whatever the software sees or hears is subject to enter their AI training databases. Users have to find a buried manual opt-out, if the software even allows opting out. Given the outsize share of the market each of these platforms cover, it is nearly impossible for an artist to avoid having their work copied.

Using an AI image generator to create a cover for your book, poster for your show, or or animation for your video is always ripping off an unconsenting visual artist who could not avoid using an industry standard Adobe suite product to meet final project requirements, or could not avoid publicizing their work on social media as the only viable route for small artists to gain visibility.

Using an AI voice to narrate your film, video, or audiobook is always ripping off an unconsenting professional actor who could not help but take a mandatory Zoom call at some point in their careers, or to be interviewed on camera in connection with publicity. Using AI generated “music” is ripping off an unconsenting professional musician in the same manner.

Passing off work as your own which was stolen, in whole or in part, from unconsenting artists is both illegal and immoral.

Generative AI companies argue that by training the model on copyrighted material, they are within fair use by “transforming” images into software data. They believe they are not legally required to pay artists or writers for using their work, or even required to seek permission to use their work.

Companies like Meta, Adobe, and other social media apps require users who post on their sites to agree, “by using the software,” to give the companies certain rights to distribute the user’s content. This was originally rolled out to protect the companies from lawsuits when they used software to generate thumbnails (or copies) of the work to display on the site. These agreements are shifting to include a blanket consent for the content to be used in training generative AI models. In some cases, artists can “opt-out” by navigating convoluted pages of settings and selecting the “opt-out” button, but the software companies are currently not being held accountable legally because of the fair use argument. And even when artists opt out, that does not erase any of their work that has already been ingested into training data.

Without any legal regulations in place, generated AI content cannot be trusted to ethically and fairly compensate artists for the use of their work. The Ubergroup disagrees with the ‘fair use’ argument. Any generative elements whatsoever disqualifies a work from being Certified Human Made.

The only instance where generative AI can arguably be legal is as a proof-of-concept or a placeholder for art that will be replaced once a human artist has finished the work, such as a writer using a watermarked stock image to explain their ideas to a human illustrator. The example still belongs to someone else and cannot be copyrighted as your own, but it is acceptable to point to it and discuss.

Generating a placeholder still requires interfacing with unethical software with significant environmental repercussions and is discouraged, but minimal use with complete transparency–you admit openly that the source does not belong to you, is only intended for discussion with human collaborators, and will be replaced–does not disqualify an artist from being Certified Human Made. Hiding use of AI in proof-of-concept is lying by omission, unfair to your potential collaborators, and will result in revoking of your certification, removal from the Ubergroup and loss of access to our network of human collaborators.

Most staff and alumni of the Ubergroup find generative AI loathsome even when used legally as proof of concept, and may be disinclined to work with you if you do it. It is strongly suggested you simply find unaltered human-produced images, voice samples, etc, to share in discussion, as to not burn bridges with potential collaborators.

To qualify as “Certified Human Made” by the Ubergroup, your project must have been created from start to finish within view of Ubergroup staff and alumni.

Personal Threads

Your personal thread on the Ubergroup server is the central aggregator of all evidence of your creation process. For ease of verification, please keep all conversation, brainstorming, edits, and iterations together on the Ubergroup server and keep at least two other human staff members or alumni who are willing to serve as a legal witness involved at all times - the more the merrier.

You are welcome to work with anyone you choose outside the Ubergroup, but please share regular status updates of work done with your collaborators inside the Ubergroup. Extensive time-stamped receipts and willing legal witnesses are your only protection against false accusations. We cannot testify to a creation process that we did not witness.

You are encouraged to export the contents of your personal thread regularly to create backup files in case of server errors, and to exchange multiple forms of contact information with your witnesses. The Ubergroup does not own the servers we operate on. We provide guidance and a network of human witnesses, but cannot be responsible for server errors beyond our control.

Legal Witnesses

All Ubergroup staff and alumni may elect to act as legal witnesses for each other. Please discuss your use of “search” and “analysis” functions regularly and openly with examples in your personal thread, so all witnesses can feel completely comfortable and in agreement that your use of software or apps which have unavoidable AI components falls under “search” and “analysis,” not “generation.” Alumni are welcome to have differing personal opinions on how to interpret an individual use case and may decline to serve as witnesses if they feel the interpretation is too liberal. You must have at least two human beings including at least one staff member willing to testify to the authenticity of your artwork and transparency of your creative process in order to earn the Certified Human Made seal.