An Ethical AI Model Must Enable Fair Competition
Intro
In the context of generative AI for creative use, it is essential that models respect copyright and are only trained on public domain, opt-in/licensed work, and Creative Commons (where terms are abided). However, consent alone is not the sole criterion for an ethical model β it must enable fair competition.
In the following, I will demonstrate the potential for an unrestricted model to disable fair competition, shed light on the economic implications if competition were undermined, and present a starting framework for an ethical model. Although this will specifically focus on visual art, the concepts are applicable to all forms of creative work and their respective AI systems.
A Harmful Network Effect
If not intentionally constrained, there is a significant risk of a network effect occurring that would undermine fair competition. It would look something like this (this would obviously coincide with the technological advancements described below):
The more artists opt-in -> the more powerful the model becomes -> the less individual artists can compete on their own -> a vicious cycle ensues whereupon artists become increasingly reliant on opting-in to secure income -> the cycle repeats until artists lose bargaining power and the ability to set their own prices.
Eventually the power dynamic between artists and the company would become inverted. This dynamic would also be reflected in the diminishing role of the artist over time. Initially, artists will use the AI as a workflow enhancement, but gradually, they will become increasingly relegated to tweaking the AIβs outputs. Eventually, at the furthest extreme, their involvement may be reduced to merely augmenting the dataset. The power of an unrestricted model is inversely related to the power of the artist (economic, bargaining).
Why Future Models Are Threatening To Competition
Itβs likely that in the not-too-distant future, models will:
β’ Be few-shot learners.
β’ Be able to accurately model long-term dependencies that can retain and utilize contextual information over extended sequences (consistent characters).
β’ Be able to transfer virtually any aspect of the input image onto the output using style transfer in img2img generation.
β’ Command a mastery over both technical and creative skill (style).
β’ Exist within user-friendly D2C UIs that allow for both autogeneration and finely detailed customization. The UIs would feature dropdown menu modifiers represented by thumbnails. For every modifier category, a preset could be chosen and the user could repeatedly autogenerate until a satisfactory result is reached. If this leaves more to be desired, parameters could be adjusted or prompting can be done. An example of a modifier category and the associated presets: composition -> rule of thirds, golden ratio, symmetry/asymmetry, leading lines, etc.
When the model achieves technical mastery, creatives will be forced into a corner and style will become their last line of defense. Yet, they will be disarmed if artists with similar styles opt-in. If that occurs, individual artists will be forced to compete against an optimized conglomerate of stylistically similar artists; by employing techniques like reinforcement learning and similarity metrics, the model could be trained to identify combinations of artists that align with a particular style. Added to that, models may offset stylistically generic outputs by:
β’ A randomness control parameter.
β’ A βblend styleβ parameter.
β’ Countless modifiers and parameters for fine-tuned customization.
β’ Being trained to account for novelty and variation.
There are also exacerbating factors:
β’ Future few-shot learners will only require a small sample of works to replicate style.
β’ If antitrust measures do not prevent a company from cornering the market, they may incentivize opt-ins for styles by offering exposure, bonuses, or paying a premium for the talent it scouts. In turn, this may incentivize people to use unethical models to replicate styles to license to the company.
Artists are therefore left in the desperate situation of banking on a thread of hope that artists with similar styles do not opt-in. Creatives are already outcompeted by the model in terms of cost, speed, and scale, and this alone will cause massive labor displacement β if similar artists enter the dataset, individual artists would likely not be able to compete at all.
When Artists Can No Longer Compete
When artists cannot compete, they obviously lose bargaining power and the ability to set their own prices. In the worst case scenario, creatives may become relegated to making βutilitarian artβ with the sole function of augmenting the dataset. Of course, there will always be fine art patrons, sustainability-orientated consumers, and art markets that feature only human work, but this will not be enough to sustain most artists. As of now, creative work is already alarmingly precarious.
It may also be worth noting that the climate of generative AI is particularly vulnerable to anticompetitive behavior, which may include: price-fixing, limit pricing, exclusive licensing, strategic partnerships, acquisition of competitors, training on unethically sourced material, training on endless derivatives from licensed works, etc. Nevertheless, even if these anticompetitive practices were curtailed, fair competition would still be undermined, albeit it would take longer. It is my belief that unfair competition is an inherent inevitability of unrestricted generative AI.
The Stages Of Economic Devaluation
Economic repercussions would coincide with artists losing their ability to compete. This is a simplified linear representation of the stages through which human art could eventually become devalued.
1.) AI augments human art and increases productivity, providing a competitive advantage.
2.) Too much similar competition leads to a saturated market: entry level jobs lost, reduced profit margins, difficulty gaining market share, damage to brand identity and loyalty, price wars begin.
3.) Supply exceeds demand leading to a flooded market. This causes a chain reaction: βrace to the bottomβ price wars -> vicious cycle of devaluation -> markets contract -> creative industries collapse -> companies that provide goods and services to those industries collapse.
4.) Tech corporations secure full ownership over the means of production. The balance of labor and capital completely shifts to capital and workers lose all bargaining power. The full realization of the tragedy of the commons emerges as human creative labor is disincentivized, resulting in the depletion and barrenness of the commons.
Ethical AI
In the following, the proposed ethical models are meant to provide a rough concept and starting point to stimulate conversation, not to serve as a literal template. Ethical models that facilitate fair competition would look something like this:
Hypothetical Models For Visual Art
1.) The model would be sandboxed into functioning solely as a tool. The generative tool would not create anything new (with some minor exceptions), but be used to repair, correct, enhance, or modify that which is already present. Possible applications in Photoshop: effects (text), layer styles, rendering, content aware fill, spot healing, etc.
2.) The model would serve as a reference tool (style transfer is imposed to default to a generic output: no color, no style, no character design, same compositional framing, same texturing, etc.)
3.) The model would be less restricted but contractually bound to conditions set forth by the artistβs license. Details are described below.
The Third Model
By default, style transfer would be imposed on the outputs of the model to make certain proprietary aspects generic, e.g., style, character design, color palette, brushwork, etc. (a list needs to be carefully designated that serves to protect aspects of artistic work without restricting the model too much). The model could be used as-is for creative purposes and additionally function as a reference tool, but users would also have the option of purchasing licenses from artists to unlock their style and design.
Hereβs how it would look technically: Style transfer is imposed on the output to make it generic. When the user purchases a license, style transfer is used on top of the generic output to translate it into the style of the artist being licensed.
Conditions Of The Third Model
It is essential to establish conditions to safeguard against the model from becoming anti-competitive. These are some possible conditions:
β’ Artists set the prices for their opt-in license and for users to license their style, not the company.
β’ Opt-in licenses are non-perpetual and can be terminated at any point. The license specifies a limited number of derivatives that can be generated.
β’ Users have to purchase individual licenses from artists (vs the company charging a subscription fee to access all styles).
β’ Even though there are works in the dataset that are public domain/CC, the model must always default to a generic output unless a user purchases a license.
β’ The model cannot train on AI generated content, including derivatives generated from the model itself.
β’ Derivatives cannot be licensed back to the company.
β’ Opt-ins cannot be AI generated or feature AI generation.
β’ Work that is used to fine-tune the model cannot be AI generated or feature AI generation.
β’ There is a robust verification process that guarantees the authenticity of an artist before they can fine-tune the model with their own work.
β’ All works that feature AI generation must be labeled as such and the label has to endure where that work is used.
β’ The company must abide by no scraping labels (e.g., robots.txt, NoAI, etc.).
An aside: regulations should ensure that as the AI systems become more powerful, their use becomes more restrictive; a balance is needed so that the creative stays competitive and is not made obsolete by the model.
Conclusion
Passionate activists are tirelessly dedicating their efforts to exposing the unethical and illegal practices perpetrated by prominent generative AI companies, aiming to hold them accountable. These individuals, grassroots volunteers who selflessly sacrifice their time, play an invaluable role that the creative community will forever be indebted to. However, it is crucial to ensure that their work is not in vain.
Presently, there are companies attempting to sway public opinion by marketing themselves as ethical. It is vital to recognize, however, that their purportedly ethical models may revert us to square one. An anticompetitive model presents a significant threat to the creative industry and has the potential to result in the widespread displacement of creative labor. We must keep at the forefront of our consciousness the understanding that any model failing to facilitate fair competition is unethical.
