This is an experiment, to see if I can collect together a more complete model of the risks associated with a single generative AI system. It is a long way from being complete, so if you notice anything I’ve missed, please feel free to email me at morungos@gmail.com, or message me on Mastodon. Not Twitter, because Elon.
As a first example, I’ve used OpenAI’s DALL-E 2 system as a starting point, and in particular I’ve drawn on the preview system card (since no other is currently available), and the content policy. That enables me to be specific and concrete about the risks. To compile the map I’ve drawn on LLM sources extensively too, and especially Bender et al.’s (2021) “Stochastic Parrots” paper. Generally, I suspect most large-scale generative AI systems (e.g., ChatGPT, Midjourney) will follow a different pattern, although the communities affected might be a little different. However, risks that do not apply to DALL-E 2 specifically, and mitigations used in other tools, I will silently pass over.
The aim is to explore the following:
A few observations so far:
The full map is below.
Category | Risk | Winners | Losers | Mitigations | Notes, sources, and examples |
---|---|---|---|---|---|
Environmental | Increased environmental impact from OpenAI | OpenAI | All | None |
See Bender et al., (2021) High resource requirements reduce competition. Competitive disadvantage is skewed globally and economically. High compute requirement is paradoxically good for OpenAI. |
Increased environmental impact from competing vendors | Other AI vendors | All | None | ||
Legal | Unlicensed copyrighted images in training data | OpenAI | Image creators | Some (filtering) |
Mitigations mainly benefit OpenAI. Data is withheld to prevent both competition and scrutiny.
Mitigations primarily protect content, and, therefore, legal exposure. See: DALL-E 2 system card |
Prompt attacks to retrieve images from the training data | Unethical users | OpenAI | Significant | ||
Identifiable people in generated images | Users | Individuals (especially well-known) portrayed in training data | None | ||
User risks | Usage to violate copyright | Unethical users and groups | Image creators | Terms and conditions only |
Generally, risks
are offloaded into users through terms and conditions: See: DALL-E 2 content policy |
Intentional use to misinformation/deception, e.g., for political gain | Unethical users and groups | All | |||
Intentional use to induce emotional reactions for propaganda/manipulation, e.g., for political gain | Unethical users and groups | All | |||
Intentional use to generate images for harassment | Unethical users and groups | Harassed individuals & their networks | |||
Intentional use to generate explicit images | Unethical users and groups | All | |||
Intentional use to generate hateful content | Unethical users and groups | Minorities | |||
Intentional use for criminal purposes, e.g., blackmail, fraud | Unethical users and groups | Victims of crime | |||
Intentional use for negative but non-criminal acts, e.g., manipulating social media | Unethical users and groups | All | |||
Use to create fake personas to conceal misinformation or propaganda | Unethical users and groups | All | None |
See: McGuffie & Newhouse (2020) May be implied under misinformation, but unclear |
|
Competitive advertising: cheap generated images drive out real products | Unethical users and groups | Traditional advertisers, customers | None | Not an acknowledged risk | |
Uploading pictures of people without consent | Unethical users and groups | Targeted individuals | Terms and conditions only |
See DALL-E 2 content policy Note application to those who cannot consent (deceased people, minors) is unclear |
|
Social | Propagation of biased content: Western-typical image content tends to supplant other content | Western-culturally aligned | Non-Western culturally aligned | None |
See Bender et al., (2021) Acknowledged in DALL-E2 system card |
Propagation of erasure: atypical image content may be erased | Typical image content | Atypical image content | None | ||
Propagation of stereotypes through generated images | Positive stereotypes | Negative stereotypes | None | ||
Propagation of antiquated values: ‘value-lock’ constrains a model to outdated values, erasure of poorly-documented social movements | Conservative values | Progressive values | None | See Bender et al. (2021) Absent from DALL-E2 system card |
|
Debasement of art, through flooding with biased content | OpenAI | Image creators, wide art community | None | Bender et al.'s "ersatz fluency" (2021) Absent from DALL-E 2 system card |
|
Economic | Reduced work for artists and creators | Users | Image creators | None | Loss of work will not be evenly distributed. Unemployment is possible for creators with less privilege, support |
Reduced demand for, e.g., models, studios | Users | Models, studios | None | ||
Reduced sales for photographic equipment | Users | Camera stores & manufacturers | None | These are uncertain. It is possible that sales could even increase due to generation of new interest | |
Reduced sales for art supplies | Users | Art stores & companies | None | ||
Conflict and loss of integrity within the creator community, e.g., debased art competitions | Users | Artists | None | See, for example, this report on the Colorado State Fair's fine arts competition | |
Early invite access enables competitive advantage for insiders | Users connected to OpenAI | Users not connected to OpenAI | None | Acknowledged in DALL-E2 system card | |
Traditional methods of image creation lose economic viability against generative AI | Users | Image creators | None | Acknowledged in DALL-E2 system card |
Photo by Aksonsat Uanthoeng