Societal Risks and Well-Being
Dual-use foundation models with widely available model weights have the potential to create benefits across society, primarily through the access to AI capabilities that such models provide. At the same time, they also pose a substantial risk of causing harms to individuals and society. As noted above, our assessment of risk is tied to a framework of marginal risk: “the extent to which these models increase societal risk by intentional misuse beyond closed foundation models or pre-existing technologies.”90 Further, due to the relative novelty of dual-use foundation models, especially models that generate output in modalities beyond text (i.e., video and image generation), combined with known difficulties in accurate reporting for societal risks, precise estimates of the extent of these risks (especially the marginal risk of open foundation models over other models) are challenging to produce.91
The societal risks and benefits discussed in this section extrapolate from existing applications of open models that do not meet the parameter size criteria for this Report.92 For example, there are no text-to-image models with widely available model weights with over 10 billion (10B) parameters available today, while there are multiple text generative based open models over the 10B parameter threshold.93 Other developers of sophisticated closed-weight multi-modal models, such as SORA from OpenAI, have not publicly announced how many parameters they have.
It is also important to note that the content discussed in this section is not created in a vacuum. Both risks and benefits accrue due to how easily and widely accessible the tools for content creation are, as well as how the content is distributed. In the case of harmful content, some measure of that risk is dependent on how effectively platforms, content distributors, and others can prevent its widespread distribution. In the case of privacy and information security, there are still open questions as to how much the model “memorizes” from the data sets it was trained on and how much of that “memorization” contains personally identifiable information.94 These risks are also embedded in our social systems. While a number of social risks and benefits arise from open foundation models, this section covers only a select few.
Societal Risks of Widely Available Model Weights
CSAM & NCII
Models with widely available weights are already used today for AI-generated child sexual abuse material (CSAM), AI-generated non-consensual intimate imagery (NCII), and other forms of abusive content.95 Such content may include the depiction of wholly fabricated individuals as well as specific individuals created using preexisting images of them. This content disproportionately affects women and teens, although any individual can be affected, and creates a hostile online environment that undermines equitable access to online services.96 Such content, even if completely AI generated, can pose both immediate and long-term harm to its targets, especially if widely distributed, and creates a systemic risk across digital platforms for gender-based harassment and intimidation.97 98
Open foundation models lower the barrier to create AI-generated CSAM and NCII. Creating such content using an open foundation model requires only a set of images to fine-tune the model, as opposed to creating a model from scratch. Open foundation models and downstream applications built from them, such as so-called ‘nudifying’ apps,99 have made it easy to create with little to no cost individually targeted NCII and CSAM, which significantly enables both the production and distribution of AI-generated (but highly realistic) NCII and CSAM.100 They also make it easier to distribute CSAM and NCII because there is no limit on the amount of content that can be created, making it possible to rapidly generate large amounts of AI-genOpen foundation models lower the barrier to create AI-generated CSAM and NCII. Creating such content using an open foundation model requires only a set of images to fine-tune the model, as opposed to creating a model from scratch. Open foundation models and downstream applications built from them, such as so-called ‘nudifying’ apps,99 have made it easy to create with little to no cost individually targeted NCII and CSAM, which significantly enables both the production and distribution of AI-generated (but highly realistic) NCII and CSAM.100 They also make it easier to distribute CSAM and NCII because there is no limit on the amount of content that can be created, making it possible to rapidly generate large amounts of AI-generated material. In contrast, closed model providers can more easily restrict or prevent the creation of CSAM and NCII through restrictions on prompts as well as on APIs.
The threat of NCII in particular is not unique to open foundation models. In one of the most publicized incidents to date—when synthetic NCII images of singer Taylor Swift spread across the Internet in early 2024—the images were created with a closed generative model. While, as demonstrated in the Taylor Swift case, malicious actors can circumvent safety features on proprietary models to produce harmful content, most CSAM/NCII content is generated by open foundation models.
Prior to the wide availability of AI systems, synthetic CSAM (e.g., non-photo based CSAM) primarily focused on non-realistic categories of material, such as anime-styled CSAM.101 Open foundation models that include pornography or CSAM in their training data, as well as downstream implementations of open models that have been finetuned on CSAM or similar material, allow for the creation of AI-generated CSAM that is realistic to the point of being easily confused with non-AI-generated images and even based on real individuals.102 NCII content specifically is often based on individuals who have never shared any form of nude images online. Creating such content prior to the release of generative AI models at the level of realism now achievable previously required both considerable skill with photo editing tools as well as a significant investment of time.
While the threat of NCII specifically is not unique to open foundation models (in one of the most publicized incidents to date—when AI-generated NCII images of singer Taylor Swift spread across the Internet in early 2024—the images were created with a closed generative model), since the emergence of open foundation models, researchers have documented significant increases in AI-created CSAM103 and NCII. For example, the release of Stable Diffusion 1.5, an open model with 860 million parameters104 (and which was revealed to have included documented CSAM in its training data),105 enables the direct creation of CSAM, and fine-tuned versions of the model have been used in down stream apps.106 This increase in harmful content enables producers to flood online platforms with enough content to overwhelm platform trust and safety teams and law enforcement’s capacity to ingest and process CSAM reports.
Due to the sheer volume and speed of production they enable, the availability of open foundation models to create CSAM and NCII represents an increase in marginal risk over both existing closed foundation models and existing technologies.107 The legal and regulatory system devoted to investigating and preventing the distribution of CSAM is not equipped to handle this influx of content.108 As open foundation models become more advanced, this threat will likely increase.
The mass proliferation of CSAM and NCII also creates a substantial burden for women, teens and other vulnerable groups to live and participate in an increasingly online and digitized society. Further, proliferation of CSAM and NCII can discredit and undermine women leaders, journalists, and human rights defenders, and the implications of this harm extend beyond the individual to society and democracy at-large.109 Again, these are risks that do not exist in a vacuum; the magnitude of harm in part depends on the ability to distribute content at scale.
Political Deepfakes
The capacity for malicious actors to use open foundation models to create convincing political deepfake content across a variety of modalities has introduced marginal risk to the integrity of democratic processes.110 Actors are already generating and disseminating political deepfakes using downstream apps built upon both open and proprietary models. For example, deepfake audio recordings created with a proprietary voice cloning AI model emerged in January 2024, mimicking President Biden discouraging voters in the New Hampshire primary via robocalls.111 This incident created immediate and widespread concerns about the potential impact on voter turnout, as well as what such incidents portended for upcoming elections and the democratic process as a whole.112
Internationally, campaigners supporting the re-election of Prime Minister Modi and other politicians in India are using open models to create synthetic videos and deliver per sonalized messages to voters by name, creating concerns that the public may not be able to discern the fake content from authentic material.113 (At least some post-election accounts indicate that this concern failed to materialize and that generative AI enabled politicians to more easily communicate with voters in the 22 official languages of In dia.114) In 2022, a deepfake of Ukrainian President Volody myr Zelensky circulated widely online, in which the false imitation of President Zelensky urges Ukrainian soldiers to lay down arms.115
In the absence of detection, disclosure, or labeling of synthetic political content, malicious actors can create deepfake videos or audio messages to unduly influence elections or enable disinformation campaigns.116 Once re leased, they can be difficult to remove from the Internet, even after they have been verified as fake, in part due to the reluctance of social media platforms to remove this content. Deepfakes can also increase the “liar’s dividend”: skepticism and disbelief about legitimate content, contrib uting to pollution of the wider information environment.117 This could undermine democratic processes by confusing voters and reducing the public’s ability to determine fake events from actual ones.
As with concerns about CSAM and NCII, most of the open models capable of producing political deepfakes today have fewer than 10 billion parameters, but evidence does exist that political deepfake content has already been cre ated and disseminated using open models under the 10B threshold. While deepfakes are a widespread source of concern, current dual-use foundation models with widely available model weights may not substantially exacerbate their creation or inflict major societal damage given the existing ability to create deepfakes using closed models, but, as open foundation models develop, this risk may in crease.
Disinformation & Misinformation
Similar to the concerns described earlier regarding CSAM and NCII, the primary risks are tied to the low barriers to entry that open models may create for greater numbers of individuals, as well as to coordinated influence operations to create content and distribute it at scale.118 Further, researchers have expressed concerns that the capabilities of generative LLMs may allow foreign actors to create targeted disinformation with greater cultural and linguistic sophistication.119 As noted in the section on geopolitical considerations, the wide availability of open U.S. models could bolster dual-use AI innovation in countries of concern, which can enable them to develop more sophisticated disinformation campaigns.
DISINFORMATION
While the release of open foundation models raises concerns about the potential to enable disinformation campaigns by adversarial actors, assessments are mixed regarding whether open models, at least at current levels of capabilities, pose risks that are distinct from proprietary models. There is evidence that open foundation models, including LLMs, are already being used today to create dis information-related content.120 Disinformation researchers anticipate that generative models “will improve the content, reduce the cost, and increase the scale of campaigns; that they will introduce new forms of deception like tailored propaganda; and that they will widen the aperture for political actors who consider waging these campaigns.”121 As a consequence of an anticipated in crease in the production of disinformation related content, one commenter expressed concerns that such content produced at significant enough of a scale would later be ingested by AI systems as training data, perpetuating its half-life.122
While many agree that open foundation models enable a larger range of adversarial actors to create disinformation, others dispute the importance of this assertion. Some re searchers argue that the bottleneck for successful disinformation operations is not the cost of creating it.123 Because the success of disinformation campaigns is dependent on effective distribution, key to evaluating marginal risk is whether the potential increased volume alone is an important factor, such that it may overwhelm the gatekeepers on platforms and other distribution venues. Some are skeptical that this is the case.124 Carnegie researchers argue that not only has disinformation existed long before the advent of AI, but that generative AI tools may prove useful to researchers and others combating disinformation.125
Misinformation
Unlike disinformation, which implies intentional malfeasance, misinformation encompasses factually incorrect information, or information presented in a misleading manner.126 All foundation models are known to create and even help propagate factually incorrect content.127 128 Malicious actors may intentionally use models to create this in formation, and models can unintentionally produce inac curate information, often referred to as “hallucinations.”129 130 The marginal risks open foundation models pose in regards to misinformation are similar to those raised by CSAM, NCII, deepfakes and disinformation: they may lower the bar for individuals to create misinformation at scale and allow for more prolific distribution of misinformation. These impacts may exacerbate the disruption to the overall information ecosystem131 and high volumes of misinformation may overwhelm information distributors’ capacity to identify and respond to misinformation. How ever, some researchers argue that consumption of misinformation is limited to individuals more likely to seek it out and that foundation models do not substantially alter the amount or impact of this content online.132
One aspect of this marginal risk that necessitates further study is how individuals react to misinformation when it is directly outputted from an AI system (e.g., a chatbot) compared to consumed on social media or another plat form. Little research to date exists that interrogates the consumption of misinformation directly from AI powered tools.133 The majority of the public does not yet seem to understand generative AI’s propensity for producing in accurate information and may place undue trust in these systems; there have been well-publicized instances of lawyers, for example, relying on ChatGPT to assist with brief writing only to be surprised and embarrassed to find that the tool fictionalized legal sources.134 In other instances, generative models have output untrue and potentially slanderous information about individuals, including public figures.135
Discriminatory Outcomes
Discrimination occurs when people are treated differently, solely or in part, based on protected characteristics such as gender, religion, sexual orientation, national origin, col or, race, or disability.136 Discrimination based on protected classes is, unfortunately, a widespread issue, impacting many groups by race, gender, ethnicity, disability, and other factors.137 There has been substantial documentation of AI models, including open foundation models, generating biased or discriminatory outputs, despite developers’ efforts to prevent them from doing so.138
Open foundation models may exacerbate this risk be cause, even if the original model has guardrails in place to help alleviate biased outcomes, downstream actors can fine-tune away these safeguards. These models could also be integrated into rights-impacting systems with little oversight and no means of monitoring their impact. Thus, it may be difficult to prevent open foundation models or their downstream applications from perpetuating biases and harmful institutional norms that may impact individuals’ civil rights.139 Bias encoded in foundation models be comes far more powerful when those models are used in decisional contexts, such as lending,140 health care,141 and criminal sentencing.
From a regulatory perspective, commercial actors that implement tools built on open foundation models will be subject to the same federal civil rights laws as those who leverage single-use models or proprietary foundation models.142 The breadth of the potential impact and the current lack of clear determination regarding how to eliminate bias from all forms of AI models indicates that more research is needed to determine whether open foundation models substantially change this risk.
Societal Benefits of Widely Available Model Weights
Releasing foundation model weights widely also introduces benefits for society. Specifically, widely available model weights can: (a) support AI use for socially beneficial initiatives; (b) promote creativity by providing more accessible AI tools for entrepreneurial or artistic creation and expression; and (c) provide greater ability to test for bias and algorithmic discrimination.
Open Research for the Public Good
Making foundation model weights widely available allows a broader range of actor researchers and organizations to leverage advanced AI for projects aimed at improving public welfare. This approach democratizes access to cutting-edge technology, enabling efforts across health care,143 environmental conservation,144 biomedical innovations, and other critical areas that benefit society.145, 146, 147 Scientists and researchers can tailor open models to better suit specific research needs or experimental parameters, enhancing the relevance and applicability of their work.148 This customization capability is crucial for advancing scientific inquiries that require specialized models to analyze unique datasets or to simulate particular scenarios.149
AI Access for Entrepreneurs and Creatives
The wide availability of model weights can catalyze creativity and innovation, providing entrepreneurs, artists, and creators access to state-of-the-art AI tools.150 By lowering barriers to access, a broader community can experiment with AI, leading to novel applications and creations. New entrants into the marketplace would not have to pay for a closed model, for example, to utilize the benefits of advanced AI systems. This democratization fuels a wide range of entrepreneurial ventures and artistic expressions, enriching the cultural landscape and reflecting a diverse array of perspectives and experiences.151 This democratization can be particularly beneficial for small and medium-sized enterprises, which may otherwise face significant barriers to accessing more advanced AI systems.
Bias and Algorithmic Discrimination Mitigation
The ability to test for bias and algorithmic discrimination is significantly enhanced by widely available model weights. A wider community of researchers can work to identify biases in models and address these issues to create fairer AI systems. Including diverse communities and participants in this collaborative effort towards de-biasing AI and improving representation in generative AI is essential for promoting fairness and equity. mitigating bias in AI. Commenters have highlighted the importance of transparency and oversight enabled by open model weights in attempting to fight bias and algorithmic discrimination in foundation models.152 As more companies, local governments, and non-profits use AI for rights-impacting activities, such as in healthcare, housing, employment, lending, and education, the need to better understand how these systems perpetuate bias and discrimination. There is a long history in the civil rights community of collaborative testing, and the wide availability of model weights enables that tradition to continue.153
Next: Competition, Innovation, and Research
90 Kapoor, S. et al., (2024). On the Societal Impact of Open Foundation Models. at 2. ArXiv.
91 Nestor Maslej, et al. Artificial Intelligence Index Report 2024. (2024 April).
92 See, e.g., Johns Hopkins Center for Health Security Comments at 2 (Acknowledging that at current, Evo, a biological design tool (BDT) “…is a 7-billion parameter model and so [is] below the EO size threshold for a dual-use foundation model…”); and The Future Society Comments at fn. 20, citing Kapoor, S. et al., (2024). On the Societal Impact of Open Foundation Models. ArXiv. (”AI-generated pornography based on Stable Diffusion offshoots quickly spread across the internet, including images resembling real people generated without their consent.”). The largest version of Stable Diffusion, Stable Diffusion XL, has only 3B parameters (HuggingFace).
93 Ecosystem Graphs for Foundation Models. (n.d.).
94 See generally, Foundation Model Privacy. IBM Research. (“Language models have an inherent tendency to memorize and even reproduce in their outputs text sequences learned during training, may this be pre-training, fine-tuning or even prompt-tuning. If this training data contained sensitive or personal information, this could result in a major privacy breach.”); Hartmann, V., et al. (2023). SoK: Memorization in General-Purpose Large Language Models. (“In many cases of interest, such as personal identifiers, social security numbers or long passages of verbatim text, it is unlikely that a model could hallucinate the target information or gain knowledge of it through reasoning.”); Huang, J., Shao, H., & Chang, K. C.-C. (2022, May 25). Are Large Pre-Trained Language Models Leaking Your Personal Information? ArXiv. (“We find that PLMs do leak personal information due to memorization.”).
95 See Thorn Comments at 1 (“One concrete risk that is already manifesting as a harm occurring today, is the misuse of broadly shared and open source foundation models to make AI-generation child sexual abuse material. This technology is used to newly victimize children, as bad actors can now easily sexualize benign imagery of a child to scale their sexual extortion efforts… . This technology is further used in bullying scenarios, where sexually explicit AI-generated imagery is being used by children to bully and harass others.”) (citations omitted).
96 2023 State of Deep Fakes. (2023).
97 Home Security Heroes. (2023). 2023 State of Deepfakes: Realities, Threats, and Impacts. Home Security Heroes. (“99% of the individuals targeted in deepfake pornography are women.”), and Eaton, A. A., Ramjee, D., & Saunders, J. F. (2023). The Relationship between Sextortion during COVID-19 and Pre-pandemic Intimate Partner Violence: A Large Study of Victimization among Diverse U.S Men and Women. Victims & Offenders, 18(2), 338–355.
98 Internet Watch Foundation (2023). How AI is being abused to create child sexual abuse imagery; Kang, C. (2024). A.I.-Generated Child Sexual Abuse Material May Overwhelm Tip Line. NYTimes.
99 How AI is being abused to create child sexual abuse imagery. (2023). Internet Watch Foundation.
100 Kapoor S. et al., (2024). On the Societal Impact of Open Foundation Models. ArXiv.
101 Thiel, D., Stroebel, M., & Portnoff, R. (2023, June 24). Generative ML and CSAM: Implications and mitigations. FSI.
102 How AI is being abused to create child sexual abuse imagery. (2023). Internet Watch Foundation.
103 Thiel, D. (2023). Generative ML and CSAM: Implications and Mitigations. Stanford.
104 Stable Diffusion Public Release. (2023, August 22). Stability AI.
105 Thiel, D. (2023). Identifying and Eliminating CSAM in Generative ML Training Data and Models. Stanford Internet Observatory.
106 Open Technology Institute Comments at 11, citing Thiel, D. (2023). Generative ML and CSAM: Implications and Mitigations. Stanford Internet Observatory.
107 See Kapoor, S. et al., (2024). On the Societal Impact of Open Foundation Models. ArXiv. (discussion of why open foundation models present an increase in marginal risk specifically for NCII).
108 Keller, M., & Dance, G. (2019, September 29). Last year, tech companies reported over 45 million online photos and videos of children being sexually abused—More than double what they found the previous year. NYTimes.
109 Gendered Disinformation: Tactics, Themes, and Trends by Foreign Malign Actors - United States Department of State. (2023, April 12). United States Department of State.
110 EPIC comment attachment p. 3-4.
111 Knibbs, K. (2024, January 26). Researchers Say the Deepfake Biden Robocall Was Likely Made With Tools From AI Startup ElevenLabs.
112 Elliott, V., & Kelly, M. (2024, January 23). The Biden Deepfake Robocall is Only the Beginning.
113 Suhasini, R. (2024, April 18). How A.I. Tools Could Change India’s Elections.
114 Christopher, N. (2024, June 5). “The Near Future of Deepfakes Just Got Way Clearer.” The Atlantic.
115 Allyn, B. (2022, March 16). Deepfake video of Zelenskyy could be “tip of the iceberg” in info war, experts warn.
116 See, “Deepfake” of Biden’s Voice Called Early Example of US Election Disinformation. (2024, January 24). Voice of America; Hartmann, T. (2024, April 16). Viral deepfake videos of Le Pen family reminder that content moderation is still not up to par ahead of EU elections. www.euractiv.com; Misinformation and disinformation. APA. (n.d.). “False information deliberately intended to mislead.”
117 Lohn, A. (2024, January 23). Deepfakes, Elections, and Shrinking the Liar’s Dividend.
118 Josh A Goldstein, et al. (2024 February) How persuasive is AI-generated propaganda?. PNAS Nexus.
119 J.A. Goldstein, et al. (2023) Generative Language Models and Automated Influence Operations: Emerging Threats and Potential Mitigations. ArXiv.
120 See Access Now Comment at 2 “Nefarious actors can access them, remove built-in safety features, and potentially misuse them for malicious purposes, from malevolent actors creating disinformation to generate harmful imagery and deceptive, biased, and abusive language at scale.”
121 J.A. Goldstein, et al. (2023) Generative Language Models and Automated Influence Operations: Emerging Threats and Potential Mitigations. ArXiv.
122 Fergusson, G., & et al. (2023). Generative Harms Generative AI’s Impact & Paths Forward (p. 3). EPIC. Comment NTIA-2023-0009-0206.
123 Kapoor, S., & Narayanan, A. (2023, June 16). How to Prepare for the Deluge of Generative AI on Social Media.
124 Kapoor, S. et al., (2024). On the Societal Impact of Open Foundation Models. ArXiv.
125 Bateman, J., & Jackson, D. (2024). Countering Disinformation: Effectively An Evidence Based Policy Guide (p. 87). Carnegie Endowment.
126 American Psychological Association. Misinformation and disinformation.
127 Perrigo, B. (2023, October 26). The Scientists Breaking AI to Make It Safer. Time.
128 Heikkilä, M. (2023, February 14). Why you shouldn’t trust AI search engines. Technology Review.
129 Bender, E., et al. (2021). On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?.
130 Chen, C. & Shu, K. (2023). Combating Misinformation in the Age of LLMs: Opportunities and Challenges. ArXiv preprint.
131 Simon, F. M., Altay, S., & Mercier, H. (2023). Misinformation reloaded? Fears about the impact of generative AI on misinformation are overblown. Harvard Kennedy School (HKS) Misinformation Review.
132 Simon, F. M., Altay, S., & Mercier, H. (2023). Misinformation reloaded? Fears about the impact of generative AI on misinformation are overblown. Harvard Kennedy School (HKS) Misinformation Review.
133 See generally: Florian Leiser, et al. (2023) From ChatGPT to FactGPT: A Participatory Design Study to Mitigate the Effects of Large Language Model Hallucinations on Users. In Proceedings of Mensch und Computer 2023 (MuC ‘23). Association for Computing Machinery, New York, NY, USA, 81–90; Marita Skjuve, Asbjørn Følstad, and Petter Bae Brandtzaeg. 2023. The User Experience of ChatGPT: Findings from a Questionnaire Study of Early Users. In Proceedings of the 5th International Conference on Conversational User Interfaces (CUI ‘23). Association for Computing Machinery, New York, NY, USA, Article 2, 1–10.
134 Neumeister, L. (2023, June 8). Lawyers blame ChatGPT for tricking them into citing bogus case law.
135 Neumeister, L. (2023, June 8). Lawyers blame ChatGPT for tricking them into citing bogus case law.
136 See generally, Wex Legal Dictionary and Encyclopedia. Discrimination. Legal Information Institute. (Last Updated November 2023).
137 See American Civil Liberties Union, Center for American Progress, and the Leadership Conference on Civil and Human Rights Comment at 3 “These examples span the range of technology that may be deemed “artificial intelligence,” and the emergence of applications such as chatbots and facial recognition technology underscores that both rudimentary and the most sophisticated AI technologies are already affecting civil rights, safety, and access to opportunities.”
138 See generally: Xiang, C. (2023, March 22). The Amateurs Jailbreaking GPT Say They’re Preventing a Closed-Source AI Dystopia; Knight, W. (2023, December 5). A New Trick Uses AI to Jailbreak AI Models-Including GPT-4; Ananya. (2024, March 19). AI image generators often give racist and sexist results: Can they be fixed?; Hofmann, V. (2024). Dialect prejudice predicts AI decisions about people’s character, employability, and criminality.
139 Wiessner, D. (2024, February 21). Workday accused of facilitating widespread bias in novel AI lawsuit.
140 Sadok, H., Sakka, F. & El Maknouzi, M. (2022). Artificial intelligence and bank credit analysis: A Review. Cogent Economics Finance, 10(1).
141 Juhn, Y. et al (2022). Assessing socioeconomic bias in machine learning algorithms in health care : a case study of the HOUSES index. Journal of the American Medical Informatics Association, 29(7), 1142-1151.
142 FTC Chair Khan and Officials from DOJ, CFPB and EEOC Release Joint Statement on AI. (2023, April 25).
143 See Connected Health Initiative Comment (“Successful creation and deployment of AI-enabled technologies which help care providers meet the needs of all patients will be an essential part of addressing this projected shortage of care workers. Policymakers and stakeholders will need to work together to create the appropriate balance between human care and decision-making and augmented capabilities from AI-enabled technologies and tools.”).
144 See Caleb Withers Comment at 9 (“Illustratively, the best coding models have either been, or been derived from, the most capable general-purpose foundation models, which are typically trained on curated datasets of coding data in addition to general training.”).
145 See Caleb Withers Comment at 9 (“Illustratively, the best coding models have either been, or been derived from, the most capable general-purpose foundation models, which are typically trained on curated datasets of coding data in addition to general training.”).
146 National Institutes of Health (n.d.). Mission and Goals. Department of Health and Human Services.
147 Bozeman, B. & Youtie, J. (2017). Socio-economic impacts and public value of government-funded research: Lessons from four UN National Science Foundation initiatives. Research Policy 46(8) 1387-1389.
148 See Center for Democracy & Technology, et al. (2024, March 25). RE: Openness and Transparency in AI Provide Significant Benefits for Society. (“Open models promote economic growth by lowering the barrier for innovators, startups, and small businesses from more diverse communities to build and use AI. Open models also help accelerate scientific research because they can be less expensive, easier to fine-tune, and supportive of reproducible research.”).
149 See ACT Association Comment (“For example, healthcare treatments and patient outcomes stand poised to improve disease prevention and conditions, as well as efficiently and effectively treat diseases through automated analysis of x-rays and other medical imaging.”) and AI Policy and Governance Working Group at 5 (“Making foundation models more widely accessible, with appropriate safeguards, could drive innovation in research and business capitalizing on the promise of public benefit. Study and use of state-of-the-art AI models, including Large Language Models and other models like AlphaFold, may lead to improvements in performance, safety, and scientific breakthroughs across various domains. These potential benefits can best be realized if other AI model assets, such as model training data, are also made widely available, and if models are not subject to restrictive licenses. Areas that stand to potentially gain from a commitment of ensuring the wide availability of AI tools and systems include, but are not limited to, innovation and novel applications in public health, biomedical research, and climate science that might be scaled in the public interest. Any decision to constrain the availability of dual-use open foundation models must carefully weigh and consider these potential societal and economic benefits.”).
150 A generative AI tool to inspire creative workers. (2024, February 14). MIT Sloan.
151 Criddle, C. & Madhumita M. (2024, May 8). Artificial intelligence companies seek big profits from ‘small’ language models. Financial Times.
152 American Civil Liberties Union, Center for American Progress, and the Leadership Conference on Civil and Human Rights Comment at 4 “In addressing AI’s risks for civil rights, safety, and access to opportunity, advocates, affected communities, and policymakers have championed a number of regulatory goals, including auditing and assessments, transparency, and explainability.”; Hugging Face Comment at 6 “Maximally open systems, including training data, weights, and evaluation protocol, can aid in identifying flaws and biases. Insufficient documentation can reduce effectiveness”; Google Comment at 7 “Openly available models also enable important AI safety research and community innovation. A diverse pool of available models ensures that developers can continue to advance critical transparency and interpretability evaluations from which the developer community has already benefited. For example, researchers have demonstrated a method for reducing gender bias in BERT embeddings.”
153 Fair Housing Testing Program. (2015, August 6). Justice.gov. (“In 1991, the Civil Rights Division established the Fair Housing Testing Program within the Housing and Civil Enforcement Section, which commenced testing in 1992.”).