Last Updated: 2023-06-29
New AI technologies represent a significant leap forward but are built upon the work of hundreds of millions of Americans. Big tech companies have used data from the American people without any compensation.
This policy would fairly compensate the American public for its contribution to generative AI technologies by requiring big tech companies to pay the American people for use of their data. It would place a fee on any generative AI technology from a big tech company where the technology incorporates data from anyone who did not explicitly assign rights for such use.
See our piece published in Politico, June 29, 2023:
AI Could Pay Dividends to Americans by Barath Raghavan and Bruce Schneier
Q: Shouldn’t we focus on potential harms from AI instead of on uncompensated use of Americans’ work and data?
A: New technologies change rapidly and it is impossible to predict which harms are most likely and who they will affect. This policy is about fairly compensating the American people for use of their data, not remedying harms. Specific harms can and should be addressed by Congress and regulators independent of this policy.
Q: Why does the policy pay the dividend to everyone?
A: AI technologies use vast quantities of data from across the Internet, and it is impossible to determine whose data and creative work is most important to any specific AI technology or use case. Such technologies are built using the American people’s work, and so the American people should be paid for their contributions.
Q: How does this interact with copyright law and ongoing cases?
A: This policy would not affect anything beyond fair use of public material. We’ll all have to see how ongoing copyright cases are resolved.
Q: What if the fees are too low? What if the fees are too high?
A: The initial fees are intended to be set at a level that provides meaningful but modest compensation to American taxpayers for their contribution to new AI technologies. Over time, if these fees are too low or do not cover new technologies, they can be updated.
Q: Shouldn’t other technologies also have a fee applied?
A: This policy will enable the Department of Commerce to include additional AI technologies as warranted, since technologies change rapidly.
Q: How will this affect AI innovation?
A: Generative models can produce content much more cheaply than humans, so there will still be strong incentives for companies to train and run models, even with the fees suggested in this proposal. Even if the fee were 100 times the cost of generating text with ChatGPT, for example, it is still likely to be affordable for most uses, so OpenAI would still have an incentive to invest in model innovation.
Q: How will this affect hobbyists, research labs, and small businesses?
A: Independent, free, non-commercial, and small business AI technology, products, and services won’t need to pay any fees. Only when public data is used by big tech, defined as an organization that exceeds an annual revenue threshold, does the fee apply. This keeps administration of the policy simple, encourages innovation, and appropriately targets the fees to those big tech companies that are profiting from the American people’s data.
Q: What happens when the company that trains an AI model isn’t the same company that generates content with it?
A: Since the fee is charged at the point where words, pixels, or other tokens are first produced, only the companies enabling AI-generated content will need to pay the fee. If big tech Company A trains a large language model and sells the model to Company B, which then uses it, Company A will be required to pay the fees for the output of its model, just as if Company A provided API-based access to the model.
Big tech companies can’t just circumvent these fees by offering public access to generated content for free. Some companies might want to offer free access to generated content as a loss leader to encourage consumers to buy their other products, but companies with over the revenue threshold will still be required to pay the fees if they do so.
Q: What types of generative AI products or services will this apply to?
A: This proposal will apply to language models (like ChatGPT or Bard), image models (like Midjourney), and other content-producing models trained on data from the public. Niche models, such as those trained on financial databases, would also be covered by this proposal if they use public data. The Department of Commerce can apply the fee to new types of products or services as they are brought to market, enabling the policy to be flexible as the technology evolves.
Q: Isn’t training data global? Why isn’t this a global policy? Why does it apply to all foreign companies of any size?
A: While AI models are trained on data from around the world, data from the US is most valuable in models used in the US, just as data from another country is most valuable for models used in that country. It makes sense for the American people to be paid for use of their data in services within the US. Other countries can put into place similar policies as they see fit. In addition, while the policy is intended to only apply to big tech companies, the policy applies to all services, with no size threshold, outside the US. This is to ensure that foreign companies (or shell companies controlled by US-based big tech) cannot gain an unfair advantage by operating outside of the US but offering the services to US-based users.
This policy will apply a licensing fee to generative AI output and return the proceeds to the American people. The policy will apply to all generative AI provided by any organization that itself has, or whose parent organization(s) have, more than $10,000,000 per year in total revenue. The policy applies whether the generative AI is offered as a service or as a non-service product, regardless of the price of the service or product. The fee applies at the point of generation of output from a generative AI technology. The policy does not apply to secondary or later uses of the output, with one exception: if the secondary or later uses of the output are themselves generative AI technologies. The fee applies to all generative AI services and products from foreign organizations, whether or not they are affiliates or subsidiaries of US-based organizations, regardless of the organization’s revenue. A product or service is exempt only if it exclusively uses private data for which explicit and affirmative rights for use in AI were granted by the creator(s) of the data at the time of creation.
Fee collection and disbursement: Companies covered under the policy must remit the fees monthly to the AI Dividend fund managed by the Department of Commerce, which will disburse the entirety of the fund each quarter (less administrative costs, which are to be no more than 1% of disbursements) in equal shares to all US adults. The first dividend will be via paper check to ensure public awareness; subsequent dividends will allow for direct deposit.
Fees: Fees for generative AI output shall be as follows, which may be increased subsequently:
The Department of Commerce may add additional covered AI technologies, products, services, and companies. It may also add additional types of fees and increase existing fees. All such fees must be remitted entirely to the AI Dividend fund and disbursed to the American people as above.
Last Updated: 2023-06-29