Adobe Comments on the National Science Foundation’s Request for Information on the Development of an Artificial Intelligence Action Plan
Undated · The document carries no date anywhere on its 11 pages — no cover date, no signature block and no footer date. The sorting date 2025-03-15 is the closing date of the Request for Information (90 FR 9088) to which it responds, i.e. the submission deadline, not a printed date; the submission can have been sent no later than that. The id year 2025 rests on the same fact and is corroborated by the NITRD file name (Adobe-RFI-2025.pdf) and by the text itself, which cites executive orders signed in January 2025 (14149, 14158, 14179). The NITRD compilation PDF was created on 2025-04-14, when the government published the responses, which is later than the submission. The host, files.nitrd.gov, is the US government's published copy of the submissions, not Adobe's own domain; no Adobe-hosted copy was found.
Not law. This is a company's own public position on AI regulation. It is not law, and it carries no legal force.
What it argues for
This is Adobe's submission to the spring 2025 consultation that fed the White House's AI Action Plan, and it reads as the policy agenda of a company whose AI business rests on two things: provenance technology it built and a generative model it markets as trained on licensed content. Its headline ask is that Washington make the Adobe-led C2PA "Content Credentials" standard the global norm for content provenance: push it through international standards bodies, direct federal agencies to attach it to all government content, get smartphone makers to build it in, and go as far as a legal duty on platforms — "We urge the Administration to require online platforms to maintain and display any Content Credentials present in digital content on their systems." On regulation generally it is deregulatory and use-based: "Adobe supports a risk-based approach to AI development which allows innovation to rapidly accelerate by reducing unnecessary regulatory barriers and inspiring consumer confidence", and it dismisses the three approaches other jurisdictions have taken — model-size analysis, catastrophic-safety scrutiny, and "implementing onerous, and trade secret-compromising, training data transparency requirements" — with "None of these approaches achieve the goal of spurring innovation and driving widespread adoption of AI by consumers." It offers Adobe's own internal review process as the model, saying it "should serve as an industry standard for all technology companies." On training data it asks the government to "take a bold, pro-innovation stance and be declarative on the issue" of fair use, paired with an opt-out standard carried by Content Credentials and a new federal anti-impersonation right for artists whose style is copied. In place of binding rules it proposes a federal AI Testing Pilot with benchmark datasets (covering violence, sexual content and self-harm), a classification system, and company self-certification that would give participants "a safe harbor against future litigation." It closes by urging faster federal procurement of "commercially safe" AI — tools trained on licensed content, which is how Adobe positions Firefly.
Stated positions (15)
- Content provenance is the lead ask: the Administration should use its convening power in international bodies to advance C2PA, since "Methods to improve awareness and transparency regarding the origins of digital content is ripe for U.S.-led global standardization."
- Frames provenance as a free-speech tool rather than content moderation: Content Credentials let the U.S. "set a global benchmark for AI transparency and help keep the government and big tech out of the censorship business."
- Asks the White House to "issue guidance directing government agencies to implement the C2PA standard and Content Credentials for all government digital content" and to "stand up a federal interagency task force" to drive adoption by the government and its partners.
- Seeks a binding duty on platforms, not just a voluntary standard: "We urge the Administration to require online platforms to maintain and display any Content Credentials present in digital content on their systems", because "Currently, Content Credentials are often stripped away on internet platforms".
- Wants provenance built into phones: "We urge the Administration to encourage smartphone manufacturers to implement provenance capabilities, so that users have the option to attach content provenance metadata to all newly created and captured content."
- Backs a risk-based regime limited to high-risk uses: "Establishing a risk-based approach to AI governance that focuses on only the most high-risk use cases will foster innovation with the added benefit of requiring organizations to conduct a more rigorous and thorough review of high-risk Al systems."
- Rejects the approaches of other jurisdictions — model-size analysis, catastrophic-safety scrutiny and "implementing onerous, and trade secret-compromising, training data transparency requirements" — saying "None of these approaches achieve the goal of spurring innovation and driving widespread adoption of AI by consumers."
- Offers its own governance as the template: Adobe asks the government "to consider Adobe’s approach to AI development as a framework that should be encouraged globally", describing an AI impact assessment and AI Review Board process that includes "ensuring models are free from ideological bias".
- Treats user feedback as a substitute for regulatory reporting: a feedback mechanism "eliminates the need for government to intervene and require onerous reporting requirements that risk compromising trade secrets."
- On copyright, asks for a clear federal answer that training is fair use: "we encourage the Administration to take a bold, pro-innovation stance and be declarative on the issue, giving American AI developers of all sizes a clear framework and certainty in the marketplace".
- Pairs that with an opt-out standard, in which creators use Content Credentials to "attach a “do not train” tag to an individual piece of content that remains associated with the work wherever it goes."
- Proposes a new federal anti-impersonation right against AI copying of an artist's style, which "would provide a new mechanism for artists to protect their livelihood from people misusing this new technology without having to rely solely on mismatched existing laws around copyright."
- Proposes a government-convened AI Testing Pilot with benchmark datasets and a classification system instead of binding rules: "Rather than imposing burdensome government requirements that could restrict private sector AI development and deployment, the Administration has an opportunity to take an alternative regulatory approach". The benchmarks "could address categories such as violence, sexual content, and self-harm."
- Wants self-certification against those benchmarks to carry legal protection: companies would participate "as it creates a safe harbor against future litigation", with third-party auditors available to verify results.
- Urges faster federal procurement of what it calls commercially safe AI: "Commercially safe AI tools, those that train only on licensed content, or content where permission is not required by law, are ripe for expedited procurement".
About this document
An 11-page PDF headed "Adobe Comments on the National Science Foundation’s Request for Information on the Development of an Artificial Intelligence Action Plan", about 4,800 words, published by the US government on files.nitrd.gov as part of its compilation of responses to the RFI. It is issued in Adobe's corporate name; no individual signs it, no author is named and there is no date. It opens with an Introduction, which includes the RFI's required statement "This document is approved for public dissemination.", then a Background section on Adobe's products and Firefly, and an Executive Summary listing five themes. Five headed sections follow, one per theme: strengthening American leadership in AI technical standards (four bulleted asks on C2PA and Content Credentials); a risk-based approach to AI development (three bullets); long-term access to training data (three bullets on fair use, an opt-out standard and an anti-impersonation right); an AI Testing Pilot and standardisation benchmarks (two bullets on benchmark datasets and self-certification); and using AI to modernise the federal government (four bullets). A short Conclusion thanks NSF and OSTP. It cites Executive Orders 14149, 14158, 14179 and 13859, names no bill and no foreign law, and has no footnotes.
How this sits against AI law
Each stance compared with what EU and US instruments actually require. Where no instrument addresses a theme, that gap is shown rather than hidden.
An open provenance standard (C2PA Content Credentials) as the global and federal norm
The US should lead global standardisation of content provenance through C2PA, direct federal agencies to attach Content Credentials to all government digital content, and set up an interagency task force to promote adoption. Adobe presents this as a way to let consumers judge content for themselves and keep "the government and big tech out of the censorship business."
Article 50(2) requires providers of AI systems that generate synthetic audio, image, video or text to mark the outputs in a machine-readable format so that they are detectable as artificially generated, which is the job Content Credentials do; the Act prescribes the outcome, not a named standard, and it covers only AI output, whereas Adobe wants provenance on all digital content.
The Action Plan as adopted in July 2025 does not mention content provenance or Content Credentials; its only synthetic-media measures concern deepfakes in the legal system — a possible NIST guideline and voluntary forensic benchmark built on the Guardians of Forensic Evidence programme, and DOJ work on evidence rules.
Platforms and devices must carry provenance data
Adobe asks the Administration to "require online platforms to maintain and display any Content Credentials present in digital content on their systems", so that credentials are not stripped on upload, and to encourage smartphone makers to let users attach provenance metadata to everything they capture.
The Act's Article 50 transparency duties fall on providers and deployers of AI systems; it places no duty on online platforms to preserve or display provenance data in content they host, and none on camera or phone makers.
The California AI Transparency Act, as amended by AB 853, requires large online platforms from 1 January 2027 to detect and preserve provenance data embedded in content they distribute and to give users an interface to inspect it, prohibits stripping provenance data, and from 1 January 2028 requires capture-device makers selling in California to offer users latent provenance disclosures in captured content — the platform and device obligations Adobe asked for, enacted at state level.
Regulate high-risk uses, not model size or catastrophic risk
Adobe backs a risk-based approach "that focuses on only the most high-risk use cases", with light review for low-risk features, and rejects regulation built on model size or catastrophic-safety scrutiny, saying "None of these approaches achieve the goal of spurring innovation and driving widespread adoption of AI by consumers."
The Act does classify high-risk systems by use (Article 6 and Annex III), which matches Adobe's first half, but it also imposes duties on general-purpose models by size and risk: Article 51 presumes systemic risk above 10^25 floating-point operations of training compute, and Article 55 then requires model evaluation, systemic-risk assessment and mitigation, incident reporting and cybersecurity — exactly the model-size and catastrophic-risk approach Adobe rejects.
SB 53 applies to frontier developers training above 10^26 operations, with the heaviest duties on those with over $500 million in revenue, and requires them to publish frameworks for managing catastrophic risk — a compute-threshold, catastrophic-risk regime of the kind Adobe says does not work.
No mandatory training-data transparency
Adobe calls "onerous, and trade secret-compromising, training data transparency requirements" one of the approaches that fail, and argues that user feedback mechanisms remove the need for government-imposed reporting that risks exposing trade secrets.
Article 53(1)(d) requires every provider of a general-purpose AI model to draw up and publish a sufficiently detailed summary of the content used for training, according to a template provided by the AI Office.
AB 2013 requires developers of generative AI systems offered to Californians to post documentation of their training data on their website, including high-level summaries of datasets, their sources, whether they contain copyrighted material or personal information, and how they were obtained.
Training data: a declared fair-use rule plus a creator opt-out
The Administration should be "declarative" that training is fair use to give developers certainty, and in parallel advance an opt-out standard under which creators attach a machine-readable "do not train" tag through Content Credentials.
The EU chose the opt-out half of Adobe's pairing rather than a blanket permission: Article 53(1)(c) requires general-purpose model providers to have a policy to comply with Union copyright law, in particular to identify and comply with rights reservations expressed under Article 4(3) of Directive (EU) 2019/790 — the machine-readable opt-out Adobe wants standardised.
The Action Plan as adopted says nothing about copyright, fair use or training-data opt-outs; its data measures concern federal scientific datasets and research data sharing. No federal instrument in our corpus settles the fair-use question.
A federal anti-impersonation right against AI copying of an artist's style
Adobe proposes a new, narrow federal right letting creators act against people who intentionally and commercially use AI tools to impersonate their work and style, which copyright does not cover.
The Act has no style or likeness right; its nearest provision, Article 50(4), only requires deployers of deep-fake image, audio or video content to disclose that it was artificially generated or manipulated.
Tennessee's ELVIS Act, the closest instrument in our corpus, extends the state's publicity right to an individual's voice and likeness and creates liability for unauthorised AI use of them, but it protects identity, not artistic style, and applies only in Tennessee; Adobe's proposal covers style and is federal.
Government-convened benchmarks and self-certification instead of binding rules
Rather than "imposing burdensome government requirements", the government should run an AI Testing Pilot with benchmark datasets and a classification system built with industry, and let companies self-certify against the benchmarks, optionally verified by third-party auditors, in exchange for a litigation safe harbour.
The Act's high-risk regime runs on the same machinery: Article 40 gives a presumption of conformity to systems that follow harmonised standards, and for most Annex III systems Article 43 allows conformity assessment by internal control, i.e. the provider's own assessment. Unlike Adobe's model, conformity is mandatory for high-risk systems and gives no immunity from litigation.
The plan's 'Build an AI Evaluations Ecosystem' section funds evaluation science, NIST and CAISI guidelines and AI testbeds, and says regulators should explore using evaluations when applying existing law, but it creates no self-certification scheme and no safe harbour from litigation.
Faster federal adoption and procurement of AI
Federal agencies should unlock AI already built into software they use, streamline procurement of "commercially safe" AI trained only on licensed or permission-free content, and use generative AI for information processing and content creation.
The Act regulates AI placed on the market or put into service and does not set a programme for public-sector adoption or procurement; public bodies appear in it as deployers with extra duties, such as the fundamental rights impact assessment in Article 27 for certain high-risk uses.
The plan's 'Accelerate AI Adoption in Government' section formalises the Chief AI Officer Council, creates a GSA-managed AI procurement toolbox and a talent-exchange programme, and separately directs procurement rules toward LLMs that are 'free from top-down ideological bias'; it does not prefer tools trained on licensed content, which is Adobe's own addition.
Adobe's agenda splits cleanly in two, and each half meets a different kind of law. On provenance it asks for more regulation than Washington has given it: the July 2025 Action Plan confines itself to deepfakes as evidence in court, while California's AI Transparency Act (as amended by AB 853) is close to a line-by-line enactment of Adobe's asks — platforms that must preserve and show provenance data from 2027, and device makers that must offer it from 2028 — and the EU AI Act requires generative systems to mark their output in machine-readable form. On everything else it asks for less: it rejects compute-threshold, catastrophic-risk and training-data-disclosure rules, which are precisely the EU AI Act's general-purpose model regime and California's SB 53 and AB 2013, and it wants self-certification to earn "a safe harbor against future litigation", something no instrument in our corpus offers. The federal plan matches its deregulatory half but says nothing on copyright, opt-outs, style impersonation or content credentials, so the provisions Adobe cares most about were not taken up.
Source
https://files.nitrd.gov/90-fr-9088/Adobe-RFI-2025.pdf- Date on the page:
- Source checked:
- opened and confirmed on 2026-09-30