As AI is increasingly integrated into our lives and national infrastructure, The Digital Chamber (TDC) has built policy principles to ensure that these vital systems preserves innovation, resiliency, safety, and a responsible federal-state regulatory harmonization that provides clear direction and protections for relevant stakeholders in the AI ecosystem.  Specifically, the United States must refine regulations to leverage decentralized AI and AI development that supports open and democratic governance – technological remedies to many legacy AI roadblocks and concerns. 

TDC supports the following principles that we intend to use to guide our AI policy engagement: 

1. Mitigating Bias: Bias Audits Should Be Effective, Outcomes-Based, Transparent, and Public 

AI systems should be tested for fair and non-discriminatory outcomes across their full lifecycle, not certified only once at launch.  

TDC supports making bias-audit results available for regulator and public review, with the FTC and state attorneys general able to act where a consistent pattern of failed audits reveals discriminatory outputs. Bias audits should be tied to reasonable mitigation requirements, and at no point should mitigating bias be treated as the introduction thereof. Responsibility for biased outputs should generally rest with the deployer, the entity closest to the person affected, rather than falling equally on every upstream developer. The goal is a standard that measures results, rather than intent, as AI is definitionally incapable of producing ‘intent’ necessary for disparate-treatment claims. 

2. Model Safety: Safety Disclosure Should Scale with Use Case, Not Model Size 

Safety testing should be continuous and include:  

• pre-deployment evaluation 

• active monitoring,  

• an established process for addressing failing systems in real time 

TDC supports disclosure requirements for safety testing results, calibrated to expected use case, user base, and associated risks, rather than model size alone. Systems deployed in healthcare, hiring, housing, or against a vulnerable population should meet a higher bar than a narrow tool with limited exposure, regardless of scale. Risk thresholds should reflect input from the people who bear the consequences of failure, and the workforce that identifies harm(s)deserves protections of its own. 

Further, catastrophic risks have received outsized attention compared to more common violations of user rights and well-being. Both catastrophic and quotidian risks must be addressed. 

3. Liability and Accountability: Accountability Should Follow Supply-Chain Role 

Liability should track function: developers and deployers should each be accountable for harms tied to their specific role, and content-neutral infrastructure providers in a decentralized AI operating system should not be shoehorned into one of these existing categories. 

• Developers supply capability and are typically liable for defects in the underlying model (e.g., Meta is the developer of the Llama series). A party that modifies an existing model also takes on developer obligations where that modification expands the capability surface, alters the action space, or degrades safeguards; 

• Deployers refine capability and put it into use. This covers parties that fine-tune a model to narrow it within the intended purposes and documented limitations, and the entity that chooses to place a system into service. Deployers sit closest to the affected person and carry the greatest share of responsibility for outcomes in use. A deployer that narrows a model should publish its intended-use documentation and test for drift within that window; 

• End Users query or operate an AI system. Where the user and the affected person are the same, accuracy limits are disclosed, and a free and prompt remedy is available, liability should not run upstream for use outside the developer’s/deployer’s stated parameters. This limit does not apply where the affected person is not the operator, as in lending, hiring, housing, healthcare, and criminal justice. 

These categories describe function, not identity, and are not mutually exclusive. A power user self-hosting a model may be developer, deployer, and end user at once. Contractual allocation should not displace the factual distribution of control. 

Obligations also presuppose a party able to discharge them. Where a layer is permissionless and no entity controls access or output, assigning developer or deployer duties there accomplishes nothing or makes recentralization the only compliance path. Accountability should attach to parties exercising control, applied equally to concentrated and distributed infrastructure. 

This chain should also make room for a distinct category of content-neutral infrastructure providers, including decentralized compute, storage, data-availability, and validation networks that do not select or control the models running on them, who should not inherit developer or deployer liability absent knowing participation in a specific harm (see Principle 9). 

Each party should be able to explain the portion of the process it controls, and affected individuals should receive that explanation without requesting it or litigating for it. 

4. Governance and Oversight: Human Oversight Is the Floor; Direct Governance Can Exceed It 

Human oversight of consequential AI decisions – in lending, hiring, healthcare, housing, and criminal justice, among other high-risk domains – remains essential, as is government accountability. 

Where a community (like a blockchain-based AI DAO) holds governance rights, that direct governance can fulfill accountability requirements.  

Worker organizations and unions should be included as participants in setting oversight standards for AI directly impacting workers in a stakeholder model. Agentic AI should inherently carry identification metadata.  

Further, that identification should extend to verifiable tamper-evident action logs, so an autonomous agent’s decisions can be attributed to a responsible party after the fact rather than reconstructed through litigation (see Principle 13). Mechanisms that build oversight directly into execution, such as network-level alignment nodes, illustrate how decentralized systems can make governance a property of the infrastructure rather than an afterthought. 

5. Data Privacy and Sovereignty: Individuals Should Hold a Strong Claim to Their Own Data 

Meaningful, rather than purely symbolic or perfunctory, control over personal and session data is essential.  

Principled data sovereignty requires that people natively own their data. Privacy-preserving infrastructure, like decentralized identity and zero-knowledge methods, can help realize these protections.  

Regulators should treat these privacy-enhancing technologies, including trusted execution environments and confidential computing, secure multi-party computation, and federated approaches, as affirmative compliance pathways deserving of safe-harbor treatment, rather than viewing their use with suspicion. 

Reasonable limits on data brokers and concentrated data control can help enhance privacy where individual-level tools fall short. TDC also supports updated de-identification standards that account for AI-enabled re-identification risk. 

6. Redefining “Democratic AI”: Legitimacy Comes from Governance, Not Governments 

Democratic legitimacy is measured by how directly an AI stack’s power and governance is checked and how widely it is distributed, rather than by which government’s flag sits on a developer’s S1.  

TDC believes AI development should build toward open, auditable, and decentralized AI, governed by meaningful direct input from the people it affects.  Where that level of direct governance isn’t yet possible, legislative direction and generic guardrails can be a reasonable alternative. 

Widespread use of a technology is not the same as widespread control over it. A system available to everyone and governed by one company is not democratic. Democratic AI should embody the following qualities: 

• Both control and usage are distributed. Many parties can run a system, but that is not the test. What matters is whether any single entity can unilaterally change the model, revoke access, or set the terms everyone else operates under. 

• The stack is open and forkable. Weights, training data, and code are published under terms that permit modification and redistribution. ‘Open source’ is a necessary but insufficient quality for Democratic AI. 

• Governance rights are enforceable. People affected by a system can change how it operates. 

Decentralization also carries a resilience and security dividend that policymakers should weigh directly: infrastructure with no single point of control or failure reduces systemic risk and concentrated dependency. Advancing this standard should include interoperability, data and model portability, and open protocols, so that no single incumbent can capture the AI stack through lock-in. 

7. Shared Societal Benefit: Democratizing AI Should Extend to Democratizing Its Gains 

AI-driven labor disruption calls for near-term investment in reskilling and a modernized safety net, tracking impact as it happens rather than after the fact. 

TDC believes that economic gains of AI should be widely distributed through consultation with civil society and affected populations. Workers and communities displaced by AI infrastructure should have a meaningful path to share the benefits of that infrastructure, rather than simply support for the disruption it causes.  

Democratizing AI should extend beyond its oversight to who benefits from the tech, which should itself be decided democratically. Concretely, this can be delivered through verifiable on-chain attribution and micropayments to the people who contribute data, compute, and storage, and through participation in the networks those contributions help run. Such mechanisms distribute gains by design rather than only redistributing them after the fact (see Principle 12). 

8. Verifiable AI: Cryptographic Proof Should Be a Recognized Path to Compliance 

Compliance should be something a regulator can check, rather than something a developer asserts. 

Audits, disclosures, and self-certification all ultimately ask the public to trust the party being regulated. TDC supports recognizing cryptographic verification, meaning proof of what model ran, on what data, and with what result, established through methods such as verifiable inference, provenance attestation, trusted execution environments, and zero-knowledge proofs, as a valid and, where available, preferred means of demonstrating compliance. A standard built on “verify, don’t just trust” lets oversight keep pace with systems too numerous and fast-moving to certify one at a time. 

9. Neutral Infrastructure: Content-Neutral Providers Should Have a Clear Safe Harbor 

Providing the pipes should not mean answering for the water. 

The compute, storage, data-availability, and validation networks that AI runs on are typically content-neutral: they do not choose, train, or control the models they carry. TDC supports a clear safe harbor establishing that operating such infrastructure does not, by itself, make a party a developer or deployer, absent known participation in a specific harm. Without this, the decentralized and distributed infrastructure that makes open AI possible cannot operate at scale in the United States. 

10. Digital Assets: Network-Incentive Tokens Deserve Regulatory Clarity 

Decentralized AI cannot run on infrastructure that has no lawful way to reward its participants. 

Decentralized networks coordinate and compensate their compute providers, storage providers, validators, and other contributors through functional tokens. TDC supports clear federal treatment confirming that tokens used to coordinate infrastructure and reward genuine network participation are not, by default, treated as securities, along with workable clarity for staking and other consensus mechanisms. Resolving this question, now under active debate in digital-asset market-structure legislation, is a precondition for a domestic decentralized-AI industry rather than a matter separate from AI policy. 

11. Open Models: Releasing Open-Weight AI Should Be Presumptively Protected 

Openness is both a safe and competitive strategy. 

Open-weight and open-source models let independent researchers inspect, test, and improve systems that would otherwise be opaque. TDC supports treating the release of open-weight models as presumptively protected, so that a developer is not automatically liable for every downstream use of an openly released model and we are cautious about blanket reporting or restriction regimes aimed at open release. Where safety concerns arise, verifiability (Principle 8) is the better answer than closure: open and verifiable is more accountable than closed and unexamined. 

12. Data Provenance: Contributors Should Be Traceable and Compensable 

If data is the raw material of AI, its providers should be visible in the record and able to share in the value. 

Verifiable data pipelines make it possible to record what data trained a model and to trace contributions back to their source. TDC supports on-chain, privacy-respecting data provenance together with mechanisms for attribution and compensation, including micropayments, to the creators and communities whose data trains AI systems. This advances content authenticity and consent while giving the shared-benefit commitment in Principle 7 a concrete mechanism. 

13. Agentic AI: Autonomous Agents Need Identity, Accountability, and Payment Rails by Design 

As agents begin to act and transact on their own, the accountability layer should be built in up front. 

Autonomous agents increasingly initiate actions and transactions with limited real-time human involvement. TDC supports requiring that agents carry verifiable, portable identity, maintain tamper-evident logs of their actions, and remain attributable to a responsible human or entity, with any agent-to-agent transaction rails designed so that accountability is verifiable rather than reconstructed after harm. Building these guarantees into the infrastructure is the most reliable way to keep an emerging agentic economy answerable to the people it affects. 

If you have any questions, please reach out to policy@digitalchamber.org.