This agreement makes AI companies both the creators of the technology and the inspectors of their own safety, so we must ask clearly: who sets the standards, who reviews the results, and who is responsible when harm occurs?
I think the risk lies not only in the technology but also in conflicts of interest. If oversight lacks an independent body, the agreement may help businesses move faster, but society must bear the cost of mistakes that are not detected in time.
This agreement makes AI companies both the creators of the technology and the inspectors of their own safety, so we must ask clearly: who sets the standards, who reviews the results, and who is responsible when harm occurs?
I think the risk lies not only in the technology but also in conflicts of interest. If oversight lacks an independent body, the agreement may help businesses move faster, but society must bear the cost of mistakes that are not detected in time.
How Is This Agreement Changing the Way AI Safety Is Governed?
The new agreement pushes technology companies to assess risks and report their own results under a framework jointly accepted by the government and the business sector.
The advantage is that rules can be established quickly and adapted to keep pace with technology. But voluntary oversight will be credible only when there is verifiable evidence and clear accountability when systems cause harm.
How Is This Agreement Changing the Way AI Safety Is Governed?
The new agreement pushes technology companies to assess risks and report their own results under a framework jointly accepted by the government and the business sector.
The advantage is that rules can be established quickly and adapted to keep pace with technology. But voluntary oversight will be credible only when there is verifiable evidence and clear accountability when systems cause harm.
When AI Creators Have to Inspect Themselves
Imagine an organization that must use AI to make important decisions, but whose entire confidence comes from assurances by the company that developed it because no external agency has independently reviewed the system. Users are therefore required to trust the reports, testing methods, and explanations provided by the same party.
The problem is that companies face pressure to launch quickly, attract customers, and maintain their business advantage, while public responsibility requires them to conduct repeated reviews when risks are discovered. Without evidence open to scrutiny, oversight may become little more than a promise that lets companies grade themselves.
When AI Creators Have to Inspect Themselves
Imagine an organization that must use AI to make important decisions, but whose entire confidence comes from assurances by the company that developed it because no external agency has independently reviewed the system. Users are therefore required to trust the reports, testing methods, and explanations provided by the same party.
The problem is that companies face pressure to launch quickly, attract customers, and maintain their business advantage, while public responsibility requires them to conduct repeated reviews when risks are discovered. Without evidence open to scrutiny, oversight may become little more than a promise that lets companies grade themselves.
Where Does This Agreement Fit into the U.S. AI Policy Landscape?
An agreement between the government and technology companies is a shared commitment that relies more on policy and reputational pressure than on legal penalties. If companies fail to comply, this framework is therefore weaker than binding law and more flexible than an executive order issued directly by the government.
Compared with existing safety standards, an agreement of this kind may help accelerate common practices, but we still need to see whether evidence is disclosed and reviewed by outsiders. Leading companies consequently occupy overlapping roles as technology developers, standard setters, and assessors of their own systems’ risks. This makes credibility depend on transparency rather than declarations alone.
Where Does This Agreement Fit into the U.S. AI Policy Landscape?
An agreement between the government and technology companies is a shared commitment that relies more on policy and reputational pressure than on legal penalties. If companies fail to comply, this framework is therefore weaker than binding law and more flexible than an executive order issued directly by the government.
Compared with existing safety standards, an agreement of this kind may help accelerate common practices, but we still need to see whether evidence is disclosed and reviewed by outsiders. Leading companies consequently occupy overlapping roles as technology developers, standard setters, and assessors of their own systems’ risks. This makes credibility depend on transparency rather than declarations alone.
From Voluntary Commitments to Truly Binding Rules?
Voluntary commitments are credible only when outsiders can inspect the evidence and there are consequences if companies fail to comply. A new approach in which companies set their own standards must therefore answer clearly who has the authority to decide and whom victims can contact to file complaints.
| Factor | Existing regulations or commitments | Companies set their own standards |
|---|---|---|
| Authority to inspect | Agencies or outsiders | The company or a group of companies |
| Transparency | A clearly disclosed framework | Depends on the company’s disclosure |
| Penalties | Consequences under the rules | May be limited to internal measures |
| Remedies | Based on mechanisms under the rules | Must rely on the company’s processes |
Without independent inspectors and clear remedies, self-enforcement remains closer to an existing commitment than to truly binding rules.
From Voluntary Commitments to Truly Binding Rules?
Voluntary commitments are credible only when outsiders can inspect the evidence and there are consequences if companies fail to comply. A new approach in which companies set their own standards must therefore answer clearly who has the authority to decide and whom victims can contact to file complaints.
| Factor | Existing regulations or commitments | Companies set their own standards |
|---|---|---|
| Authority to inspect | Agencies or outsiders | The company or a group of companies |
| Transparency | A clearly disclosed framework | Depends on the company’s disclosure |
| Penalties | Consequences under the rules | May be limited to internal measures |
| Remedies | Based on mechanisms under the rules | Must rely on the company’s processes |
Without independent inspectors and clear remedies, self-enforcement remains closer to an existing commitment than to truly binding rules.
How Well Will Safety Promises Work in Real-World Situations?
Pre-launch testing must take place whenever a model performs beyond what was anticipated, much like checking an iPhone 17 Pro Max with a 3nm chip, 12GB of RAM, and a 120Hz display for both performance and risks before putting it on sale.
When vulnerabilities are found, companies should report them immediately and share incident information with other platforms because AI threats do not stop at a single company. When risks increase faster than control measures can keep up, launches must be delayed or restricted rather than leaving users to bear the consequences themselves.
How Well Will Safety Promises Work in Real-World Situations?
Pre-launch testing must take place whenever a model performs beyond what was anticipated, much like checking an iPhone 17 Pro Max with a 3nm chip, 12GB of RAM, and a 120Hz display for both performance and risks before putting it on sale.
When vulnerabilities are found, companies should report them immediately and share incident information with other platforms because AI threats do not stop at a single company. When risks increase faster than control measures can keep up, launches must be delayed or restricted rather than leaving users to bear the consequences themselves.
How Does U.S. Self-Governance of AI Compare with Other Options?
The U.S. voluntary agreement stands out for its ability to adapt quickly, but accountability and inspection authority still depend largely on companies themselves. It is therefore better suited to rapidly changing risks than to rigid forms of control.
| Factor | U.S. voluntary agreement | European Union AI Act | U.K. voluntary approach | U.S. risk management framework |
|---|---|---|---|---|
| Strictness | Moderate | High | Low | Moderate |
| Speed of adaptation | High | Low | High | High |
| Legal accountability | Low | High | Low | Low |
| Authority of independent agencies | Low | High | Moderate | Low |
The European option is stricter and has clearer enforcement authority, while the U.S. approach is more flexible. But it must disclose inspection criteria and remediation results in a verifiable way to avoid becoming merely a promise.
How Does U.S. Self-Governance of AI Compare with Other Options?
The U.S. voluntary agreement stands out for its ability to adapt quickly, but accountability and inspection authority still depend largely on companies themselves. It is therefore better suited to rapidly changing risks than to rigid forms of control.
| Factor | U.S. voluntary agreement | European Union AI Act | U.K. voluntary approach | U.S. risk management framework |
|---|---|---|---|---|
| Strictness | Moderate | High | Low | Moderate |
| Speed of adaptation | High | Low | High | High |
| Legal accountability | Low | High | Low | Low |
| Authority of independent agencies | Low | High | Moderate | Low |
The European option is stricter and has clearer enforcement authority, while the U.S. approach is more flexible. But it must disclose inspection criteria and remediation results in a verifiable way to avoid becoming merely a promise.
Strengths and Vulnerabilities of Letting Companies Govern Themselves
Allowing companies to govern themselves helps them respond quickly to risks and draw on the expertise of the teams that directly develop AI. Standards can therefore be adjusted more flexibly than rules that fail to keep pace with technology.
But conflicts of interest remain a vulnerability. Companies may disclose only the information they consider necessary, standards may differ across companies, and without clear penalties, oversight may stop at the level of promises.
Pros
- +Respond quickly to risks
- +Use expertise from developers directly
- +Adjust standards more flexibly than outdated rules
Cons
- −Risk of conflicts of interest
- −Uneven standards across companies
- −Disclosure may be limited
- −Lack of clear penalties
Strengths and Vulnerabilities of Letting Companies Govern Themselves
Allowing companies to govern themselves helps them respond quickly to risks and draw on the expertise of the teams that directly develop AI. Standards can therefore be adjusted more flexibly than rules that fail to keep pace with technology.
But conflicts of interest remain a vulnerability. Companies may disclose only the information they consider necessary, standards may differ across companies, and without clear penalties, oversight may stop at the level of promises.
Pros
- +Respond quickly to risks
- +Use expertise from developers directly
- +Adjust standards more flexibly than outdated rules
Cons
- −Risk of conflicts of interest
- −Uneven standards across companies
- −Disclosure may be limited
- −Lack of clear penalties
The Price Society May Have to Pay Even Without Higher Service Fees
Voluntary commitments may appear lighter than regulations, but the real cost falls on those affected. Victims must prove for themselves that AI caused the harm, while companies determine the inspection methods and may report incidents late.
Large companies have more legal teams and resources than smaller players, creating unequal standards of accountability. If commitments are insufficient, the public must bear the burden of filing complaints, following up, and protecting themselves.
Pros
- +Create a framework for companies to begin taking responsibility
- +Reduce part of the government’s oversight burden
- +Allow companies to resolve problems more quickly
Cons
- −Victims must prove the harm themselves
- −Inspection is not independent
- −Large companies have an advantage over smaller players
- −Incident reporting may be delayed
- −The public bears the burden when commitments are insufficient
The Price Society May Have to Pay Even Without Higher Service Fees
Voluntary commitments may appear lighter than regulations, but the real cost falls on those affected. Victims must prove for themselves that AI caused the harm, while companies determine the inspection methods and may report incidents late.
Large companies have more legal teams and resources than smaller players, creating unequal standards of accountability. If commitments are insufficient, the public must bear the burden of filing complaints, following up, and protecting themselves.
Pros
- +Create a framework for companies to begin taking responsibility
- +Reduce part of the government’s oversight burden
- +Allow companies to resolve problems more quickly
Cons
- −Victims must prove the harm themselves
- −Inspection is not independent
- −Large companies have an advantage over smaller players
- −Incident reporting may be delayed
- −The public bears the burden when commitments are insufficient
Questions This Agreement Still Cannot Answer
Who will verify that companies are actually honoring their commitments, and how independent are the inspectors from the companies? If a problem occurs, who will decide that a model is too dangerous and must be taken offline or modified?
Another issue is the extent to which safety information will be disclosed to the public. Companies may invoke trade secrets as a reason not to reveal details.
Finally, will the government intervene when companies breach the agreement, and must it wait until harm occurs? Without clear answers, this agreement still relies more on voluntary participation than on genuinely verifiable protection.
Questions This Agreement Still Cannot Answer
Who will verify that companies are actually honoring their commitments, and how independent are the inspectors from the companies? If a problem occurs, who will decide that a model is too dangerous and must be taken offline or modified?
Another issue is the extent to which safety information will be disclosed to the public. Companies may invoke trade secrets as a reason not to reveal details.
Finally, will the government intervene when companies breach the agreement, and must it wait until harm occurs? Without clear answers, this agreement still relies more on voluntary participation than on genuinely verifiable protection.
The Real Test Is Not the Statement but the Day an Incident Occurs
The credibility of the agreement will be measured on the day a company must choose an option that costs it revenue or a competitive advantage—not on the day when everyone still benefits from an attractive declaration.
Going forward, we should track clear evidence, including testing reports, reported incidents, and penalties when standards are violated. Without these things, self-governance remains merely a commitment that is difficult to verify.
The Real Test Is Not the Statement but the Day an Incident Occurs
The credibility of the agreement will be measured on the day a company must choose an option that costs it revenue or a competitive advantage—not on the day when everyone still benefits from an attractive declaration.
Going forward, we should track clear evidence, including testing reports, reported incidents, and penalties when standards are violated. Without these things, self-governance remains merely a commitment that is difficult to verify.
How Is This Agreement Changing the Way AI Safety Is Governed?
The new agreement is shifting AI governance away from rules imposed by the state and toward voluntary oversight by technology companies. Each company must announce its own testing standards and approaches to handling risks.
The key issue is therefore not merely the commitment but verifiable evidence, such as testing reports, incident notifications, and remediation when systems encounter problems. If the government still lacks clear monitoring authority, this agreement may help establish common standards, but it is not the same as actual enforcement.
How Is This Agreement Changing the Way AI Safety Is Governed?
The new agreement is shifting AI governance away from rules imposed by the state and toward voluntary oversight by technology companies. Each company must announce its own testing standards and approaches to handling risks.
The key issue is therefore not merely the commitment but verifiable evidence, such as testing reports, incident notifications, and remediation when systems encounter problems. If the government still lacks clear monitoring authority, this agreement may help establish common standards, but it is not the same as actual enforcement.
When AI Creators Have to Inspect Themselves
Imagine an organization preparing to use AI to screen applications or assist with important decisions. Executives must trust assurances from the developer that the system is safe, even though no external agency has independently inspected it.
The risk is that the company must choose between disclosing problems to the public and preserving customer confidence and revenue. A voluntary agreement may help companies establish rules and report incidents more effectively, but when the creators are also the inspectors, transparency still depends on the company’s incentives.
When AI Creators Have to Inspect Themselves
Imagine an organization preparing to use AI to screen applications or assist with important decisions. Executives must trust assurances from the developer that the system is safe, even though no external agency has independently inspected it.
The risk is that the company must choose between disclosing problems to the public and preserving customer confidence and revenue. A voluntary agreement may help companies establish rules and report incidents more effectively, but when the creators are also the inspectors, transparency still depends on the company’s incentives.
Where Does This Agreement Fit into the U.S. AI Policy Landscape?
An agreement between the government and technology companies is a voluntary collaboration. It differs from laws that apply to every company and from executive orders through which the government directly sets policy. Existing safety standards often serve as reference frameworks, while this agreement is more closely tied to companies’ commitments and reporting.
Leading companies therefore occupy three overlapping roles. They are AI developers, co-creators of standards, and assessors of their own systems’ risks. The advantage is that rules can be adjusted quickly, but the weakness is that inspection authority remains close to the people who created the technology rather than being clearly separated from them.
Where Does This Agreement Fit into the U.S. AI Policy Landscape?
An agreement between the government and technology companies is a voluntary collaboration. It differs from laws that apply to every company and from executive orders through which the government directly sets policy. Existing safety standards often serve as reference frameworks, while this agreement is more closely tied to companies’ commitments and reporting.
Leading companies therefore occupy three overlapping roles. They are AI developers, co-creators of standards, and assessors of their own systems’ risks. The advantage is that rules can be adjusted quickly, but the weakness is that inspection authority remains close to the people who created the technology rather than being clearly separated from them.
From Voluntary Commitments to Truly Binding Rules?
Voluntary commitments allow companies to move quickly, but credibility still depends on disclosure and external inspection. An approach in which companies enforce standards among themselves may be more flexible, but it must answer who has the authority to challenge them and through which channel victims can seek help.
| Factor | Existing regulations or commitments | Companies set standards among themselves |
|---|---|---|
| Authority to inspect | An agency or external signatories | The company and its working groups |
| Transparency | A verifiable reporting framework | Depends on the information disclosed by the company |
| Penalties | Consequences under the rules or agreement | A clear mechanism must be established |
| Remedies | Based on legal procedures | The responsible party still needs to be identified |
From Voluntary Commitments to Truly Binding Rules?
Voluntary commitments allow companies to move quickly, but credibility still depends on disclosure and external inspection. An approach in which companies enforce standards among themselves may be more flexible, but it must answer who has the authority to challenge them and through which channel victims can seek help.
| Factor | Existing regulations or commitments | Companies set standards among themselves |
|---|---|---|
| Authority to inspect | An agency or external signatories | The company and its working groups |
| Transparency | A verifiable reporting framework | Depends on the information disclosed by the company |
| Penalties | Consequences under the rules or agreement | A clear mechanism must be established |
| Remedies | Based on legal procedures | The responsible party still needs to be identified |
How Well Will Safety Promises Work in Real-World Situations?
Pre-launch testing should work like checking how well an iPhone 17 Pro Max with an Apple A19 Pro chip (3 nm) and 12GB of RAM can handle demanding tasks before allowing people to use it in the real world. If unexpected behavior is found, the launch must be delayed or capabilities restricted first.
When vulnerabilities are found, companies should report them with verifiable details rather than waiting until the problem affects large numbers of users. As with devices available with 256GB, 512GB, or 2TB of storage, limitations must be clearly specified from the beginning.
If a threat affects multiple platforms, companies should share incident information with one another. Oversight will be credible only when it includes evidence, alerts, and workable measures for shutting down systems.
How Well Will Safety Promises Work in Real-World Situations?
Pre-launch testing should work like checking how well an iPhone 17 Pro Max with an Apple A19 Pro chip (3 nm) and 12GB of RAM can handle demanding tasks before allowing people to use it in the real world. If unexpected behavior is found, the launch must be delayed or capabilities restricted first.
When vulnerabilities are found, companies should report them with verifiable details rather than waiting until the problem affects large numbers of users. As with devices available with 256GB, 512GB, or 2TB of storage, limitations must be clearly specified from the beginning.
If a threat affects multiple platforms, companies should share incident information with one another. Oversight will be credible only when it includes evidence, alerts, and workable measures for shutting down systems.
How Does U.S. Self-Governance of AI Compare with Other Options?
Voluntary agreements can adapt quickly because companies determine their own responses, but legal accountability and inspection authority from independent agencies remain limited. This differs from the European Union’s AI Act, which has clearer requirements. The United Kingdom and the U.S. risk management framework are more flexible but rely primarily on companies’ cooperation.
| Factor | U.S. voluntary agreement | European Union AI Act | U.K. voluntary approach | U.S. risk management framework |
|---|---|---|---|---|
| Strictness | Low | High | Moderate | Moderate |
| Speed of adaptation | Fast | Slow | Fast | Fast |
| Legal accountability | Limited | Clear | Depends on existing law | Limited |
| Authority of independent agencies | Limited | High | Limited | Limited |
The trade-off is clear: the U.S. gains flexibility, but credibility depends on companies’ disclosure and compliance with their commitments.
How Does U.S. Self-Governance of AI Compare with Other Options?
Voluntary agreements can adapt quickly because companies determine their own responses, but legal accountability and inspection authority from independent agencies remain limited. This differs from the European Union’s AI Act, which has clearer requirements. The United Kingdom and the U.S. risk management framework are more flexible but rely primarily on companies’ cooperation.
| Factor | U.S. voluntary agreement | European Union AI Act | U.K. voluntary approach | U.S. risk management framework |
|---|---|---|---|---|
| Strictness | Low | High | Moderate | Moderate |
| Speed of adaptation | Fast | Slow | Fast | Fast |
| Legal accountability | Limited | Clear | Depends on existing law | Limited |
| Authority of independent agencies | Limited | High | Limited | Limited |
The trade-off is clear: the U.S. gains flexibility, but credibility depends on companies’ disclosure and compliance with their commitments.
Strengths and Vulnerabilities of Letting Companies Govern Themselves
Allowing companies to govern themselves helps them respond quickly to risks and draw on knowledge from the teams that directly develop AI. Rules can also keep pace with technology more effectively than laws that take a long time to enact.
But credibility is weakened when companies have to inspect themselves because conflicts of interest may arise, standards may vary between organizations, and insufficient information may be disclosed. Without clear penalties, the agreement may amount to nothing more than a promise.
Pros
- +Respond quickly to risks
- +Use expertise from developers directly
- +Reduce the burden of rules that cannot keep pace with technology
Cons
- −Risk of conflicts of interest
- −Uneven safety standards
- −Limited disclosure
- −Lack of clear penalties
Strengths and Vulnerabilities of Letting Companies Govern Themselves
Allowing companies to govern themselves helps them respond quickly to risks and draw on knowledge from the teams that directly develop AI. Rules can also keep pace with technology more effectively than laws that take a long time to enact.
But credibility is weakened when companies have to inspect themselves because conflicts of interest may arise, standards may vary between organizations, and insufficient information may be disclosed. Without clear penalties, the agreement may amount to nothing more than a promise.
Pros
- +Respond quickly to risks
- +Use expertise from developers directly
- +Reduce the burden of rules that cannot keep pace with technology
Cons
- −Risk of conflicts of interest
- −Uneven safety standards
- −Limited disclosure
- −Lack of clear penalties
The Price Society May Have to Pay Even Without Higher Service Fees
A voluntary system may appear economical, but when harm occurs, victims may have to gather evidence and prove the case themselves. At the same time, inspections conducted by the same company may not be sufficiently independent.
Large companies have stronger legal teams and more resources, while smaller players may struggle to meet the standards. If incident reporting is delayed or commitments are incomplete, the ultimate burden may fall on the public.
Pros
- +Provide a framework for companies to take responsibility
- +Allow problems to be resolved more quickly
Cons
- −Victims must prove the harm themselves
- −Inspections may not be independent
- −Large companies have an advantage over smaller players
- −Incident reporting may be delayed
- −The public bears the burden when commitments are insufficient
The Price Society May Have to Pay Even Without Higher Service Fees
A voluntary system may appear economical, but when harm occurs, victims may have to gather evidence and prove the case themselves. At the same time, inspections conducted by the same company may not be sufficiently independent.
Large companies have stronger legal teams and more resources, while smaller players may struggle to meet the standards. If incident reporting is delayed or commitments are incomplete, the ultimate burden may fall on the public.
Pros
- +Provide a framework for companies to take responsibility
- +Allow problems to be resolved more quickly
Cons
- −Victims must prove the harm themselves
- −Inspections may not be independent
- −Large companies have an advantage over smaller players
- −Incident reporting may be delayed
- −The public bears the burden when commitments are insufficient
Questions This Agreement Still Cannot Answer
Who will verify that companies are actually honoring their commitments, and how independent are the inspectors from the companies? If companies refuse to disclose information, how can outsiders conduct inspections?
Another issue is who decides that a model is too dangerous and what criteria are used. How much safety information will be disclosed to the public without compromising trade secrets?
If a company breaches the agreement, will the government intervene, and how clear will the penalties be? The answers will determine whether self-governance carries real weight or is merely a commitment that is difficult to verify.
Questions This Agreement Still Cannot Answer
Who will verify that companies are actually honoring their commitments, and how independent are the inspectors from the companies? If companies refuse to disclose information, how can outsiders conduct inspections?
Another issue is who decides that a model is too dangerous and what criteria are used. How much safety information will be disclosed to the public without compromising trade secrets?
If a company breaches the agreement, will the government intervene, and how clear will the penalties be? The answers will determine whether self-governance carries real weight or is merely a commitment that is difficult to verify.
The Real Test Is Not the Statement but the Day an Incident Occurs
The credibility of the agreement will become clear on the day a company must choose an option that costs it revenue or a competitive advantage—not on the day when everyone still benefits from a joint declaration.
Going forward, three types of evidence should be tracked: testing reports, reported incidents, and penalties when standards are violated. These will reveal whether the agreement has real effect or is merely a commitment that is difficult to verify.
The Real Test Is Not the Statement but the Day an Incident Occurs
The credibility of the agreement will become clear on the day a company must choose an option that costs it revenue or a competitive advantage—not on the day when everyone still benefits from a joint declaration.
Going forward, three types of evidence should be tracked: testing reports, reported incidents, and penalties when standards are violated. These will reveal whether the agreement has real effect or is merely a commitment that is difficult to verify. This agreement makes AI companies both the creators of the technology and the inspectors of their own safety, so we must ask clearly: who sets the standards, who reviews the results, and who is responsible when harm occurs?
I think the risk lies not only in the technology but also in conflicts of interest. If oversight lacks an independent body, the agreement may help businesses move faster, but society must bear the cost of mistakes that are not detected in time.
This agreement makes AI companies both the creators of the technology and the inspectors of their own safety, so we must ask clearly: who sets the standards, who reviews the results, and who is responsible when harm occurs?
I think the risk lies not only in the technology but also in conflicts of interest. If oversight lacks an independent body, the agreement may help businesses move faster, but society must bear the cost of mistakes that are not detected in time.
How Is This Agreement Changing the Way AI Safety Is Governed?
The new agreement pushes technology companies to assess risks and report their own results under a framework jointly accepted by the government and the business sector.
The advantage is that rules can be established quickly and adapted to keep pace with technology. But voluntary oversight will be credible only when there is verifiable evidence and clear accountability when systems cause harm.
How Is This Agreement Changing the Way AI Safety Is Governed?
The new agreement pushes technology companies to assess risks and report their own results under a framework jointly accepted by the government and the business sector.
The advantage is that rules can be established quickly and adapted to keep pace with technology. But voluntary oversight will be credible only when there is verifiable evidence and clear accountability when systems cause harm.
When AI Creators Have to Inspect Themselves
Imagine an organization that must use AI to make important decisions, but whose entire confidence comes from assurances by the company that developed it because no external agency has independently reviewed the system. Users are therefore required to trust the reports, testing methods, and explanations provided by the same party.
The problem is that companies face pressure to launch quickly, attract customers, and maintain their business advantage, while public responsibility requires them to conduct repeated reviews when risks are discovered. Without evidence open to scrutiny, oversight may become little more than a promise that lets companies grade themselves.
When AI Creators Have to Inspect Themselves
Imagine an organization that must use AI to make important decisions, but whose entire confidence comes from assurances by the company that developed it because no external agency has independently reviewed the system. Users are therefore required to trust the reports, testing methods, and explanations provided by the same party.
The problem is that companies face pressure to launch quickly, attract customers, and maintain their business advantage, while public responsibility requires them to conduct repeated reviews when risks are discovered. Without evidence open to scrutiny, oversight may become little more than a promise that lets companies grade themselves.
Where Does This Agreement Fit into the U.S. AI Policy Landscape?
An agreement between the government and technology companies is a shared commitment that relies more on policy and reputational pressure than on legal penalties. If companies fail to comply, this framework is therefore weaker than binding law and more flexible than an executive order issued directly by the government.
Compared with existing safety standards, an agreement of this kind may help accelerate common practices, but we still need to see whether evidence is disclosed and reviewed by outsiders. Leading companies consequently occupy overlapping roles as technology developers, standard setters, and assessors of their own systems’ risks. This makes credibility depend on transparency rather than declarations alone.
Where Does This Agreement Fit into the U.S. AI Policy Landscape?
An agreement between the government and technology companies is a shared commitment that relies more on policy and reputational pressure than on legal penalties. If companies fail to comply, this framework is therefore weaker than binding law and more flexible than an executive order issued directly by the government.
Compared with existing safety standards, an agreement of this kind may help accelerate common practices, but we still need to see whether evidence is disclosed and reviewed by outsiders. Leading companies consequently occupy overlapping roles as technology developers, standard setters, and assessors of their own systems’ risks. This makes credibility depend on transparency rather than declarations alone.
From Voluntary Commitments to Truly Binding Rules?
Voluntary commitments are credible only when outsiders can inspect the evidence and there are consequences if companies fail to comply. A new approach in which companies set their own standards must therefore answer clearly who has the authority to decide and whom victims can contact to file complaints.
| Factor | Existing regulations or commitments | Companies set their own standards |
|---|---|---|
| Authority to inspect | Agencies or outsiders | The company or a group of companies |
| Transparency | A clearly disclosed framework | Depends on the company’s disclosure |
| Penalties | Consequences under the rules | May be limited to internal measures |
| Remedies | Based on mechanisms under the rules | Must rely on the company’s processes |
Without independent inspectors and clear remedies, self-enforcement remains closer to an existing commitment than to truly binding rules.
From Voluntary Commitments to Truly Binding Rules?
Voluntary commitments are credible only when outsiders can inspect the evidence and there are consequences if companies fail to comply. A new approach in which companies set their own standards must therefore answer clearly who has the authority to decide and whom victims can contact to file complaints.
| Factor | Existing regulations or commitments | Companies set their own standards |
|---|---|---|
| Authority to inspect | Agencies or outsiders | The company or a group of companies |
| Transparency | A clearly disclosed framework | Depends on the company’s disclosure |
| Penalties | Consequences under the rules | May be limited to internal measures |
| Remedies | Based on mechanisms under the rules | Must rely on the company’s processes |
Without independent inspectors and clear remedies, self-enforcement remains closer to an existing commitment than to truly binding rules.
How Well Will Safety Promises Work in Real-World Situations?
Pre-launch testing must take place whenever a model performs beyond what was anticipated, much like checking an iPhone 17 Pro Max with a 3nm chip, 12GB of RAM, and a 120Hz display for both performance and risks before putting it on sale.
When vulnerabilities are found, companies should report them immediately and share incident information with other platforms because AI threats do not stop at a single company. When risks increase faster than control measures can keep up, launches must be delayed or restricted rather than leaving users to bear the consequences themselves.
How Well Will Safety Promises Work in Real-World Situations?
Pre-launch testing must take place whenever a model performs beyond what was anticipated, much like checking an iPhone 17 Pro Max with a 3nm chip, 12GB of RAM, and a 120Hz display for both performance and risks before putting it on sale.
When vulnerabilities are found, companies should report them immediately and share incident information with other platforms because AI threats do not stop at a single company. When risks increase faster than control measures can keep up, launches must be delayed or restricted rather than leaving users to bear the consequences themselves.
How Does U.S. Self-Governance of AI Compare with Other Options?
The U.S. voluntary agreement stands out for its ability to adapt quickly, but accountability and inspection authority still depend largely on companies themselves. It is therefore better suited to rapidly changing risks than to rigid forms of control.
| Factor | U.S. voluntary agreement | European Union AI Act | U.K. voluntary approach | U.S. risk management framework |
|---|---|---|---|---|
| Strictness | Moderate | High | Low | Moderate |
| Speed of adaptation | High | Low | High | High |
| Legal accountability | Low | High | Low | Low |
| Authority of independent agencies | Low | High | Moderate | Low |
The European option is stricter and has clearer enforcement authority, while the U.S. approach is more flexible. But it must disclose inspection criteria and remediation results in a verifiable way to avoid becoming merely a promise.
How Does U.S. Self-Governance of AI Compare with Other Options?
The U.S. voluntary agreement stands out for its ability to adapt quickly, but accountability and inspection authority still depend largely on companies themselves. It is therefore better suited to rapidly changing risks than to rigid forms of control.
| Factor | U.S. voluntary agreement | European Union AI Act | U.K. voluntary approach | U.S. risk management framework |
|---|---|---|---|---|
| Strictness | Moderate | High | Low | Moderate |
| Speed of adaptation | High | Low | High | High |
| Legal accountability | Low | High | Low | Low |
| Authority of independent agencies | Low | High | Moderate | Low |
The European option is stricter and has clearer enforcement authority, while the U.S. approach is more flexible. But it must disclose inspection criteria and remediation results in a verifiable way to avoid becoming merely a promise.
Strengths and Vulnerabilities of Letting Companies Govern Themselves
Allowing companies to govern themselves helps them respond quickly to risks and draw on the expertise of the teams that directly develop AI. Standards can therefore be adjusted more flexibly than rules that fail to keep pace with technology.
But conflicts of interest remain a vulnerability. Companies may disclose only the information they consider necessary, standards may differ across companies, and without clear penalties, oversight may stop at the level of promises.
Pros
- +Respond quickly to risks
- +Use expertise from developers directly
- +Adjust standards more flexibly than outdated rules
Cons
- −Risk of conflicts of interest
- −Uneven standards across companies
- −Disclosure may be limited
- −Lack of clear penalties
Strengths and Vulnerabilities of Letting Companies Govern Themselves
Allowing companies to govern themselves helps them respond quickly to risks and draw on the expertise of the teams that directly develop AI. Standards can therefore be adjusted more flexibly than rules that fail to keep pace with technology.
But conflicts of interest remain a vulnerability. Companies may disclose only the information they consider necessary, standards may differ across companies, and without clear penalties, oversight may stop at the level of promises.
Pros
- +Respond quickly to risks
- +Use expertise from developers directly
- +Adjust standards more flexibly than outdated rules
Cons
- −Risk of conflicts of interest
- −Uneven standards across companies
- −Disclosure may be limited
- −Lack of clear penalties
The Price Society May Have to Pay Even Without Higher Service Fees
Voluntary commitments may appear lighter than regulations, but the real cost falls on those affected. Victims must prove for themselves that AI caused the harm, while companies determine the inspection methods and may report incidents late.
Large companies have more legal teams and resources than smaller players, creating unequal standards of accountability. If commitments are insufficient, the public must bear the burden of filing complaints, following up, and protecting themselves.
Pros
- +Create a framework for companies to begin taking responsibility
- +Reduce part of the government’s oversight burden
- +Allow companies to resolve problems more quickly
Cons
- −Victims must prove the harm themselves
- −Inspection is not independent
- −Large companies have an advantage over smaller players
- −Incident reporting may be delayed
- −The public bears the burden when commitments are insufficient
The Price Society May Have to Pay Even Without Higher Service Fees
Voluntary commitments may appear lighter than regulations, but the real cost falls on those affected. Victims must prove for themselves that AI caused the harm, while companies determine the inspection methods and may report incidents late.
Large companies have more legal teams and resources than smaller players, creating unequal standards of accountability. If commitments are insufficient, the public must bear the burden of filing complaints, following up, and protecting themselves.
Pros
- +Create a framework for companies to begin taking responsibility
- +Reduce part of the government’s oversight burden
- +Allow companies to resolve problems more quickly
Cons
- −Victims must prove the harm themselves
- −Inspection is not independent
- −Large companies have an advantage over smaller players
- −Incident reporting may be delayed
- −The public bears the burden when commitments are insufficient
Questions This Agreement Still Cannot Answer
Who will verify that companies are actually honoring their commitments, and how independent are the inspectors from the companies? If a problem occurs, who will decide that a model is too dangerous and must be taken offline or modified?
Another issue is the extent to which safety information will be disclosed to the public. Companies may invoke trade secrets as a reason not to reveal details.
Finally, will the government intervene when companies breach the agreement, and must it wait until harm occurs? Without clear answers, this agreement still relies more on voluntary participation than on genuinely verifiable protection.
Questions This Agreement Still Cannot Answer
Who will verify that companies are actually honoring their commitments, and how independent are the inspectors from the companies? If a problem occurs, who will decide that a model is too dangerous and must be taken offline or modified?
Another issue is the extent to which safety information will be disclosed to the public. Companies may invoke trade secrets as a reason not to reveal details.
Finally, will the government intervene when companies breach the agreement, and must it wait until harm occurs? Without clear answers, this agreement still relies more on voluntary participation than on genuinely verifiable protection.
The Real Test Is Not the Statement but the Day an Incident Occurs
The credibility of the agreement will be measured on the day a company must choose an option that costs it revenue or a competitive advantage—not on the day when everyone still benefits from an attractive declaration.
Going forward, we should track clear evidence, including testing reports, reported incidents, and penalties when standards are violated. Without these things, self-governance remains merely a commitment that is difficult to verify.
The Real Test Is Not the Statement but the Day an Incident Occurs
The credibility of the agreement will be measured on the day a company must choose an option that costs it revenue or a competitive advantage—not on the day when everyone still benefits from an attractive declaration.
Going forward, we should track clear evidence, including testing reports, reported incidents, and penalties when standards are violated. Without these things, self-governance remains merely a commitment that is difficult to verify.
How Is This Agreement Changing the Way AI Safety Is Governed?
The new agreement is shifting AI governance away from rules imposed by the state and toward voluntary oversight by technology companies. Each company must announce its own testing standards and approaches to handling risks.
The key issue is therefore not merely the commitment but verifiable evidence, such as testing reports, incident notifications, and remediation when systems encounter problems. If the government still lacks clear monitoring authority, this agreement may help establish common standards, but it is not the same as actual enforcement.
How Is This Agreement Changing the Way AI Safety Is Governed?
The new agreement is shifting AI governance away from rules imposed by the state and toward voluntary oversight by technology companies. Each company must announce its own testing standards and approaches to handling risks.
The key issue is therefore not merely the commitment but verifiable evidence, such as testing reports, incident notifications, and remediation when systems encounter problems. If the government still lacks clear monitoring authority, this agreement may help establish common standards, but it is not the same as actual enforcement.
When AI Creators Have to Inspect Themselves
Imagine an organization preparing to use AI to screen applications or assist with important decisions. Executives must trust assurances from the developer that the system is safe, even though no external agency has independently inspected it.
The risk is that the company must choose between disclosing problems to the public and preserving customer confidence and revenue. A voluntary agreement may help companies establish rules and report incidents more effectively, but when the creators are also the inspectors, transparency still depends on the company’s incentives.
When AI Creators Have to Inspect Themselves
Imagine an organization preparing to use AI to screen applications or assist with important decisions. Executives must trust assurances from the developer that the system is safe, even though no external agency has independently inspected it.
The risk is that the company must choose between disclosing problems to the public and preserving customer confidence and revenue. A voluntary agreement may help companies establish rules and report incidents more effectively, but when the creators are also the inspectors, transparency still depends on the company’s incentives.
Where Does This Agreement Fit into the U.S. AI Policy Landscape?
An agreement between the government and technology companies is a voluntary collaboration. It differs from laws that apply to every company and from executive orders through which the government directly sets policy. Existing safety standards often serve as reference frameworks, while this agreement is more closely tied to companies’ commitments and reporting.
Leading companies therefore occupy three overlapping roles. They are AI developers, co-creators of standards, and assessors of their own systems’ risks. The advantage is that rules can be adjusted quickly, but the weakness is that inspection authority remains close to the people who created the technology rather than being clearly separated from them.
Where Does This Agreement Fit into the U.S. AI Policy Landscape?
An agreement between the government and technology companies is a voluntary collaboration. It differs from laws that apply to every company and from executive orders through which the government directly sets policy. Existing safety standards often serve as reference frameworks, while this agreement is more closely tied to companies’ commitments and reporting.
Leading companies therefore occupy three overlapping roles. They are AI developers, co-creators of standards, and assessors of their own systems’ risks. The advantage is that rules can be adjusted quickly, but the weakness is that inspection authority remains close to the people who created the technology rather than being clearly separated from them.
From Voluntary Commitments to Truly Binding Rules?
Voluntary commitments allow companies to move quickly, but credibility still depends on disclosure and external inspection. An approach in which companies enforce standards among themselves may be more flexible, but it must answer who has the authority to challenge them and through which channel victims can seek help.
| Factor | Existing regulations or commitments | Companies set standards among themselves |
|---|---|---|
| Authority to inspect | An agency or external signatories | The company and its working groups |
| Transparency | A verifiable reporting framework | Depends on the information disclosed by the company |
| Penalties | Consequences under the rules or agreement | A clear mechanism must be established |
| Remedies | Based on legal procedures | The responsible party still needs to be identified |
From Voluntary Commitments to Truly Binding Rules?
Voluntary commitments allow companies to move quickly, but credibility still depends on disclosure and external inspection. An approach in which companies enforce standards among themselves may be more flexible, but it must answer who has the authority to challenge them and through which channel victims can seek help.
| Factor | Existing regulations or commitments | Companies set standards among themselves |
|---|---|---|
| Authority to inspect | An agency or external signatories | The company and its working groups |
| Transparency | A verifiable reporting framework | Depends on the information disclosed by the company |
| Penalties | Consequences under the rules or agreement | A clear mechanism must be established |
| Remedies | Based on legal procedures | The responsible party still needs to be identified |
How Well Will Safety Promises Work in Real-World Situations?
Pre-launch testing should work like checking how well an iPhone 17 Pro Max with an Apple A19 Pro chip (3 nm) and 12GB of RAM can handle demanding tasks before allowing people to use it in the real world. If unexpected behavior is found, the launch must be delayed or capabilities restricted first.
When vulnerabilities are found, companies should report them with verifiable details rather than waiting until the problem affects large numbers of users. As with devices available with 256GB, 512GB, or 2TB of storage, limitations must be clearly specified from the beginning.
If a threat affects multiple platforms, companies should share incident information with one another. Oversight will be credible only when it includes evidence, alerts, and workable measures for shutting down systems.
How Well Will Safety Promises Work in Real-World Situations?
Pre-launch testing should work like checking how well an iPhone 17 Pro Max with an Apple A19 Pro chip (3 nm) and 12GB of RAM can handle demanding tasks before allowing people to use it in the real world. If unexpected behavior is found, the launch must be delayed or capabilities restricted first.
When vulnerabilities are found, companies should report them with verifiable details rather than waiting until the problem affects large numbers of users. As with devices available with 256GB, 512GB, or 2TB of storage, limitations must be clearly specified from the beginning.
If a threat affects multiple platforms, companies should share incident information with one another. Oversight will be credible only when it includes evidence, alerts, and workable measures for shutting down systems.
How Does U.S. Self-Governance of AI Compare with Other Options?
Voluntary agreements can adapt quickly because companies determine their own responses, but legal accountability and inspection authority from independent agencies remain limited. This differs from the European Union’s AI Act, which has clearer requirements. The United Kingdom and the U.S. risk management framework are more flexible but rely primarily on companies’ cooperation.
| Factor | U.S. voluntary agreement | European Union AI Act | U.K. voluntary approach | U.S. risk management framework |
|---|---|---|---|---|
| Strictness | Low | High | Moderate | Moderate |
| Speed of adaptation | Fast | Slow | Fast | Fast |
| Legal accountability | Limited | Clear | Depends on existing law | Limited |
| Authority of independent agencies | Limited | High | Limited | Limited |
The trade-off is clear: the U.S. gains flexibility, but credibility depends on companies’ disclosure and compliance with their commitments.
How Does U.S. Self-Governance of AI Compare with Other Options?
Voluntary agreements can adapt quickly because companies determine their own responses, but legal accountability and inspection authority from independent agencies remain limited. This differs from the European Union’s AI Act, which has clearer requirements. The United Kingdom and the U.S. risk management framework are more flexible but rely primarily on companies’ cooperation.
| Factor | U.S. voluntary agreement | European Union AI Act | U.K. voluntary approach | U.S. risk management framework |
|---|---|---|---|---|
| Strictness | Low | High | Moderate | Moderate |
| Speed of adaptation | Fast | Slow | Fast | Fast |
| Legal accountability | Limited | Clear | Depends on existing law | Limited |
| Authority of independent agencies | Limited | High | Limited | Limited |
The trade-off is clear: the U.S. gains flexibility, but credibility depends on companies’ disclosure and compliance with their commitments.
Strengths and Vulnerabilities of Letting Companies Govern Themselves
Allowing companies to govern themselves helps them respond quickly to risks and draw on knowledge from the teams that directly develop AI. Rules can also keep pace with technology more effectively than laws that take a long time to enact.
But credibility is weakened when companies have to inspect themselves because conflicts of interest may arise, standards may vary between organizations, and insufficient information may be disclosed. Without clear penalties, the agreement may amount to nothing more than a promise.
Pros
- +Respond quickly to risks
- +Use expertise from developers directly
- +Reduce the burden of rules that cannot keep pace with technology
Cons
- −Risk of conflicts of interest
- −Uneven safety standards
- −Limited disclosure
- −Lack of clear penalties
Strengths and Vulnerabilities of Letting Companies Govern Themselves
Allowing companies to govern themselves helps them respond quickly to risks and draw on knowledge from the teams that directly develop AI. Rules can also keep pace with technology more effectively than laws that take a long time to enact.
But credibility is weakened when companies have to inspect themselves because conflicts of interest may arise, standards may vary between organizations, and insufficient information may be disclosed. Without clear penalties, the agreement may amount to nothing more than a promise.
Pros
- +Respond quickly to risks
- +Use expertise from developers directly
- +Reduce the burden of rules that cannot keep pace with technology
Cons
- −Risk of conflicts of interest
- −Uneven safety standards
- −Limited disclosure
- −Lack of clear penalties
The Price Society May Have to Pay Even Without Higher Service Fees
A voluntary system may appear economical, but when harm occurs, victims may have to gather evidence and prove the case themselves. At the same time, inspections conducted by the same company may not be sufficiently independent.
Large companies have stronger legal teams and more resources, while smaller players may struggle to meet the standards. If incident reporting is delayed or commitments are incomplete, the ultimate burden may fall on the public.
Pros
- +Provide a framework for companies to take responsibility
- +Allow problems to be resolved more quickly
Cons
- −Victims must prove the harm themselves
- −Inspections may not be independent
- −Large companies have an advantage over smaller players
- −Incident reporting may be delayed
- −The public bears the burden when commitments are insufficient
The Price Society May Have to Pay Even Without Higher Service Fees
A voluntary system may appear economical, but when harm occurs, victims may have to gather evidence and prove the case themselves. At the same time, inspections conducted by the same company may not be sufficiently independent.
Large companies have stronger legal teams and more resources, while smaller players may struggle to meet the standards. If incident reporting is delayed or commitments are incomplete, the ultimate burden may fall on the public.
Pros
- +Provide a framework for companies to take responsibility
- +Allow problems to be resolved more quickly
Cons
- −Victims must prove the harm themselves
- −Inspections may not be independent
- −Large companies have an advantage over smaller players
- −Incident reporting may be delayed
- −The public bears the burden when commitments are insufficient
Questions This Agreement Still Cannot Answer
Who will verify that companies are actually honoring their commitments, and how independent are the inspectors from the companies? If companies refuse to disclose information, how can outsiders conduct inspections?
Another issue is who decides that a model is too dangerous and what criteria are used. How much safety information will be disclosed to the public without compromising trade secrets?
If a company breaches the agreement, will the government intervene, and how clear will the penalties be? The answers will determine whether self-governance carries real weight or is merely a commitment that is difficult to verify.
Questions This Agreement Still Cannot Answer
Who will verify that companies are actually honoring their commitments, and how independent are the inspectors from the companies? If companies refuse to disclose information, how can outsiders conduct inspections?
Another issue is who decides that a model is too dangerous and what criteria are used. How much safety information will be disclosed to the public without compromising trade secrets?
If a company breaches the agreement, will the government intervene, and how clear will the penalties be? The answers will determine whether self-governance carries real weight or is merely a commitment that is difficult to verify.
The Real Test Is Not the Statement but the Day an Incident Occurs
The credibility of the agreement will become clear on the day a company must choose an option that costs it revenue or a competitive advantage—not on the day when everyone still benefits from a joint declaration.
Going forward, three types of evidence should be tracked: testing reports, reported incidents, and penalties when standards are violated. These will reveal whether the agreement has real effect or is merely a commitment that is difficult to verify.
The Real Test Is Not the Statement but the Day an Incident Occurs
The credibility of the agreement will become clear on the day a company must choose an option that costs it revenue or a competitive advantage—not on the day when everyone still benefits from a joint declaration.
Going forward, three types of evidence should be tracked: testing reports, reported incidents, and penalties when standards are violated. These will reveal whether the agreement has real effect or is merely a commitment that is difficult to verify.