AI Safety Pledges: Who Checks Compliance?

AI safety pledges—such as those made by major AI labs—are public commitments to develop and deploy artificial intelligence responsibly. This article explains who, if anyone, verifies that companies actually follow these pledges, and it reveals that, as of now, there is no independent, legally binding enforcement mechanism. Instead, compliance is largely self-reported, with oversight coming from a patchwork of internal governance, academic scrutiny, and voluntary external audits.

Formula

The calculator above computes the annualized rate of public safety incidents relative to a company’s AI safety pledge. The formula is:

Incidents per year = (Number of incidents) / (Time since pledge in months / 12)

Where:

• Time since pledge (t) – the number of months that have elapsed since the company signed the pledge (unit: months).
• Number of public safety incidents (r) – the count of publicly reported safety-relevant events (unit: count).
• Incidents per year – the result, expressed as incidents per 12-month period.

Worked Example

Suppose a company made a safety pledge 12 months ago and there have been 3 public safety incidents during that time. Using the default values:

1. Convert months to years: 12 months ÷ 12 = 1 year.
2. Divide incidents by years: 3 ÷ 1 = 3.
3. Result: 3 incidents per year.

The calculator shows 3.00 incidents per year.

Background: The Rise of AI Safety Pledges

Since the mid-2010s, several AI research organizations have published formal safety pledges. These documents typically commit to principles like transparency, fairness, accountability, and avoiding harmful applications. Notable examples include the Asilomar AI Principles (2017) and the Montreal Declaration for Responsible AI (2018), which were signed by hundreds of researchers and companies. More recently, in 2023, the White House secured voluntary commitments from leading AI developers to focus on safety testing and watermarking. These pledges are often seen as a way to preempt regulation and build public trust, but they also raise a critical question: who holds these companies accountable?

How Compliance Is (and Isn’t) Verified

Currently, there is no single independent body that audits AI safety pledges. Instead, verification relies on a mix of mechanisms:

• Self-reporting: Companies publish annual reports, safety whitepapers, and incident logs. They decide what to disclose, which can lead to underreporting or selective reporting.
• Internal oversight: Many firms have internal ethics boards or safety teams. These groups have access to internal data but may lack independence from management.
• Academic and NGO audits: Researchers and civil society organizations analyze public releases and API access to test for bias, robustness, or compliance with stated restrictions. These audits are often limited by lack of access to proprietary systems.
• Governmental regulation: In some jurisdictions, such as the EU’s AI Act (adopted 2024), certain high-risk AI systems require conformity assessments by notified bodies. However, most safety pledges are not legally binding, and regulatory oversight is still evolving.

In practice, the most common “check” is the court of public opinion and the financial market. A scandal or a widely publicized safety failure can damage a company’s reputation and stock price, creating an indirect incentive to comply.

Common Problems with Pledge Enforcement

Several structural issues make it hard to verify compliance:

• Lack of clear metrics: Pledges often use vague language like “safety” or “fairness,” which is difficult to measure objectively.
• Trade secrets: Companies frequently cite intellectual property concerns to avoid sharing details of their AI systems, making external verification nearly impossible.
• No legal teeth: Most pledges are voluntary and do not include penalties for non-compliance. Breaching a pledge rarely leads to legal consequences.
• Incentive misalignment: The same companies that make pledges also face pressure to release products quickly and maximize profits, which can conflict with safety commitments.
• Rapidly changing technology: Even if a pledge is specific, the technology evolves so fast that the original terms may become outdated.

Comparisons: Voluntary vs. Regulatory Oversight

Voluntary pledges differ significantly from regulatory frameworks. For example, the EU AI Act imposes mandatory requirements for high-risk systems, including risk management, data governance, and human oversight. Violations can result in fines up to €35 million or 7% of global turnover. In contrast, a voluntary pledge has no such penalties. Regulatory bodies like the U.S. Federal Trade Commission (FTC) can act on deceptive practices—if a company makes a public pledge and then violates it, that could be considered false advertising. However, the FTC has not yet brought a major case specifically about an AI safety pledge. Thus, while regulation is stricter, it is also slower and less comprehensive than the scope of most pledges.

How to Evaluate a Company’s Compliance Yourself

If you want to assess whether a company is following its safety pledge, consider these steps:

1. Read the pledge carefully: Note the exact commitments and any defined terms.
2. Compare with public actions: Look for product releases, research papers, and public statements that might contradict the pledge.
3. Check for third-party audits: See if the company has published results from independent evaluators.
4. Monitor incident reports: Follow AI news for any safety incidents and see how the company responds.
5. Look for transparency mechanisms: Does the company release model cards, datasheets, or bias reports?
6. Consider the track record: Has the company been accused of violating similar commitments in the past?

Remember that a lack of public information is not proof of compliance; it could also indicate a lack of transparency.

Costs and Benefits of Compliance

Compliance with safety pledges is not free. Companies may need to invest in:

• Safety testing infrastructure – including red-teaming and adversarial testing.
• Diverse hiring – to reduce bias and improve fairness.
• Documentation and reporting – to demonstrate compliance.
• Legal and ethical review – for each new model or feature.

These costs can be substantial, but they may be offset by long-term benefits such as public trust, reduced regulatory risk, and avoiding costly scandals. For example, a company that invests in robust safety testing may avoid a major incident that could lead to a product recall or a ban.

Alternatives to Self-Regulation

Given the limitations of voluntary pledges, several alternatives have been proposed:

• Independent auditing bodies: Similar to financial auditors, these organizations would have legal authority to inspect AI systems and issue compliance certificates.
• Mandatory reporting: Laws requiring companies to disclose safety metrics and incident data.
• Liability regimes: Holding companies legally responsible for harms caused by their AI systems, which would create a strong financial incentive to comply.
• Open-source requirements: Requiring that certain AI models be released publicly for scrutiny, though this raises its own security concerns.
• International treaties: Multi-country agreements that set binding safety standards, similar to nuclear non-proliferation treaties.

Each alternative has trade-offs between effectiveness, feasibility, and unintended consequences.

Frequently Asked Questions

Are AI safety pledges legally binding?

No, most are not. They are voluntary commitments with no legal enforcement mechanism. However, if a company makes a pledge and then engages in deceptive practices that harm consumers, it could potentially be prosecuted under consumer protection laws, but that is indirect and rare.

What happens if a company breaks its pledge?

There are usually no direct consequences. The company might face public criticism, loss of customer trust, or a drop in stock price. In extreme cases, governments could use the breach as evidence for future regulation, but immediate legal penalties are uncommon.

Can I file a complaint about a company violating its pledge?

You can file complaints with consumer protection agencies (like the FTC in the U.S.) or with industry bodies, but unless the violation also breaks a specific law, the agency may not take action. It’s still worth documenting and publicizing the issue.

Are there any independent watchdogs that monitor AI safety pledges?

Several non-profits and academic groups track AI safety, such as the AI Now Institute, the Partnership on AI, and various university labs. However, they rely on public information and do not have the authority to compel companies to change their behavior.

Do smaller AI companies face more scrutiny than big ones?

In practice, larger companies receive more media attention and public scrutiny, but they also have more resources to manage their image. Smaller companies may be overlooked, but they also have fewer resources to implement robust safety measures.

What is the best way to ensure AI safety if pledges don’t work?

Most experts agree that a combination of mandatory regulation, independent auditing, and liability for harm is more effective than voluntary pledges alone. Until such systems are in place, the burden falls on consumers and researchers to remain vigilant.

Related

• Windows 11 Black Screen of Death: Causes & Fixes
• Hot Work Gear Set Guide for Arknights Endfield
• Time Synchronization Failed? Fix It Now

in t | Time since pledge (months) | months | 12 | 0
in r | Number of public safety incidents | incidents | 3 | 0
out rate | Incidents per year | per year | r / (t / 12) | 2