When AI Safety Commitments Meet Reality: What Charities Need to Know

Update, June 2026: since this was written, Anthropic refused the Pentagon's revised terms and was designated a supply-chain risk. It is now excluded from US Department of Defense contracts while the dispute continues through the courts. The underlying lesson for charities still stands. A provider's safety commitments can shift under commercial and political pressure, however good its intentions.

If you use Claude, you may have noticed the growing attention around the company behind it, Anthropic, and its links with the US Department of Defense. In early 2026, its Responsible Scaling Policy was overhauled. Commentators are divided, while headlines lean toward alarm. As a charity leader, you might be asking what any of this has to do with your organisation.

In short, this is bigger than one company. It is about what happens when safety commitments meet commercial and political reality, and why that matters for any organisation that relies on a provider's promises.

Why Anthropic’s safety record matters

Before we get into the headlines, it is worth understanding why this story matters in the first place. It matters because Anthropic has, more than any other major AI company, built its identity around safety.

The company operates as a public benefit corporation, a legal structure that commits it to considering societal impact alongside profit. It was the first AI developer to publish a Responsible Scaling Policy, a public framework committing the company to pause development if its systems crossed defined risk thresholds.

“Among the companies, Anthropic is the most concerned with safety, and for many people, that encourages them to use Anthropic.”

Professor Geoffrey Hinton,

IASEAI 2026

What happened between Anthropic and the Pentagon?

Two things happened in quick succession, and the timing made each one louder than it might have been on its own.

The Debate on Defence Contracts

In summer 2025, the United States Department of Defense awarded defence contracts to AI companies. Anthropic was the first to be cleared for classified use, with its contract reportedly including restrictions prohibiting certain applications, including autonomous weapons and mass surveillance of American citizens.

The US Defense Secretary has since issued a new AI strategy requiring that Defense Department AI contracts remove company-specific guardrails and instead permit “any lawful use” of AI for departmental purposes.

Anthropic was given until 27 February 2026 to agree to the revised contract terms. CEO Dario Amodei publicly stated that the company would not comply, saying "we cannot in good conscience accede to their request." The government had signalled it could invoke the Defense Production Act or designate the company a supply-chain risk, effectively ending all its defence contracts, which is what subsequently happened (see the update note above).

The debate centres around whether AI companies retain enforceable limits on how their systems are used once they enter national security contracts, or whether any lawful use overrides company-imposed safeguards.

Anthropic’s Responsible Scaling Policy: what changed and why it matters

In the same week, Anthropic released version 3.0 of its Responsible Scaling Policy, the document that explains how it manages risks as its AI systems become more powerful.

One change drew particular attention from AI safety researchers. Earlier versions of the policy included a clear commitment: Anthropic would not train more advanced systems unless appropriate safety measures were already in place. If those safeguards were judged to be insufficient, development could be paused. Version 3.0 removes that firm pledge.

In an interview with TIME , Anthropic’s Chief Science Officer, Professor Jared Kaplan, said that unilateral commitments of that kind were no longer appropriate if competing companies continued to push ahead.

Instead, Anthropic says it will match or exceed competitors’ safety practices, publish more detailed and frequent risk reports, and release roadmaps outlining planned safeguards. It will delay development only if it judges itself to be the leading developer and determines that the risk of catastrophic harm is significant.

The update shifts the Responsible Scaling Policy from fixed internal thresholds to a more conditional, competition-aware framework.

What Recent Events in AI Safety Commitments Mean for Charities

If you are a charity CEO or operations director reading this, you may be wondering why a Pentagon contract dispute matters to you. Here is why: the Anthropic situation is not really about Anthropic. It is about the ecosystem your charity operates in when it uses AI. It reveals three things about that ecosystem that are relevant to every organisation relying on AI providers’ safety promises.

1. Safety commitments can change, and not always by choice

Anthropic’s situation shows that a provider’s safety commitments exist within a wider system of pressures: competitive dynamics, government policy, investor expectations, and regulatory uncertainty. A provider may genuinely intend to maintain its commitments and still find them unsustainable. If you chose an AI provider partly because of its safety reputation, it's important to bear in mind that reputation is not a fixed asset. It is subject to forces outside the provider’s control.

2. The safest available option is still only a relative claim

There are still a lot of unknowns around AI safety, with knowledge gaps likely to increase due to the rate of innovation. It is important to remember that the “safest available option” and “safe enough for your specific context” are two very different things. A charity deploying AI in mental health support or safeguarding needs assurance that goes beyond which provider has the best reputation in a given month.

3. The absence of regulation creates risk for everyone

The current contract standoff likely exists in part because there are no clear legal rules for military AI use. The same absence of clear regulation affects charities. Until regulation catches up, every organisation using AI is making governance decisions in a vacuum. That vacuum combined with the rate of change is what makes provider-level safety commitments so fragile.

What you can do about it

None of this means you must stop using AI, or stop using a specific provider. It means you should build your own layer of assurance, so that your AI governance does not depend entirely on any single provider’s promises. This is exactly the gap I call Silent Drift: the space between the tool your board approved and the tool you are actually using once a provider quietly changes it. My Silent Drift Audit is a practical way to check.

• Monitor your provider’s policy changes. Most post changes on their websites and notify users by email. Assign a team member to review them rather than skipping through. Add a standing item to your quarterly governance review: have any AI providers updated safety policies, terms of service, or model behaviour since the last review? If so, evaluate the impact on your specific use cases.

  
  

• Test independently and regularly. Run your own scenario-based checks using prompts that reflect your service users’ real interactions. Research by Princeton University, recently presented at IASEAI 2026, suggested AI models may change their behaviour significantly over a short amount of time, meaning published safety evaluations may not always reflect the real time experience of people using technologies such as chat bots.

  
  

• Design for portability. The field of AI is changing rapidly. Where possible, avoid deep dependency on a single AI provider. Ask your technology partner: if we needed to switch AI provider within 60 days, would it be possible and what would that involve?

  
  

The bigger picture: why AI safety governance matters for everyone

Anthropic built something genuinely important: a public, measurable commitment to safety at a time when its competitors made no equivalent promises. The fact that this commitment is now under pressure from a combination of competitive dynamics and regulatory absence does not erase what Anthropic contributed. However, it reveals the limits of what any single company can sustain alone.

For me, this situation underscores that the rules of the road for AI are still being written. Charities that work with vulnerable populations have a legitimate and valuable perspective to contribute to this discussion. Sector bodies, umbrella organisations, and policy networks are the right channels for this.

For charities whose work is built on trust and duty of care, the question is not only operational but ethical: what do we owe the people we serve when we cannot fully verify the safety of the tools we use to serve them?

Explore related guides on AI safety and governance:

• AI Governance for Charities: Practical Lessons from the International AI Safety Report 2026 - Deep dive into governance frameworks and structures.

  
  

• Safe AI Innovation in Your Charity - A Step-by-step implementation guide for responsible AI adoption.

  
  

• Service Lifecycle Management for Charities - Integrate AI safety into your operational strategy.

Thinking about AI in your charity? It's a lot to weigh up, and it helps to talk it over with someone who's sat on both sides of the table. Let's talk it through.

Explore the Silent Drift Audit

Note: These insights are based on practitioner experience and do not constitute legal or regulatory advice. Always review your specific funder contracts, data protection policies (GDPR) and safeguarding policies before making significant changes to operations. Examples are for illustrative purposes only; no official affiliation with the organisations or tools mentioned is claimed.