Anthropic's new policy lets it share chats with police: why on-device AI sidesteps the question

Can a cloud AI vendor share your conversations with police?

Yes. Anthropic's consumer privacy policy, effective 8 July 2026, permits it to proactively share Free, Pro and Max conversations with law enforcement based on an internal good-faith belief, without a court order first being served. Around the same time Microsoft restricted employee use of its latest model while it reviewed the change. The plainest way to remove that risk is structural: run the model offline, on hardware you control, so no vendor is ever holding your conversation to disclose. That is what a sovereign, on-device system does, and it is the honest core of the Mickai argument.

We want to be precise, because the detail matters. The policy in question covered consumer tiers. Anthropic's enterprise and API tiers were excluded from it. So the sharper, truthful point is not that enterprise chats got shared to police. It is that the architecture of hosted AI puts a third party between you and your own words, and that third party can rewrite the terms of disclosure whenever it chooses. Control and auditability are the real issue, not a single headline.

Why does a hosted model create this exposure at all?

Because someone other than you is holding the data. When you type into a cloud assistant, your prompt, the model's reply and often a great deal of surrounding metadata land on infrastructure the vendor owns and governs. Their policy, their jurisdiction and their internal judgement then decide what happens next. A good-faith belief standard means a human or an automated system inside the vendor can form a view that disclosure is warranted, and act on it, before any judge has seen the matter.

For a regulated organisation this is not a comfort. A bank, a hospital, a defence contractor or a government team cannot outsource the decision about when their own records reach a police force to a supplier's discretion. That decision is theirs to make, on legal advice, on the record. The problem is not that vendors are acting in bad faith. It is that the wiring puts the choice in the wrong hands.

How does on-device AI actually change the picture?

It removes the middle party. When the model, the weights and the inference all sit inside your own walls, offline, on machines you control, there is no external service quietly retaining a copy of the conversation. A police request, lawful or otherwise, has to come to you, the data controller, and you answer it the way you answer any other legal demand: with counsel, against the actual warrant or order, inside your own compliance process.

That is the whole shift. Disclosure stops being a setting in a supplier's policy and becomes a decision you own. Mickai is built for exactly this posture. The system runs on the customer's own hardware, offline, with no dependence on a cloud vendor to process or store the conversation. There is no upstream account that can be served, mined or repurposed under a changed terms of service.

What does Mickai add beyond just being offline?

Being offline stops the leak. Auditability answers the next question, which is what you disclose when you legitimately must. Mickai seals every consequential action into a post-quantum signed audit ledger, using ML-DSA-65 under FIPS 204. That means when a lawful request does arrive, you can produce a precise, tamper-evident record of what the system did and when, signed with cryptography designed to survive a future quantum attacker. You are not scraping logs from a third party and hoping they are complete. You hold the ledger, and it is yours to reason about.

This matters in two directions. It protects the organisation, because the record is defensible and controlled. It also disciplines the organisation, because sealed actions cannot be quietly rewritten later. Sovereignty is not the same as secrecy. It is the ability to say, precisely and provably, what happened on your own systems, and to decide yourself what leaves them.

Does this make an operator exempt from the law?

No, and we will not pretend otherwise. Running Mickai does not put you above lawful process. A valid warrant or court order still binds you, exactly as it binds any data controller holding sensitive records. What changes is who holds the data and who makes the disclosure decision. Instead of a vendor forming a good-faith belief and acting on it, you receive the request, test it against the law, and respond. That is a better position, not an escape hatch, and honesty about the difference is the point.

It also does not make you immune from your own obligations. You still owe your regulators, your customers and your staff proper data governance. Mickai gives you the tools to meet those duties on your own terms. It does not dissolve them.

What should a regulated buyer take from all this?

Read the disclosure clauses of any AI you rely on, and read who they exclude. Ask a simple question of every assistant your people use: if this vendor changes its policy overnight, what of ours can it now hand over, and to whom, and on whose judgement? For consumer-grade tools the answer is often uncomfortable. For hosted enterprise tools it is better but still not fully in your hands, because the data still leaves your estate.

The sovereign answer is to keep the model, the data and the decision inside your own perimeter. That is the design Mickai ships: 50 brains, around 60 studios that replace the cloud and SaaS stack, running offline on operator-owned hardware, with 104 filed UK patent applications and 2,340 formal claims behind the architecture. It is built and live, not a slide.

We are an ally to the people scoring these vendors, not a magic bullet. What we do about Anthropic's policy change is straightforward: we make sure there is nothing for a vendor to share, and we hand you a sealed, post-quantum audit trail so that when a lawful request comes, the answer is yours to give.

“Anthropic's consumer privacy policy effective 8 July 2026 permits proactively sharing Free, Pro and Max conversations with law enforcement on an internal good-faith belief, without a court order.”

Frequently asked questions

Did Anthropic share enterprise or API conversations with police?

No. The consumer privacy policy effective 8 July 2026 covered Free, Pro and Max tiers, and enterprise and API tiers were excluded from it. The broader concern is structural: a hosted vendor sits between you and your own words and can change its disclosure terms.

What exactly does the new policy allow?

It permits Anthropic to proactively share consumer-tier conversations with law enforcement based on an internal good-faith belief, without a court order first being served. Microsoft restricted employee use of its latest model while it reviewed the change.

How does running AI on-device stop this?

When the model runs offline on hardware you control, no external vendor is holding your conversation, so there is no upstream account for anyone to serve or disclose. Any lawful request comes to you, the data controller, and you answer it through your own legal process.

Does Mickai make my organisation exempt from lawful requests?

No. A valid warrant or court order still binds you as it binds any data controller. What changes is that you hold the data and make the disclosure decision yourself, rather than a vendor acting on its own good-faith belief.

What is the audit ledger and why does it matter?

Mickai seals every consequential action into a post-quantum signed audit ledger using ML-DSA-65 under FIPS 204. When you must lawfully disclose, you can produce a precise, tamper-evident record you control, rather than reconstructing logs held by a third party.

Is this about secrecy?

No. Sovereignty means you can prove exactly what your systems did and decide yourself what leaves your estate. You still owe your regulators, customers and staff proper governance; Mickai lets you meet those duties on your own terms.