The Law Doesn't Ask If It's Fake

The NPCC Policing Problem Book makes two claims about synthetic content: that deepfakes complicate investigations, and that they undermine trust in digital evidence. The second is the important one, and it points somewhere other than where the response usually points. Detection restores trust in content. It does nothing to restore trust in capture. In a large and growing class of casework, capture is the part that is actually contested.

Start with the diagnosis

The NPCC Policing Problem Book makes two claims about synthetic content. The first is that deepfakes complicate investigations. The second, and the more important one, is that they undermine trust in digital evidence.

We think the second claim is the correct one, and we think it points somewhere other than where the response usually points. The intuitive response is to ask how policing can detect and counter synthetic content in real time. We understand why it is asked. But detection restores trust in content. It does nothing at all to restore trust in capture. And in a large and growing class of casework, capture is the part that is actually contested.

This piece argues that the highest-value intervention available to policing on synthetic content is not a better detector. It is to make the act of publication independently provable, so that the evidence of the offence no longer rests on an officer's word about what appeared on their screen.

Where detection genuinely earns its place, and where it does not

We should be precise about what we are not saying.

Detection has real value in triage and intelligence. Where a force holds a large volume of seized or reported media and needs to prioritise it, a classifier that flags likely synthetic material is doing useful work, and the consequences of an error are proportionate. Detection also matters in fraud, in impersonation, and in voice-cloned social engineering, where the fabrication is the mechanism of the offence rather than incidental to it. And detection will matter defensively, when the defence alleges that a genuine piece of footage is synthetic and the prosecution must answer.

What detection cannot do is carry evidential weight in the offence category driving the numbers. The Problem Book cites a sixfold increase in non-consensual deepfake pornography. For that offence, Parliament has already removed the question detection would be answering.

Parliament has already answered the detection question

Section 66B of the Sexual Offences Act 2003, inserted by the Online Safety Act 2023 and in force since 31 January 2024, criminalises intentionally sharing a photograph or film which shows, or appears to show, another person in an intimate state, where that person does not consent and the defendant does not reasonably believe they consent.

That phrase was drafted deliberately. It is how deepfake imagery was brought inside the offence without creating a separate one. Its effect is that the prosecution never has to establish whether the image is genuine. The point is made more explicitly still in the associated threats offence at section 66B(4), where section 66B(7) provides that the prosecution need not prove that the photograph or film mentioned in the threat exists, or, if it does exist, that it is in fact one which shows or appears to show a person in an intimate state.

Read that again, because it is the whole argument. Parliament wrote the offence so that nobody would ever have to prove whether the image was real.

The baseline offence turns on three things: that the defendant shared the image, that the victim did not consent, and that the defendant had no reasonable belief in consent. The prosecution must still prove that the image appears to show the victim in an intimate state. What it never has to prove is whether the image is real. The statute asks what the image appears to be, not what it is, and everything else it asks about is an act of publication.

A real-time deepfake detector, deployed in support of a section 66B prosecution, would therefore produce expert evidence establishing a fact that is not in dispute, at considerable cost, resting on proprietary models whose error rates a court cannot interrogate, in an arms race against generation that it is structurally losing.

Meanwhile the fact that genuinely is in dispute, that this account published this content at this time, currently rests on a screenshot taken on a police officer's laptop.

The takedown duty is destroying the evidence of the offence

There is a second problem, and it is one policing has been handed rather than one it created.

Intimate image abuse, including deepfake material, is a priority offence under the Online Safety Act: section 66B was added to Schedule 7 by the Online Safety Act 2023 (Priority Offences) (Amendment) Regulations 2024. Platforms are under a duty to prevent it proactively and remove it swiftly, and Ofcom can enforce against those that do not.

The direction of travel is only one way. On 6 February 2026, section 138 of the Data (Use and Access) Act 2025 came into force, inserting sections 66E to 66H into the Sexual Offences Act 2003 and criminalising the creation, and the requesting of the creation, of a purported intimate image of an adult without consent. The government accelerated that commencement in response to the Grok controversy, in which a widely used generative model was made to produce sexualised images of real people without their consent. Political pressure to make this material disappear faster is now a permanent feature of the landscape.

This is unambiguously right for the victim. Every hour that material stays up is further harm, and the regulatory pressure to take it down fast is doing exactly what it should.

But look at what Parliament did in the same breath. New section 66H extends the time limit for these offences, so that a prosecution may be brought within six months of sufficient evidence coming to the prosecutor's knowledge, and within three years of the offence itself. Parliament has deliberately built a three-year evidential tail into offences whose evidence the regulatory regime is deleting within hours. That is the whole problem in a single sentence, and it comes straight from the statute book.

Consider what this does to the evidential picture. The primary evidence of the offence is the publication event itself: this content, served from this URL, at this time, associated with this account. The statutory framework is now designed to destroy that as quickly as possible. What survives is whatever an officer or an analyst captured before it disappeared, and what they captured is a screenshot, or at best a screen recording, produced on their own machine.

Removal and prosecution should not be in tension. Right now they are.

"But we get the publication data from the platform"

This is the first objection any experienced investigator will raise, and it deserves a direct answer.

Sometimes you do. Where the platform is cooperative, in jurisdiction, and responsive to legal process, subscriber and posting records are obtainable and the publication event is well evidenced without anything Shorenet does.

The problem is the residual, and the residual is not small. Much of this material moves through platforms that are offshore, slow, selective in their responses, or actively hostile to legal process. Much of it moves through closed groups, disappearing messages, and sites whose business model is precisely that they do not answer letters. In those cases the request goes out, the takedown happens first, the response comes back months later or not at all, and the case is left standing on the analyst's capture. That is the case that gets discontinued, and it is the case nobody counts.

Even where the platform does respond, its records tell you what its logs say. They are hearsay from a third party with its own interests, produced in a foreign jurisdiction, and increasingly contested. An independent cryptographic attestation of the transmission itself is a different and stronger species of evidence, and it is available on day one rather than in month nine.

What has to be trusted, and why it currently cannot be

Ask a plain question of any web capture tool in use in policing today, from a screenshot upward. What, exactly, are we trusting?

The answer, in every case, is the operator's own machine. That machine holds the TLS session keys for the encrypted connection. Whoever holds the session keys could, in principle, produce a capture of a page that was never served, complete with a valid hash and a valid timestamp. The cryptography attests to the file. It does not attest to the event. This is true of screenshots, of screen recordings, and of every web capture product on the market, however well engineered. We set the same argument out in narrower technical terms in What does a screenshot actually prove?

That was a tolerable weakness when nobody seriously suggested evidence was fabricated. It is not tolerable in a world where the defence can point at the same synthetic media panic that motivates the concern and ask the jury why they should believe the officer's file.

Designing out the crime: prove the publication

The Problem Book asks a further question, and it is the one that matters: what design standards can be co-developed with industry to design out crime in new products?

Here is our answer. Make publication provable, independently of the machine doing the looking.

When an investigator views a page, encrypted traffic travels from the server, across the network, to their device. Every existing tool sits at the far end of that journey, on the device. Move the witness into the network instead. A relay operated separately from the investigating party can attest to what a server sent and in which direction, with the record then bound to an independently certified time. Direction is established from the network tunnel itself, not from anything an endpoint could assert.

Critically, that relay sees ciphertext and nothing else. It witnesses that the server sent this, with no capacity to see what "this" is. For material of this kind, that is not a limitation. It is the only property that makes independent witnessing lawful and deployable at all.

Combine that attestation with the target server's own cryptographic tags and an independently issued RFC 3161 timestamp, and the record is no longer a claim about what an officer saw. It is a proof about what a server did: that this URL served these bytes to this client, witnessed in transit by a party the investigator does not control, and sealed no later than an independently certified time. The timestamp bounds when the record was sealed rather than fixing the instant of transmission, and the distinction is one we would rather draw ourselves than have drawn for us.

Two properties follow, and both matter to policing directly.

It survives the takedown. Because it is a proof about a past transmission rather than a copy of a present page, the platform can remove the content in minutes, as it should, and the evidence of publication remains intact and independently verifiable.

It is deterministic, and the defence can check it. This is the point on which we would invite the sharpest scrutiny, because we have just criticised detection for asking a court to accept an error rate it cannot interrogate, and we should expect the same test applied to us. Our claim is a cryptographic verification with a binary outcome. The artefact verifies against its independent witness, or it does not. Any party holding the artefact can run that check, including the defence expert, without access to our servers or our goodwill. There is no error rate to accept on trust, because there is no model to trust.

Stating our own boundaries

Our standing rule in this section is to state the boundary before opposing counsel does, and it applies to the argument above as much as to anything we criticise.

We do not detect deepfakes and we make no claim to. If a genuine server serves a fabricated video, our evidence proves faithfully that the server served it, and says nothing whatever about whether the depicted event occurred. Content authenticity and capture provenance are different questions. Conflating them is how a sector talks itself into buying probabilistic classifiers to answer evidential questions, and we would rather draw the boundary clearly than blur it in our favour.

We do not identify people. The record establishes that a page as served displayed a given account handle. It does not establish who was at the keyboard. Attribution to a person still comes from subscriber data, device examination and the rest of the investigation.

The relay is independent of the investigating party, not of us. It is operated by Shorenet on Shorenet infrastructure, and its authority to sign comes from a certificate we issue, with a fixed validity window, which can be revoked. Its signing key is held in software rather than in tamper-resistant hardware, so a compromise of that host, or extraction of that key, would permit signatures indistinguishable from genuine ones until revocation. What the design removes is the ability of the party holding the session keys to fabricate unilaterally. It raises the requirement from compromising one party to compromising two. It is not an absolute guarantee, and a supplier telling you otherwise should be asked for the custody model in writing.

Cryptography proves the bytes, not the interpretation. Rendering captured bytes back into a readable page or message is parsing, which is engineering rather than mathematics. Where a reconstruction is uncertain we fail closed and quarantine it rather than present a best guess, on the view that a wrong exhibit reaching a courtroom is far worse than a missing one.

Our claim is narrow, and it is verifiable: that a specific server sent specific bytes to a specific client, sealed no later than an independently certified time, attested in transit by a party that cannot read the content and does not answer to the investigating party. We are not a disinterested party in the general sense, and we will not claim to be one: we sell this, and we operate and certify the relay. The independence that matters evidentially is independence from whoever is making the case, and that is the independence the architecture provides.

Where else this argument reaches

We have written this against synthetic content because that is where the attention currently sits. But the underlying gap, that policing's record of a web publication is an unattested artefact produced on a police laptop, does not respect the boundaries of any one problem statement. Four others touch it.

Cybercrime, malicious communications and digital stalking. Most domestic abuse cases now feature a technology component, and malicious communications and stalking offences are rising sharply. Every one of those cases involves an investigator capturing content from a platform, and none of those captures is independently attested today. The evidential structure is identical, without the deepfake framing to draw attention to it. We would expect the caseload here to dwarf the synthetic content caseload, and we would welcome being shown the figures either way.

Online sexual offences and child sexual exploitation. Capture that survives takedown matters most where takedown is fastest and most obligatory. There is a second question here worth putting on the table: whether attested capture can be combined with privacy-preserving matching, so that evidential records can be checked against known material without either party disclosing the underlying content.

Capture and analysis of digital evidence. This is framed, understandably, around material policing holds: devices, CCTV, audio, cloud accounts, at petabyte scale. We would gently observe that the taxonomy has a gap. The fastest-growing category of digital evidence is material policing does not hold and cannot seize, which exists only on a third party's server, which will be deleted before the case is charged, and which currently enters the evidential chain as a screenshot. Improved evidential quality and maintained evidential integrity are explicit goals. Neither is achievable for this category under current practice.

Automating the evaluation of innovations for legal admissibility. Independent, deterministic, defence-checkable verification is what such a tool would have to be built on. A method whose validity can only be asserted by its supplier cannot be automatically evaluated by anyone.

What we would ask of policing

Shorenet is a UK company. The capture engine is running, the patents are filed, and the technology exists today. We are pre-deployment, and we are not going to point at customers or case studies we do not have. What we cannot manufacture for ourselves is policing's judgement about where the evidential threshold actually sits in contested casework.

We would welcome the chance to work with the NPCC, the Office of the Chief Police Scientific Adviser, the Forensic Capability Network and the CPS on three questions:

In image-based abuse and malicious communications casework, how often is the evidence of publication already the weak link, and what is it costing in attrition and in cases never charged?
What would a standard for evidentially attested web capture need to specify, so that industry builds to it rather than around it?
Where could this be tested against real casework, under proper controls, at a scale sufficient to settle the question either way?

If your unit would consider hosting an evaluation, we would like to hear from you. The mechanism is described in full in how it works, and nothing on this site is gated.

The Problem Book asks for diagnosis before prescription. Our diagnosis is the Problem Book's own: synthetic content undermines trust in digital evidence. Our prescription differs from the one being reached for. Trust in digital evidence will not be restored by proving that content is real. It will be restored by proving which server served which bytes, witnessed in transit by a party the investigator does not control and sealed under an independently certified time, in a way that does not require the court to take policing's word for it. Note the wording: bytes, not what a screen displayed. Turning bytes back into a rendered page is reconstruction, which is engineering rather than cryptography, and we keep the two apart on purpose.

That is a solvable problem. We would like to help solve it, and we are equally prepared to be told we are wrong.

Ant Rose-Edwards is the founder of Shorenet Ltd, a UK evidence technology company building evidentially attested web capture for law enforcement, counter-fraud and government.