No Mistakes
In 2023/24 police in England and Wales recorded reports of 5.48 million offences.
However, for various reasons, around or just over 60% of them ended up marked No Further Action (NFA). 2.15 million were so marked on account of “no suspect identified”. Between 1.10 and 1.37 million fell by the wayside later along the investigative trail because of other “evidential difficulties” (aka “outcome 16”).
To provide some sort of perspective, in the same year the Crown Courts and Magistrates’ Courts processed around 1.5 million cases.
Wasting time and money
Weeding out reports which end up as NFAs has probably been part of police work for as long as police forces have existed.
False positives will form an unknowable proportion of the NFAs in 2023/24 and every year. Do we ask the police not to record or start to investigate anything until they are satisfied it is not a false positive or until they are sure they can go all the way in securing the evidence and the witnesses? Of course not. That would be ridiculous.
Stay with me on this. In my corner of the forest some people seem paralysed by the spectre of false positives.
“That seems like a good idea… but what about the risk of false positives, whether arising from a report mistakenly made, even if in good faith, or deliberately, maliciously manufactured?”
That concern, real or feigned, too often becomes an alibi for inaction: a reason to keep things exactly as they are. Change is postponed, not for weeks or months, sometimes for years, even decades.
Now some good news
So imagine a tool existed which only identifies items that are definitely criminal or are exceptionally likely to be and, when it does, it will simultaneously always identify a suspect, or point to a strong evidential chain.
A tool that works in this way is called “PhotoDNA”.
It works with child sexual abuse material (csam), both still pictures and videos. All csam is illegal in the UK and, for practical purposes, in every other country.
The original version
PhotoDNA was developed and released in 2009 by Microsoft and Professor Hany Farid, then at Dartmouth College, now at Berkeley. It is given away free to qualifying organizations (although of course there will always be set-up costs and potentially systems integration costs).
Today there are several products which work identically to PhotoDNA. When I refer to PhotoDNA I am including all of them. Some of these other products may not be free at the point of acquisition.
The process
First an image is reported and examined by appropriately qualified humans (note the plural) who work for a recognised authority.
The humans must agree the image is csam.
Maybe in future AI tools will be developed which match human capabilities in terms of confirming a particular image is an illegal child sexual abuse image - we have to hope so - but we ain’t there yet. I am sticking with the here and now. For now.
Once confirmed as an illegal image, PhotoDNA creates a unique digital signature of it. This digital signature, known as a “hash”, can be placed in a database of hashes.
There are other hashing systems but PhotoDNA has singular properties, which is why I am writing about it
Operationally PhotoDNA works in a severely circumscribed, narrow way. It doesn’t and cannot make guesses. It doesn’t and cannot scan or otherwise look at or for anything else. It makes no judgments. It flags matches of digital signatures in the database. In other words it detects copies. That’s it. Nothing else.
Professor Farid has estimated the risk of a false positive to be around 1 in 50 billion. Others have said it is even lower (I have no way of judging the maths involved).
But the the key, indisuptable point is, since 2009, there have been
Zero wrongful arrests or convictions
Despite the fact PhotoDNA first started being used in 2009 there has not been a single reported case anywhere in the world of anyone being arrested, much less charged or convicted, because of a false positive generated by PhotoDNA.
The scale of photo and video uploads is gargantuan
Still pictures and videos are being uploaded on a truly gargantuan scale.
As you will see in the section below on a global reporting hub, on a voluntary basis all the major platforms (except Apple) use PhotoDNA in at least parts of their operations. It works in milliseconds. It has to or it would be no use.
If the wind blows from the north-east in Mongolia
Thus, even if smart-Alecs could show how, in laboratory conditions, when there is an r in the month and the wind is blowing from the north-east in Mongolia while five geeks hold hands in Palo Alto with the lights turned off for 18.7 minutes, a false positive could be intentionally engineered and used to “poison” a database, as far as anybody knows this hasn’t actually happened, at least not in a way that produced real world consequences for anyone.
But not using PhotoDNA has consequences
To my knowledge, Meta is the only platform that has ever adopted PhotoDNA and then publicly declared its intention to scale back its use.
The company is rolling out end-to-end encryption (E2EE) in parts of its virtual spaces where previously it didn’t exist. By not putting in any safeguards e.g. client-side scanning, Meta is therefore knowingly sidelining PhotoDNA.
On Facebook Messenger, E2EE is now on by default. Instagram Direct was going to go the same way but Meta seems to have paused that, at least for now. It can, though, be turned on by individuals.
Meta completed the introduction of E2EE in Messenger in the summer of 2024. In NCMEC’s 2025 report the inevitable result became clear.
There were 6.9 million fewer reports from Meta when, other things being equal, we would have expected to see an overall increase.
Meta at least had the good grace not to try to claim the fall off in reports meant there had been a reduction in the amount of csam being circulated by them.
All Meta had done was blind itself.
The company says it is doing other things to help reduce threats to children and I’m sure that is true. But they could do those anyway. The company’s “pivot to privacy” is the reason it has blinded itself. That pivot comes at the expense of children for sure, the numbers show that, but I very much doubt children will be the only losers.
Last time it was 58%
Lest we forget on 20th December 2020, because of a technical argument within the EU about the applicability of the e-Privacy Directive, Facebook (as it then was) stopped using PhotoDNA altogether for several months.
Microsoft didn’t. Google didn’t. As far as we know most companies that were already using it didn’t.
In respect of other laws, Facebook has paid billions to the EU by way of fines because it ignored or broke them. But in respect of a measure that protects children, showing an over-abundance of caution, it took no chances. Go figure.
The result? A 58% overall drop in all reports of CSAM from the EU.
How many children are being safeguarded by PhotoDNA?
It is not possible to say with any certainty how many children are being identified, located and safeguarded as a result of work helped along or started by PhotoDNA. In that respect, referring to Facebook’s/Meta’s actions in stopping using it, I guess we will never know how many children who could have been identified and safeguarded weren’t. But we do have general, indicative or adjacent data.
For example, in 2019, an ONS Report (section 6) reveals UK law enforcement identified 552 UK-based children who appeared in csam and some of these “ will have been identified with the assistance of the CAID” (CAID being the name of the UK police’s image database).
A 2025 report tells us in the UK 400,000 searches are carried out for csam every month. In 2021 the National Crime Agency sent 20,000 image-related cases to the UK’s local police forces for further investigation
NCMEC - a global reporting hub
Under US Federal law every US company that discovers csam on its platform must report it to the National Center for Missing and Exploited Children (NCMEC). Some non-US based platforms have also chosen to report to NCMEC. NCMEC has therefore become, in effect, a key global reporting hub for csam.
According to NCMEC, in one instance abusive imagery connected to one particular child has been found and reported 1.3 million times over 19 years.
In a blogpost from February 2021 Meta’s Global Head of Safety told us
“Facebook and Instagram… conducted an in-depth analysis of the illegal child exploitative content we reported to the National Center for Missing and Exploited Children (NCMEC) in October and November of 2020. We found that more than 90% of this content was the same as or visually similar to previously reported content. And copies of just six videos were responsible for more than half of the child exploitative content we reported in that time period.” (emphasis added by me).
We have a repeats problem
In 2024, of the 28 million confirmed csam reports to NCMEC 56% had been seen previously. About 75% of the videos likewise had been seen before.
The 2024 Annual Report of INHOPE (the global association of internet hotlines) shows 63% of the images they processed were copies.
Since it was established in 1998 NCMEC has received more than 195 million reports of material suspected of being csam. Each report can reference more than one item. NCMEC has reviewed more than 425 million images and videos. Overwhelmingly (99%+) of these reports came from companies that found the suspected csam themselves on their own virtual premises. Less than 1% came from the public.
The reports received by INHOPE come from their global network of national hotlines and generally are derived from the wider internet.
The recent growth in AI-generated material and in so-called “self-generated” content may change the proportions of originals v repeats but repeats will remain a major part of the problem for the foreseeable future unless and until we achieve a much wider deployment of PhotoDNA. Hosting services in particular need to up their game.
Every image is deadly. Every single image counts
As a reminder, let’s look at why repeats matter.
Whether an image is caught at the moment of its second or third time of publication or on the occasion of its 1.3 millionth, its effects on the victim depicted can be absolutely deadly. Literally. If anyone is in any doubt about that just direct them to the testimony of the Phoenix11.
The damage caused by the original abusive acts is bad enough. Learning images of your pain and humiliation are now potentially available to be viewed by anyone with an internet connection expands, compounds and magnifies the original harm.
Privacy? Whose privacy?
Moreover, these images always represent a gross violation of the depicted child’s right to privacy. When some people argue against the use of tools like PhotoDNA they tend to do so on the grounds of its potential to be used in ways which could infringe someone’s right to privacy. They rarely, if ever, mention cases where there is direct and irrefutable evidence of an actual and major invasion of someone’s privacy, namely the children in the images.
And it’s not only that.
Risks to children as yet unharmed
To the extent the continued circulation of csam encourages or helps sustain abusive or paedophilic behaviour and abusive or paedophile networks, the images also pose a threat to children as yet unharmed everywhere in the world where the internet is present, which is pretty much everywhere in the world.
Never mind what it says about our collective ability or determination to combat online crime, even of this level of awfulness and seriousness.
Similar Tools, Wider Context
On a comparable or even greater scale, all of us already use, tolerate or do not question the use of technologies which work in highly proximate ways to PhotoDNA.
Think about:
Content recognition tools which block suspected copyright infringing material e.g. music, films and live sports.
Anti-virus, anti-phishing and anti-ransomware software which scans every email or text message we receive to warn us of scams or malware.
Spam filters make judgments based on message content and sender history. Emails are examined then directed to spam folders where in all probability they will never be seen again.
In respect of various actions we carry out online, these days it is almost routine to be told to “check spam folders”. Why? Because of the high incidence of spam filters getting it wrong.
And of course we allow all kinds of programmes to read and digest practically everything we do online, or we transmit data about our physical location, particularly through the mobile phones we carry around with us.
No need to “repurpose” anything
One of the utterly disingenuous arguments you hear against the deployment of PhotoDNA is that it could be “repurposed”, by malevolent forces. Really? With all the other tools and scanning methods that already exist and have existed a lot longer than PhotoDNA?
It is very hard to imagine, with the power and resources they already have, or could buy in from globetrotting techies-for-hire, Mr Dictator or Mr Oppressive Boss is just waiting to bully or require companies and organizations to slip in new criteria to their deployment of PhotoDNA so they can find and suppress dissidents or whistleblowers they have not been able to find in other ways.
The “repurpose” argument has no merit.
What’s the problem then?
If PhotoDNA were being more widely deployed the repeats problem would start to fade away, maybe quite rapidly. We need to identify the platforms and companies that are not using it and publish their names. Maybe the regulators could help with that.
We also need to focus on companies that, once notified there is csam on their platform, nevertheless act very slowly to remove it. Are they constantly removing the same images, in other words are they taking no steps to prevent material being reuploaded? Who are the owners of these companies? Who are their bankers?
In the UK the Online Safety Act 2023 allows for the possibility of platforms being directed to deploy PhotoDNA although, given its widespread adoption on a voluntary basis, it seems unlikely such a direction will be necessary any time soon (although see the reference to Australia below).
Within the European Union, however, things are not that straightforward.
The only reason platforms are allowed to use PhotoDNA in the EU is by virtue of a temporary derogation from the e-Privacy Directive. It expires on 3rd April next year and right now there is a risk it will lapse because of larger arguments led principally by the small number of Pirate Party MEPs. If the temporary derogation is not renewed (why can’t it be made permanent?), if the past is anything to go by, Meta will stop all use of PhotoDNA immediately and this time others will likely follow.
Grooming and images not yet detected
Up to now this blog has only addressed cases where repeats of known images are involved. I hope I have shown an overwhelming case for deploying PhotoDNA to locate, delete and prevent those images being reupoladed.
But now tools also exist which, with execptionally high levels of accuracy, can detect both grooming behaviour and images which have not yet been seen and categorized as csam. These latter tools are referred to as “classifiers”. Google has developed its own, so has Meta. Ditto an excellent organization called Thorn.
With all the advances being made with AI, these already extremely good tools will become even better, but this blog is already too long. It won’t be hard for you to work out how I feel about classifiers. Their use will spread. Think of them in the same way you think about the protective tools you already use to warn you of fraud or fakes.
A world of zero trust
To state the obvious, we now live in a world of zero trust.
When it comes to doing anything which seeks to make the internet a better and safer place, there is widespread suspicion which is directed at almost all governments, intergovernmental bodies and their agencies, as well as Big Tech, and little tech for that matter. That’s what zero trust means.
Children have become unintended casualties, collateral damage, in this high-stakes conflict, otherwise the broader use of tools like classifiers and programmes such as PhotoDNA would be an absolute no-brainer.
Navigating this deeply political landscape in an era defined by zero trust is challenging—but absolutely vital. And children’s groups cannot do it on their own.
A worrying message from Australia
As I was finishing this blog a report from Australia’s e-Safety Commissioner landed in my inbox. That will be the subject of my next blog. The report from Oz appears to show that child protection tools are not being deployed in all areas of platforms’ operations where we might hope and expect them to be.