Machine learning saves £4.4M in UK.gov work and pensions fraud detection The UK government's Department for Work and Pensions (DWP) has saved £4.4 million over three years by using machine learning to tackle fraud, according to the National Audit Office (NAO). However, the public spending watchdog found the department's ability to expand this work is limited by fragmented IT systems and poor cross- …
Here's fraud detection fucking up for people in Northern Ireland.
Admittedly, this was HMRC (which administers the benefit in question) and it's not clear if machine learning has been used. But I'm sure DWP won't make any these kind of mistakes. Nah, who am I'm kidding?! Of course they will accuse legitimate users of fraud.
People processing everything have a chance to learn, machines and other processing methods show how much people have learnt to sent them up originally and then applied them to the situation. People think, but machines and coding only process the environment, and make money but never think.
As kids we all learn for years a lot to think reasonably well - machines and everything else are only updated to look like they have a good result (hopefully).
You've just reminded me of this story where someone claimed a £763 advance in the name of a teacher and left the teacher on the hook for the repayment. That was DWP. And the fraudster attended face-to-face meetings to verify their identity...
So, not only did the fraudster get away with it, an innocent person footed the bill.
"That seems like an edge case a remotely competent person would have flagged when designing the system."
You're assuming such a person actually was aware of the existance of the Common Travel Area (CTA) and how it functions. You can't design around something that you're not aware of.
As someone from Northern Ireland I've found that GB government agencies, GB companies, and many GB people seem to have a Schrodinger's Cat view of NI (that it is simultaneously "part of" and "not part of" the UK).
It's not an NI specific problem: it also caught someone who booked a flight, but never took it. And someone who flew out by plane but came back by Eurostar. So NI is a bit more exposed. But it's a flaw in their thinking.
They are continuing to justify it because it found more fraudsters than false positives. Of course, at this point, we don't know how many of the false positives have yet to come forwards. (Or if they will ever come forwards.)
"it also caught someone who booked a flight, but never took it."
That sounds like a flaw with the E-Borders system - AFAIK the details of passengers on a flight are passed to gov *before* a flight takes off. Sounds like that data is not later updated once the flight actually takes off to update with the actual (in transit) passenger list. Indeed that article says "suggesting that the Home Office immigration operations could have had access to passenger records" which would appear to be a reference to E-Borders.
"And someone who flew out by plane but came back by Eurostar."
AFAIK Eurostar (and Ferries) are also part of E-Borders and so HMRC should have had access to those records as well. Again sounds like a flaw with E-Borders system.
The Northern Ireland situation is different as AFAIK E-Borders is *not* involved in commercial transport services (buses and taxis) crossing the NI/Ireland border and obviously private cars are not either. Whilst ANPR cameras close to the border will record vehicles crossing there are no person checks.
Less "innocent unless proven guilty" and instead "you ARE guilty, of what we are yet to decide, but we WILL get you, don't think those wishy washy courts will save you either, once we get the arrest powers we deserve, our next step will be DWP run welfare courts with powers to jail and no need for any appeal options"
I wish I was being dramatic but the DWP apparently are convinced that they are the ministry of love from 1984.............
1) Overpayment and fraud of £9.5bn is 3.3% of total benefits paid - I'd reckon a well run system should achieve something of the order 0.3-0.7%
2) DWP identify £4.5bn of overpayment and fraud stopped currently...so
3) Manual processing stops about a third of benefits error and fraud, and...
4) Machine learning has saved one-thousandth of the amount saved by manual processes
And finally, total benefits error and fraud is about one fifth of the error and fraud in tax collection of £47bn in 24/25 (which doesn't include tax avoiding big tech firms).
A footnote: The link "DWP's IT systems are not fully integrated" leads to a story that is not about the operational systems that are implicated in overpayment and fraud, but refers to back office support services systems like HR, procurement, accounting.
No they don't conflate them, that's just how it appears in the NAO report and in my comment. According to DWP published figures, for 24/25
Total benefit expenditure overpaid in FYE 2025 was 3.3% (£9.5bn).
Overpayments due to Fraud were 2.2% (£6.5bn) in FYE 2025.
Overpayments due to Claimant Error were 0.7% (£1.9bn) in FYE 2025.
Overpayments due to Official Error were 0.4% (£1.0bn) in FYE 2025.
Underpayments were 0.4% in FYE 2025 (£1.2bn), these are all classed as Official Error.
https://www.gov.uk/government/statistics/fraud-and-error-in-the-benefit-system-financial-year-2024-to-2025-estimates/fraud-and-error-in-the-benefit-system-financial-year-ending-fye-2025#total-estimates-of-fraud-and-error-across-all-benefit-expenditure
Funnily enough I don't have ANY trust in figures generated by an agency that has wilfully defied courts, wilfully obstructed parliamentary committees, wilfully destroyed documentation they were ordered to preserve, denied having documentation that they not only have but are fully aware of being in possession of.
The DWP not only distort the truth but do this wilfully and maliciously, while treating vulnerable claimants with utter and total disdain to the point of slow walking making tribunal ordered back payments, miscalculate payments, fail to make uprated payments on an ongoing basis - all to be as obstructive as possible in knowledge a reasonable percentage will either give up or not question matters, leaving them out of pocket.
What's worse is where DWP have lost at tribunal and then retailiated by "reassessing" the claim and reinstating their original decision to force claimants to go back to tribunal again.
Private company tried anything like that, then they would end up with court appointed bailiffs at their doors or the responsible staff member(s) ordered to appear in court (and in handcuffs if that was required) on a charge of "contempt of court" - judges not being especially tolerant of those sticking up 2 fingers to the authority of the court.
Most of the "overpayment" is due to DWP Error including giving deliberately wrong information and is NOT due to fraud.
They also don't mention the large amount of welfare / social security that people are entitled to that goes unclaimed due to wilful misinformation by the DWP and the "hostile environment" the DWP has created around social security - where claimants are treated as a criminal underclass to be kept under category A prison level surveillance and their lives deliberately and intentionally made as wretched and miserable as possible 24/7/365, where the DWP likely pop the champagne corks everytime a claimant takes their own life or leave a "welfare monkey" on someone's desk anytime a claim is approved (ala the border agency)
"It is interesting how government is diligent in chasing the poor who might have cheated the system out of few quid."
£6.5bn a year might be your definition of a few quid, it certainly isn't mine. I also doubt that many genuinely poor people are making fraudulent claims.
"Oooh, me 'ip is playing up som't rotten and I canna 'ford the gas ne more; I'd betta mek a fictitishis claim fr'm the Nash wi' me dead grandads nemh. 'E paid in is ole life, so its only fair"
I assume you refer to tax error, avoidance and evasion, and if that's the case then Government publish estimates on this under the term of "tax gap", and that's £47bn. That doesn't wholly cover "legitimate" tax avoidance because if its done "properly" then by definition that can't be part of the tax gap, although HMRC are still including avoidance as a category of the tax gap. The report (that I won't link because you could have done your own homework) helpfully breaks down who the big dodgers are. And the winner is........small business, responsible for 55-60% of the total. Large businesses are responsible for 12%, and the very wealthy for 5%.
I suspect that the tax avoidance of large corporations and the wealthy is understated because they can use 100% legal paths to avoid UK taxes - eg claiming that anything bought from Amazon is bought from a Belgian company, likewise all the family trusts, overseas investments, capital gains deductions etc of the wealthy.
But this has what to do with the thread?
Because if you actually read the reports, you will learn that HMRC doesn't actually know the figures for big corporations and the rich. The numbers presented are pulled out of the arse, if you pardon me using scientific statistics term.
Whereas the numbers for the poor are precise in comparison. This shows where the focus of government is.
What is the machine leaning for?
You have an output which is benefits and you have an input which is tax paid. The system checks your tax paid then determines eligibility for benefit automatically. In cases of overpayments the system already looks back, flags them up and requests payment. It does not do the same for underpayments btw.
In cases of fraud the tax paid would have to be false but they wouldn't have that information. What information are they using to detect fraud? There isn't any that would come from the system itself. Patterns aren't reliable because what pattern would be looking for? Even if you did look for a pattern it would then require confirmation and lengthy investigation which based on patterns would be in my opinion a huge waste of resource.
Bizarre. Is this ML for ML's sake? or is it justification for the standardisation of data for easy extraction to Palantir?
If only things were so simple.
The reality of tax and benefits is that both are incredibly complicated. In the case of benefits, much of the relevant information is declared by the claimant and many details aren't recorded on the tax system (eg savings, marital status, household status and occupancy, various eligibility criteria).
£4.4 million over 3 years is a rounding error for the DWP. I wonder how much they spent on the system to get that saving?
Yes, you could try to argue that the system will continue to find more savings. I doubt it though - it'll require fundamental architecture changes to scale, it's known to have unfairness built right into it, and when/if the data formats are standardised (probably around the point the Porcine Red Arrows display is overhead) it'll need a rewrite anyway.
In an other area of commerce returns like that would be so dramatically and obviously indicate that the solution was unfit for purpose that you would expect the contract to be dissolved and the provider to scuttle off back under a rock in abject shame.
But not with AI/ML. Keep shoveling the 50 pound notes into the furnace...
Why do you think the DWP and Chancellor Sturmer are so intent on getting access to everyone's bank accounts, getting rid of cash, rolling out live facial recognition across as much of the country as possible and where it's all feeding the same database?
Facial data is a doddle - most people have a driving licence and/or passport, those who don't will be tackled in another way - ID card that isn't an ID card, facial recognition tied to something like a doctor's appointment etc...
Then it will be Chinese style social credit scoring and for those who are deemed awkward - well no power or gas for you, oh it's minus 10 outside? Well you should have thought twice before making those posts or saying those unacceptable things (whether in public or in private)
Though doesn't matter who is Reichschancellor really, faces and parties change, policies stay the same.....
We all know that many of the gov systems have components supplied or administered from the commercial sector. Besides the commercial sector there are still staff costs.
I've been a contractor on a gov project in the past so I've got to ask how much did this savings cost?
The real cost will probably look nothing like the spreadsheet that someone has hastily knocked together for management to show that everything is in control, yep, nothing to see here. All the costs are accounted for.
If its true, then they should have it independantly audited where the government has no way to influence the output, then we might trust things a bit more.
But ... Conversely, if they have pulled a phoenix out of the burning pile of government IT systems and made something work well, then keep it up, take a look at HMRC and fix the mess there. If you could reduce our costs of wasted time having to spend 45 minutes a go on hold, during which time we are played constant messages to use the on-line tools that don't work, then talk to a person that promises something in x weeks time. Rince and repeat, wasting 6-9 months getting anything done for very simple queries. This might actually save all of us a bunch of time, but nobody measures the waste that governments push onto their subjects.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
