Older forms of AI like machine learning have been used by tax departments for decades. Often called Automated Decision Systems, these systems are used by tax administrators to evaluate returns and identify red flags.
The machine learning model is trained to look for certain features that could indicate a return…
1 – May be fraudulent
2 – That income is underreported
3 – Or that a taxpayer didn’t file a return they should have by cross-referencing data.
A group of Stanford researchers indicated that the IRS’s automated systems had potential bias issues, particularly those evaluating people claiming refundable tax credits. The system was picking disproportionate numbers of taxpayers claiming the earned income tax credit, forcing large numbers of African American Taxpayers into audits.
The Government Accountability Office also examined the IRS system and determined that it could have similar bias issues due to how the machine learning model was trained.
A similar scandal in the Netherlands involved a biased algorithm that caused tens of thousands of people to be denied or forced to repay benefits to which they were entitled. This AI scandal was known as the Dutch Childcare Benefits Scandal.
Between 2013 and 2019, a discriminatory algorithm used by the Dutch Tax Authority incorrectly flagged over 26,000 families for alleged welfare fraud. The system subjected parents to severe financial ruin and ultimately caused the entire Dutch government to resign in 2021.
Now several State Taxation Agencies are using Generative AI, to do their thinking for them. Examples of this include…
1 – The California Franchise Tax Board recently adopted and implemented machine learning models within the last couple of years. They began using them to vet tax returns and to help the agency determine which ones to audit.
2 – The New York State Department of Taxation and Finance uses a system called the Case Identification and Selection System (CISS). This system has two basic purposes being fraud detection and collections. By one 2012 estimate, the system helped the state collect more than $2 billion in payments.
3 – California also began the Enterprise Data to Revenue Project, a long-term IT upgrade by the Franchise Tax Board to overhaul its information technology architecture.
Phase two of the project includes implementing new machine learning models to help the agency identify taxpayers who may be underreporting income, not filing returns, or committing fraud. The idea is that by utilizing more advanced tools, they can trawl through millions of returns and identify patterns indicating the need for staff to take a closer look.
The system sorts through the information, flags items, and then either processes them further with other automated systems or sends them to staff for review. The system has not undergone outside review that could indicate bias.
In other words, the nightmare that occurred in the Netherlands could happen here as well. No one is watching the machines.
A full transcript of the original conversation can be found at the following web address…
https://www.taxnotes.com/tax-notes-live/tax-notes-podcasts/tax-notes-talk/how-ai-bias-affects-state-audit-selection/7w4qz
Let me leave you with this…
Consider the implications of a biased machine learning system choosing the wrong taxpayer for an audit. The burden on taxpayers is particularly high.
Some experts noted that one effect of these more efficient fraud-identification systems is that the new system makes them much more efficient and faster at targeting and identifying individuals to pursue.
Much of this relates to electronic filing. Let’s remember that the IRS only began allowing electronic filing so they wouldn’t have to pay someone to type the numbers we put on returns into the system for checking, flagging, and ultimate audit selection.
With the mass firings and early retirements of almost one third of all IRS Employees, do you really think they aren’t using these systems more now than ever before?
I noticed years ago one result of machine learning choosing audit targets and the speed at which it occurs. On the returns that we don’t complete where an audit is intiated, the timing differences are staggering.
If a person mails in a return that’s being audited, the timing on receiving an audit notice is normally 30 to 35 months after submission. That’s because the statute of limitations on income taxes is 36 months.
If the Service wanted to audit outside that period, they would need to go to a Judge to show cause before extending the statute past three years.
However, electronically filed returns normally skip that review period. Those audit notices are normally received within six months of a filing date.
I can already hear my naysayers screaming that there isn’t a difference between electronic filing and mailing in a return. A reasonable person would say the IRS only needs to scan a mailed return for the system to check it.
And in a reasonable world, you’d be correct, but we’re talking about the IRS. Most of the Services’ computer systems are programmed in FORTRAN. They don’t have scanning capabilities.
The easiest way to avoid an audit in the US is to invest in a $0.76 cent stamp. At that point, these biased machines probably won’t get a chance to look at you.
If you’re having problems with your accounting and tax work, I’d love to help. Please contact us today.
We’re all going to get through this. Let’s get through it together…
Accounting Solutions Ltd. stands ready to complete our mission and purpose of protecting you, your family, and your business. Whether you need Payroll Services, or Accounting and Tax Work, you have but to ask. I’m here and I remain,
Sincerely yours,
Chris Amundson
President
Accounting Solutions Ltd.
773-267-7500
888-310-0300
www.AccountingSolutionsLtd.com
Disclaimer: The content on our website or newsletter is provided solely for general informational purposes and should not be construed as tax, accounting, legal, investment, or professional advice of any kind. Accessing this information does not create, and is not intended to create, an accountant-client relationship. This information may not reflect the most current tax laws, accounting standards, or regulatory developments and may not apply to your specific jurisdiction or circumstances. It is not a substitute for consulting qualified professionals. Before making any decisions or taking any actions, you should seek advice from a professional who is fully informed of all relevant facts pertaining to your situation.
Tax-related content on this site is not intended, nor may it be used by any taxpayer, to avoid penalties that may be imposed under applicable tax laws. To comply with IRS requirements, we inform you that any U.S. federal tax advice contained herein is not intended or written to be used, and cannot be used, for the purpose of avoiding tax penalties or promoting, marketing, or recommending any transaction or matter addressed herein.
All information is provided “as is,” without any guarantee of completeness, accuracy, or timeliness, and without any warranty, express or implied, including but not limited to warranties of performance, merchantability, or fitness for a particular purpose. We disclaim all liability for any loss or damage arising from reliance on this information.
Links to third-party websites are provided for convenience only; we do not endorse or assume responsibility for their content. All materials are the property of our firm and may not be reproduced without prior written consent.