Yes, the IRS increasingly uses statistics, machine learning and linked taxpayer data to decide which returns deserve examination resources. No, there is not one universal “IRS algorithm,” and a model score is not a finding that you underpaid tax. Different systems screen different populations, flag possible issues or rank cases for examiner attention. Auditors and managers can add or remove issues after reviewing the return.
Recent oversight shows both expansion and limits. The Government Accountability Office (GAO) counted 126 active IRS artificial-intelligence use cases in an inventory dated June 2025, but 77 were still in development. Separate reviews by GAO and the Treasury Inspector General for Tax Administration (TIGTA) found gaps in performance testing, data quality, equity analysis, staffing and strategic management.
As an Amazon Associate I earn from qualifying purchases.
What “data-driven” means in IRS audit selection
IRS selection has several routes rather than a single automated gate. The agency says it uses random selection and computer screening, develops statistical norms from a statistically valid sample audited through the National Research Program, and updates selection information from those results. It can also select a return because it is related to an issue or transaction on another taxpayer’s return.
The IRS explicitly says selection does not necessarily indicate a problem. A refund is not, by itself, an audit trigger, and filing an amended return does not remove the original return from consideration. These statements describe the agency’s general process; they do not reveal a formula that lets an individual calculate audit odds.
#1 Best Overall
Statistical screening and related returns
Statistical norms give examiners a way to identify returns that differ from patterns in comparable filings. Related-return examinations can arise when the IRS is already examining a business, partnership or individual connected to other taxpayers. Those routes may use conventional statistical analysis rather than machine learning.
Machine-learning classification and issue recommendations
TIGTA’s May 19, 2025 review found that IRS divisions had integrated statistical and machine-learning approaches into examination selection. In the reviewed programs, models classified returns or recommended issues for examination. The output helps prioritize scarce staff time; it does not make the final legal determination.
Does the IRS use AI to choose audits?
It uses AI in multiple operational areas, including examination-related work, but the public inventory does not show that every audit is AI-selected. GAO’s March 2026 report counted 126 active IRS AI use cases as of June 2025. Of those, 77 (61%) were listed as in development. Sixty-five use cases were either too sensitive for public reporting or exempt because they were research and development. GAO noted that the inventory had grown from 10 listed use cases in August 2022.
Those figures are a snapshot of a June 2025 inventory reviewed by GAO, not a live count for October 2026. More than one-quarter of listed cases lacked information about their intended agency benefit, and GAO identified AI-enabled tools that were omitted from the inventory. The report also said the IRS lacked an agency-wide process to ensure AI investments supported strategic goals.
Rank #2
- Ideal for Gifting
- Ideal for a bookworm
- Compact for travelling
How the main selection systems differ
Calling all of these tools “the IRS algorithm” obscures important differences in who they cover, what data they use and how people intervene.
| Program or model | Population or coverage | Primary inputs and purpose | Human role | Reported measure or status |
|---|---|---|---|---|
| Dependent Database (DDB) | Returns claiming certain refundable credits | Automated checks and risk flags intended to identify potential noncompliance | IRS processes the flagged returns; GAO called DDB the primary automated selector for these returns | No-change rate was the primary metric; IRS piloted a score that also used revenue gained from audits |
| Form 1040 return-classification model | Individual income-tax returns | Classifies returns and identifies three issues for possible examination | Examiners and managers may independently add or remove issues after risk analysis | TIGTA said IRS had not established a process to show whether the model outperformed earlier methods |
| Large asset corporation (LAR) model | Large corporations with substantial assets | Risk-based selection for examination resources | Model output supports examiner and management decisions | Comparative performance result was not stated in TIGTA’s review |
| Large Partnership Compliance (LPC) model | Complex partnership returns | Development intended to identify potential compliance risks | Model output is subject to IRS examination processes | TIGTA said results were not yet available because the related examinations had not closed |
The table reflects distinct programs, not an apples-to-apples ranking. A no-change rate, revenue protected, issue productivity and other measures answer different questions about effectiveness.
Refundable-credit screening: what the Dependent Database does
GAO’s April 2024 review identified the DDB as the primary automated selector for specific returns claiming refundable credits. The system flags potential risk; it does not establish that a taxpayer is ineligible. GAO said the IRS primarily measured performance by the no-change rate—how often reviewed returns ended without a change. IRS officials were also piloting a scoring approach that included revenue gained from audits.
Equity questions raised by GAO
GAO found that DDB reviews did not comprehensively examine the data inputs and assumptions that could produce different effects across demographic groups. The IRS does not collect race or ethnicity information on taxpayers, although a selection criterion can still have unequal effects. GAO recommended systematic review of inputs, bias assessment, additional outcome measures and reliability checks for the databases used by DDB. The IRS agreed with all six recommendations; agency updates reported in January 2025 are not independent confirmation that every action was completed.
In the same report, GAO cited an academic study attributing 78% of the overall estimated racial disparity in audit rates to audits of Earned Income Tax Credit (EITC) returns. That is the study’s estimate as reported by GAO in 2024, not a new GAO measurement or a current audit-rate statistic. GAO also reported that some DDB risk scores varied by sex and had not been updated since 2001, a finding specific to that refundable-credit review.
Why model output is not an audit finding
TIGTA described the Form 1040 model as identifying three issues for examination, while allowing examiners and managers to add or remove issues based on independent risk analysis. This human review matters because a model can recognize patterns without understanding every fact, explanation or document in a taxpayer’s file.
An audit can therefore begin with a statistical flag, a machine-learning recommendation, a related-party connection or a conventional random screen. The eventual result depends on the records, law and facts examined—not simply on a risk score.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Oversight has found weaknesses in measurement and feedback
Performance testing and model drift
TIGTA said the IRS had not established processes to determine whether the reviewed AI models performed better than prior methods or met their intended objectives. The agency could not readily demonstrate, in real-world context, that AI was improving overall examination compliance efforts.
Rank #4
TIGTA recommended using examination results in model monitoring and refinement, considering ensemble machine-learning methods where appropriate, and setting metrics to detect drift and test expected benefits. The IRS agreed with recommendations on monitoring and metrics subject to staffing constraints and anticipated Treasury guidance. On ensemble methods, IRS said it had already tested and implemented them where appropriate. Those responses do not establish that every model now has a complete, independently validated performance record.
The delayed feedback problem
Models trained on current return data may not immediately receive the results of completed examinations. Many audits close too late to influence the next model cycle. Completed cases could show which suggested issues produced tax adjustments, which productive issues examiners found that models missed, and where machine and human risk assessments differed. TIGTA identified those as potential improvement routes, not proof that the IRS has already resolved the gap.
Staffing and data quality
GAO reported that the IRS Research, Applied Analytics and Statistics group lost 63 employees working full- or part-time on AI during 2025. At the time of GAO’s review, IRS officials had not identified the skills needed to support AI or created a workforce plan to address gaps. Incomplete inventory entries and omitted tools add another management risk: decision-makers cannot evaluate a portfolio they cannot fully see.
Recommended Free Tools
High-income and complex-entity enforcement
Data-driven selection also extends beyond ordinary individual returns. GAO’s February 2024 review of high-income and high-wealth audits examined individual audits closed in fiscal years 2012 through 2022. It reported that Treasury directed the IRS in 2020 to audit at least 8% of returns filed by individuals with income of $10 million or more. GAO found that audit activity in this segment had increased, but the IRS still needed to assess research data, selection models, staffing and training, and centralized management. The IRS could not provide documented evaluation plans for several selection models at that time.
Best Value
- It can be a gift option
- Comes with secure packaging
- Helpful in various ways
The 8% figure is a historical Treasury-directed target, not a personal probability and not a guarantee of the current audit rate. TIGTA’s 2025 review also described plans to use data science and machine learning for large partnerships and planned increases in large-corporation and partnership audit rates, while warning that resource cuts made those plans less certain. Planned rates are not verified realized rates.
Does an IRS audit mean I did something wrong?
No. The IRS’s taxpayer guidance says, “Selection for an audit does not always suggest there’s a problem.” Selection means the return entered an examination process; it is not a legal conclusion that tax was underpaid. The IRS may later find no change, request clarification, or propose an adjustment based on evidence and tax law.
What happens if you are selected?
How am I notified?
The IRS says it initially notifies taxpayers by mail and will not initiate an audit by telephone. An authentic notice identifies the return or issues under review and explains how to respond. Treat an unexpected call demanding immediate payment as a potential scam rather than as proof of an audit.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →How will the IRS conduct my audit?
Audits are generally conducted by correspondence or in person. The notice tells you the format, the response deadline and the contact information for the assigned function or examiner. Follow the notice rather than sending documents to an unverified address.
What do I need to provide?
The IRS will make a written request for documents. Provide the records that support the items identified in the notice, and ask the auditor which electronic records are acceptable if your files are digital. Keep a copy of everything submitted and a log of dates and communications.
How long should records be kept?
IRS guidance generally calls for retaining records used to prepare a return for at least three years. The IRS may generally examine returns filed within the last three years and usually does not go back more than six years, although a substantial error can lead to additional years. Keep source documents organized by tax year so you can reconstruct how each reported amount was calculated.
Quick Recap
A practical response checklist
- Authenticate the notice. Confirm that it is an IRS letter, then use only the contact instructions and deadlines printed on it.
- Identify the scope. List each tax year, form, line item and document requested; do not send unrelated personal information.
- Rebuild your file. Gather income statements, receipts, bank or brokerage records, invoices, mileage logs and worksheets used for the return.
- Reconcile the numbers. Tie each document to the filed return and note legitimate differences, amended information or corrected forms.
- Respond on time. Request an extension through the channel allowed by the notice if you cannot meet the deadline.
- Get help for complex issues. A qualified tax professional can help with partnerships, business deductions, international matters, substantial proposed adjustments or representation before the IRS.
What taxpayers cannot reliably do
- There is no public calculator that can produce a dependable individual audit probability from the IRS’s disclosed information.
- Changing lawful reporting, buying a particular product or choosing specific tax software is not established as a way to avoid selection.
- A risk score, refund, amended return or computer flag is not itself evidence that a tax position is wrong.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
