Is that the IRS with your refund? No, it’s a ChatGPT scam

THE AI TAX SCAMMERS COMETH: When Sergey Shykevich, a threat intelligence manager at the cybersecurity company Check Point Research, prompted ChatGPT to produce an example of a tax scam email containing malware, the AI chatbot spit back a grammatically immaculate email on the Employee Retention Credit that asked the recipient for their Employer Identification Number, payroll information and a list of their employees and their social security numbers.

Another request to ChatGPT for an example of a call that could be used to “educate elderly people on tax scams” produced a transcript between a scammer impersonating an IRS agent pressuring a senior to pay a fake tax debt.

Both responses were produced instantly and at the prompting of simple requests inputted by Shykevich.

And that’s just the beginning.

Cyber security experts that are researching the burgeoning field of AI told Weekly Tax in interviews that chatbots threaten to advance the credibility, sophistication and success of tax scams in leaps and bounds, not only in phishing and malware scams that take the form of emails but also in calls where so-called “deep fakes” could imitate the voice of an IRS agent.

BUT FIRST: We are hurtling at full speed towards Tax Day on April 18, and all the tax groups that have skin in the game are looking to make their voices heard on Capitol Hill as the IRS closes the filing season.

Indeed, it’s promising to be a hectic day, with more than two dozen “Patriotic Millionaires” planning to show up outside the Capitol with life-sized cutouts of Elon Musk, Mark Zuckerberg and Jeff Bezos calling for tax hikes on the rich.

Meanwhile, Americans for Tax Reform will be holding a conference that afternoon with Majority Leader Steve Scalise (R-La.), Budget Chair Jodey Arrington (R-Texas) and Republican Study Committee Chair Kevin Hern (R-Okla.), among other lawmakers, that will surely involve some measure of roasting of the IRS’s new $80 billion windfall.

Send us your best Tax Day vibes and notes.

Email: [email protected], [email protected] and [email protected].

You can also reach us on Twitter at @tobyeckert, @brian_faler, @ben_guggenheim, @POLITICOPro and @Morning_Tax.
Want to receive this newsletter every weekday? Subscribe to POLITICO Pro. You’ll also receive daily policy news and other intelligence you need to act on the day’s biggest stories.

I’M SORRY, DAVE. I’M AFRAID YOU OWE TAXES: Well, it’s not necessarily that cyber experts are afraid of a hyper-intelligent HAL 9000 impersonating a tax collector just yet, but rather that the availability of bots like ChatGPT will make tax scams already used by foreign cyber criminals much more efficient and credible.

Foreign criminals, especially those in Eastern Europe, target Americans with tax scams every year, but poor grammar and inaccurate references to the U.S. tax system can make those scams easy to detect.

“Their revenue from tax scams is millions every tax season,” Shykevich said. “So even if they integrate even some parts of the [ChatGPT] engine, their operation would become much more [competitive].”

The scams produced by ChatGPT, as shown below, are more credible because they’re well-written, can reference information about certain credits and deductions founds on IRS.gov and even in theory track down data online on a specific taxpayer, like those pertaining to a taxpayer’s credit score and business.

Here’s are some examples prompted by Shykevich that are partially redacted so they can’t be easily replicated. The first involves ChatGPT producing a fake email from the IRS claiming that a taxpayer is eligible for a $1,450 refund:

Here’s an example of a tax scam call prompted by Shykevich that shows a transcript with an elderly person and a fake IRS agent:

Notably, according to Shykevich, there doesn’t seem to be many guardrails in place yet against eliciting tax scams from ChatGPT— while Open AI developers have trained ChatGPT to identify abusive requests regarding bank information that employ malicious code, Shykevich had no problem prompting schemes from ChatGPT like those noted on the IRS’s Dirty Dozen list.

Alarmingly, a list by BeenVerified found that tax and IRS schemes were among the fastest-growing scams of 2022 and increased 55.4 percent compared with the year before.

DATA, DATA, DATA: While we didn’t get many details from the new strategic operating plan on exactly how the IRS plans to select high net-worth individuals and large corporations for audit (in large part, according to the agency, because the IRS doesn’t want to give tax cheats helpful clues), we do know that the phrase “analytics” was used a whole lotta times in the report.

And given the IRS’s limited resources, it wouldn’t be terribly surprising if the agency is planning to leverage algorithms and computer scientists to manage the audit selection process.

That raises all sorts of interesting solutions, but also new potential pitfalls, for the IRS as it tries to capture an estimated half-trillion dollars that escapes the agency’s grasp every year.

“So what the IRS is doing, as far as we know any details, is they are having these algorithms look at huge masses of data — and that’s tax return data, financial data, data from third-party reporting and information they’re getting from treaty partners and international information sharing—and they are using that to locate targets,” Rob Kovacev of Miller & Chevalier said.

The process could entail, for instance, looking at credits claimed by businesses of a certain size and targeting a business that claimed a credit several standard deviations larger than the average, according to Kovacev.

What that might look like: Data analytics are run by the IRS’s in-house research organization, called the office of Research, Applied Analytics, and Statistics (RAAS), which reports directly to Deputy Commissioner for Operations Support Jeff Tribiano.

And here’s the intriguing thing: RAAS has been dabbling over the years in some pretty cutting-edge stuff that combines psychology, neuroscience and behavioral economics to improve tax compliance and administration.

For instance, in a report published by the RAAS on behavioral insights research, the office cited a study in which the New Mexico state government identified taxpayers who would be most likely to improperly report their unemployment benefits based on the time of day they go online.

“Machine learning methods are increasingly used to segment populations based on characteristics, behaviors, or attitudes,” the report said.

The dangers: So the IRS could start doing some pretty savvy things once it hires new technologists to write and tweak algorithms for audit selection, but it also comes with the risk that the IRS won’t be able to fully explain how certain taxpayers were selected for scrutiny—as when a Stanford study published earlier this year indicated that Black taxpayers are three to five times more likely than non-Black taxpayers to be audited.

If an algorithm wants to maximize collections, it could conclude going after low-income taxpayers is the most efficient way to do so, since those taxpayers don’t have the resources that higher income folks do to fight collections, Kovacev said by way of example.

“Because these algorithms are black boxes, this could happen without anyone ever intending that result, or even that it’s happening before it’s already happened,” Kovacev said.

WERFEL’S FIRST TEST: Speaking of which, Commissioner Danny Werfel promised to lawmakers at his confirmation hearing that he would get to the bottom of the racial disparities in auditing illuminated by the Stanford study within 60 days, which could prove complicated with all those aforementioned opaque IRS algorithms.

In an April 13 letter addressed to Werfel and Treasury Secretary Janet Yellen, Sen. Elizabeth Warren (D-Mass.) reiterated her concern about racial inequities in tax enforcement — noting that 80 percent of the disparity in audit rates between Black and non-Black taxpayers alone can be chalked up to racial inequities within audit rates of Earned Income Tax Credit recipients.

The Stanford Study determined that the differences cannot be fully explained by racial differences in income, family size or household structure.

“The same report suggests this disparity may be due to IRS audit selection computer algorithms (or assigned scores),” which target under-reporting by taxpayers and overpayments of refundable credits, Warren said.

Around the World

Guardian: “Indonesia considers tourist tax to curb bad behaviour in Bali"

Reuters: “Mexico to help Pemex with $2 billion tax deferral"

Tax Notes: “A Pillar 2 Tour Around the World"

Around the Nation

TIME: “There’s a New Push to Create a U.S.-Run Free Electronic Tax-Filing System"

WSJ: “U.S. Buyback Tax Could Hit More Foreign Firms Than First Expected"

Bloomberg: “Tax Day Brings Some Extra Wrinkles for Crypto Investors"

Also Worth Your Time

Reuters: “New report looks at what corporate tax departments are up against in 2023"

WSJ: “Taliban Prove to Be Formidable Tax Collectors, Putting Squeeze on Afghans"

NBC: “Startup aims to help climate companies unlock billions in Inflation Reduction Act tax credits"

Did you know?

March 1 was the original filing deadline when lawmakers first established the federal income tax.