Good morning. Andrew here. We’ve got a fun story for you this morning: Employees are using artificial intelligence to create fake receipts to submit fraudulent expense reports. My colleague Sarah Kessler goes deep on this new phenomenon and the race by software companies to develop detection systems. Sarah also has an interview with an entrepreneur who just opened a private school in New York City that uses A.I. to teach its academic classes. (Bill Ackman is a fan.) And make sure to take this week’s quiz. (Was this newsletter forwarded to you? Sign up here.)
‘A cat-and-mouse game’It wasn’t long after ChatGPT began generating realistic images that Anant Kale started seeing posts on social media that explained how it could be used to generate a pretty convincing fake receipt. That was, he recognized, his problem. As the chief executive of AppZen, a software used by finance teams to manage expenses, he had overseen the creation of fraud-detection tools that flagged A.I.-generated receipts. But this was different. “We were like, oh, shoot, this is too easy,” he told DealBook. AppZen immediately started developing a tool to detect fake receipts generated by chatbots. It’s not the only one: The expense management app Expensify added ways to detect A.I.-generated receipts in April, and SAP Concur’s automated expense-auditing tool, Verify, expanded a similar capability to all users this month. This summer, when announcing new efforts to flag A.I.-generated receipts, Nicolas Ritz, who works on product development at the corporate travel software company Navan, summed up the dilemma: “A.I.-generated receipts will only get better from here,” he wrote. “To combat fraudulent A.I., we need to use A.I.” Expense fraud can be a slippery slope. Kale said it’s common for employees to generate their first fake receipt to account for a legitimate expense. Maybe they lost the receipt. But when they don’t get caught, they do it again. Occasionally, the fraud is egregious. AppZen once detected a batch of A.I.-generated receipts submitted by a company employee for hotels and airfare in Bangkok — a city that, upon further investigation, the employee had not visited. The Association of Certified Fraud Examiners, which certifies about 5,000 new examiners each year, regularly asks members to submit the largest case of occupational fraud they’ve investigated in the last 18 months. In the most recent survey, about 13 percent of the cases involved employees who submitted inflated or invented expenses, which can lead to criminal charges. The median loss was $50,000. Fake receipts make it easier. About 30 percent of fraudulent receipts that AppZen catches are now generated by A.I. chatbots, rather than through an image editor or a template service, the company said, and the number of fraudulent receipts it catches overall has increased about 30 percent since May 2024. Expensify said it detects hundreds of A.I.-generated receipts each month, out of the millions of receipts it processes. SAP Concur flagged about 1 percent of receipts audited by Verify as potentially generated by A.I. “I have definitely heard from members and other anti-fraud experts that A.I. is directly resulting in not only an increase in this type of fraud by volume, but also making this type of fraud more difficult to detect,” Mason Wilder, the research director at the fraud examiners association, told DealBook. Convincing fakes have sparked a new tech escalation. Not entirely dissimilar from when, say, home printers created a need for new expense-fraud-detection methods, software companies have built an arsenal of methods for catching a new tier of fake receipts generated by A.I. Chatbots leave a fingerprint in the metadata of the images they generate, but if an employee takes a photo or screenshot of the image, that signal disappears. An algorithm can compare A.I.-created receipts with real receipts from the same vendor. It might pick up on slight differences in font or spacing, for example, that a human eye wouldn’t. Like many expense-auditing software tools, AppZen has relied on identifying suspicious patterns — like spending that is unusual for the time of day or employee’s role — to flag receipts that warrant a closer look. Those suspicious receipts are submitted to its newer second layer of auditing, which looks for patterns that signal that a chatbot may have produced them. While generating restaurant receipts, for example, did the employee always ask the chatbot to use the same server name, or dish order? It’s not a single technique that can detect such receipts, Kale said: “It has to be layers and layers. It’s a cat-and-mouse game.”
The labor market hit a snag this summer. Yesterday’s jobs report came in well below economists’ estimates, with employers making just 22,000 hires in August (the FactSet forecast called for 110,000) — and a revision showed a decline in June payrolls. The S&P 500 had a volatile trading session yesterday, though, as investors weighed the prospect of rate cuts against weaker economic growth. The question now: If the economy is slowing, will the Federal Reserve lower borrowing costs multiple times this year? A new Tesla pay package could make Elon Musk a trillionaire. The carmaker unveiled a plan that could bestow its C.E.O. with unimaginable riches and more control — if it is able to hit lofty goals over the next decade, including an $8.5 trillion market value and a tripling of earnings. Behind the proposal, according to Tesla’s board, is a desire to keep Musk focused on the company as he juggles his responsibilities at his half-dozen other businesses. But critics of Tesla’s last compensation plan for Musk, which a Delaware judge blocked on corporate governance grounds, are likely to protest that this one is also too lavish. Google escapes a breakup. A federal judge ruled that the tech giant doesn’t need to sell its Chrome browser, nor must it stop paying smartphone and browser makers for prime placement of its search engine, despite being found to be running an illegal monopoly over web search. The decision came as a relief to shareholders of Google and of Apple, which receives an estimated $20 billion a year from the search behemoth; it also suggests that efforts by the Biden and Trump administrations to rein in Big Tech may be constrained by the courts. The U.S. Open comes to a close. The women’s singles final will be played today, and the men’s singles final will be played tomorrow. Spotted at the tournament this year: a carefully curated cast of celebrities, bedazzled Labubus and the most profitable cocktail in sports.
The entrepreneur using A.I. to upend school as we know itThe first students at the New York location of Alpha School started classes on Wednesday. And by traditional education standards, they had a very unusual day. Alpha School, which has more than 20 locations across the United States, uses generative A.I. to tailor lesson plans to each student — which makes it possible, it says, to squeeze academic learning into two hours a day. Adult leaders in the room are instructed to primarily focus on motivation, and the remainder of the day is spent working on what the school calls life skills, like grit and teamwork, through projects driven partly by the students’ interests. It’s an approach that has some fans on Wall Street, most notably Bill Ackman. And some critics. The tech entrepreneur Joe Liemandt, who serves as the principal of Alpha School, has invested in building its technology. An entity created by the school’s co-founder MacKenzie Price is developing its software — which it hopes to eventually distribute widely beyond its own classrooms. DealBook spoke with Price. The interview has been condensed and edited. You founded the company in 2014. Obviously that was before generative A.I. How has generative A.I. changed what you’re doing? We realized we could precisely assess what a student knows and doesn’t know and then build lesson plans customized toward that student. We can use a vision model to measure how efficiently and effectively a student is learning. In the next one to two years, we’ll be able to overlay that with the interest graph. So for example, that child who loves fashion design, you can incorporate that in math lessons. Or baseball statistics. We do not use chatbots. Most students will use those to shortcut or cheat. Can the academic work be accomplished in two hours because the instruction is personalized? In a traditional classroom, if you’ve got 20 kids in a fifth-grade class, some have got knowledge in fourth-, third-, second-grade math. Some already know fifth-grade math material and need to be more advanced. And it’s impossible for a teacher to get through all of the teaching that needs to happen for each and every kid. In the life-skills curriculum, there seems to be a focus on entrepreneurship. There is. We pay students for hitting their academic goals, and we start that in kindergarten. So students are able to earn “Alpha bucks.” They’re learning how to earn, how to save, they learn to invest, they learn about how to spend, because they can spend those Alpha bucks at our school Emporium. As students get older, those Alpha bucks turn into real money that they’re able to put toward their passion projects. What are your ambitions for scaling? This is something that I think is going to impact a billion kids. In the next five years, I think every kid is going to have access to a tablet for $1,000 a year that’s going to be able to give them personalized lesson plans. Is the plan to sell the software outside of Alpha School? We already have entrepreneurs who have started private schools that are using it. We also have established private schools who are using our program. And then on the public side, we do have one virtual charter school, called the Unbound Academy. The public system needs to make sure that they’ve got research and data to show this works. And so it’ll be an interesting thing to see how they can adopt it. But I hope in the next 10 years, this will become more widely used in the public system. Joe Liemandt said recently that what he called “the A.I. burn” for the learning platform is $10,000 a student. Is the sort of scale you’re talking about possible at that price point? No, I think that price point is going to come down. The cost of A.I. is continuously going down. And right now, we’re always using the best and the latest to understand how great we can get our learning engine to be. So right now it is $10,000. We’re kind of testing it out in this model and showing the results, and then I believe this is going to go down in price and that’s what’s going to help this become more accessible. Am I right to assume that expanding the software outside of Alpha School is the main commercial opportunity? I don’t know if that’s totally true. Obviously software has always got a high value, but one thing to understand about Alpha is we are a for-profit model. Normally in education, people don’t like for-profit. I think there are going to be more people who get excited about the idea of starting schools and having schools that are delivering good academic outcomes, delivering life skills experience and leveraging the software to do it. The brand of a school that’s delivering great results is very valuable. Critics of this approach tend to argue it will teach to a test and lack depth. What do you say when you hear that? Well, I will definitely say a student who scores a 5 on the A.P. Physics exam knows more than a student who doesn’t. And from there you can go deep. We have a student who did the physics curriculum, and then was able to spend time in the afternoon working on his Alpha X project, which was water-skiing, and using physics principles to analyze how he could take length off the rope and increase his speed. To be able to engage in physics at that level, you have to have base physics in your head, and that’s something that can absolutely be taught very efficiently. If you think about the amount of time in a traditional classroom that’s actually spent on good knowledge transfer, it is extraordinarily little. It’s not that all of the hours in a school year are going toward in-depth education. It’s a wildly inefficient system. So you’re saying you could apply the same criticism to traditional education. Only a third of students in our country are doing math or reading at grade level. That’s the challenge we have. Quiz: Tech execs praise Trump, againTop tech executives gathered at the White House on Thursday for dinner with President Trump and the first lady, Melania Trump. The president opened his remarks with praise for the assembled business leaders. “This is definitely a high-I.Q. group, and I’m very proud of them,” he said. In recent months, tech leaders have sought to curry favor with the president — seeking looser regulation and exemptions from Trump’s punishing tariffs. Thursday’s dinner was no different: One by one, the C.E.O.s lavished praise on Trump for his policies on A.I. and business and promised to invest billions in the United States. Which C.E.O. was notably missing from the bunch? A. Sundar Pichai B. Mark Zuckerberg C. Jensen Huang D. Sam Altman Niko Gallogly contributed reporting. We hope you’ve enjoyed this newsletter, which is made possible through subscriber support. Subscribe to The New York Times. Thanks for reading! We’ll see you Monday. We’d like your feedback. Please email thoughts and suggestions to dealbook@nytimes.com. Quiz answer: C.
|
