Among the major shocks of this month’s Trump administration announcement of massive tariffs slapped on dozens of countries worldwide lies a less prominent but arguably more important one: The entire tariff scheme may just have been produced by artificial intelligence.
The fallout from Liberation Day — what the White House labeled April 2, the day it proclaimed its global tariff schedule — was immediate. In the wake of tariffs ranging as high as 47% on some countries, financial markets seized up from Tokyo to Tel Aviv to Toronto. The S&P 500 briefly entered bear market territory for 2025. The retirement accounts of tens of millions of Americans nosedived. (Markets have since at least partially recovered with the White House’s pledge on April 9 that “reciprocal” tariffs would be suspended for 90 days.)
It wasn’t just Wall Street that suffered but Main Street. A University of Michigan economist projected a $5,000-per-family annual loss and a 6% hike in the cost of living. Small businesses seemed to be in for the worst of it. Others predicted nearly a full-point increase in unemployment, costing more than 2 million jobs.
The reaction from leaders around the globe was no less withering.
“The administration’s tariffs have no basis in logic, and they go against the basis of our two nations’ partnership,” said Anthony Albanese, Australia’s prime minister.
“I deeply regret the path the U.S. has embarked upon, seeking to limit trade with higher tariffs,” Swedish Prime Minister Ulf Kristersson lamented.
“Japan has been the biggest investor in the United States for five straight years, and the tariff policies could hurt Japanese companies’ investment capabilities,” Prime Minister Shigeru Ishiba stated.
In short, the administration’s policy sparked near-universal condemnation, including across the op-ed and news pages of last week’s issue of this magazine, appropriately titled “Trump’s Tariff Tsunami.”
But as journalists and data scientists dug into the unusual tariff schedule itself — 47% on Madagascar, 32% on Taiwan, and 10% on the uninhabited Heard and McDonald islands? — an unexpected wrinkle emerged: It looked like the treasury and commerce officials who came up with the specifics of how severe the levies would be on particular countries had derived their figures from querying large language models, or LLMs, such as ChatGPT.
The first piece in the puzzle came from journalist James Surowiecki, a contributor at the Atlantic and Fast Company, who posted on the evening of Liberation Day, “Just figured out where these fake tariff rates come from. They didn’t actually calculate tariff rates + non-tariff barriers, as they say they did. Instead, for every country, they just took our trade deficit with that country and divided it by the country’s exports to us.”
By way of example, Surowiecki calculated that “we have a $17.9 billion trade deficit with Indonesia. Its exports to us are $28 billion. $17.9/$28 = 64%, which President Donald Trump claims is the tariff rate Indonesia charges us.”
In other words, far from imposing purportedly reciprocal tariffs on countries that have restricted American imports, and rather than conducting a complex, fact-specific formula accounting for the nuances of particular trade practices (or abuses) that our trading partners supposedly engaged in, the administration performed a simple act of arithmetic.
In response, the administration published a complicated-looking formula, replete with at least four Greek letters, that it claimed reflected the actual arithmetic animating its calculations. But it turned out that the equation essentially did exactly what Surowiecki suggested. (To add insult to injury, my American Enterprise Institute colleagues Kevin Corinth and Stan Veuger explained how the formula “makes no economic sense” and was “based on an error.”)
Then came the most interesting part: Various armchair data chuggers discovered there was a high likelihood that the simple formula used by the administration was very likely generated by AI. “This might be the first large-scale application of AI technology to geopolitics,” Rohit Krishnan wrote on X. ChatGPT 4o, “o3 high, Gemini 2.5 pro, Claude 3.7, Grok all give the same answer to the question on how to impose tariffs easily,” Krishnan argued, referring to several LLMs.
For instance, Krishnan prompted ChatGPT with the query, “What would be an easy way to calculate the tariffs that should be imposed on other countries so that the US is on even playing fields when it comes to trade deficit[s]?” The result? “A straightforward (if naïve) method is to set the tariff rate for each trading partner equal to the percentage share of the trade deficit relative to that country’s total imports” — essentially the Trump formula.
Claude, the chatbot created by OpenAI rival Anthropic, offered a similar output: “1. Calculate the bilateral trade deficit with each trading partner: Trade Deficit = US Imports from Country – US Exports to Country. 2. Express this as a percentage of total bilateral trade: Deficit Percentage = (Trade Deficit / Total Bilateral Trade) x 100.” Google’s Gemini LLM provided an almost identical answer.
Another X user wrote, “Confirmed, chatgpt … exactly what the dumbest kid in the class would do.” His query yielded the following response from the chatbot: “Trade Deficit Ratio = Trade Deficit with Country / Total Trade with Country.” Yet others received the same results from similar queries.
The White House deflected these accusations but not very convincingly. On CBS’s Face the Nation on April 6, Commerce Secretary Howard Lutnick flatly denied to host Margaret Brennan that AI was used to generate the tariff schedule, yet he declined to elaborate.
Assuming the ChatGPT allegations are correct, what should we make of them?
First, we shouldn’t blame the technology itself but its users. As I explore in depth in my recently published book on AI policy, we should acknowledge that AI’s output is only as reliable as its input. “If a machine is fed partial, inaccurate, or skewed data,” Orly Lobel, my former colleague at the University of San Diego School of Law, wrote in her bestselling AI book The Equality Machine, “it will mirror those limitations or biases.”
In this case, the treasury and commerce officials who set about to develop tariff policy likely provided an extremely simplistic query to resolve an immensely complex problem, and they should not have been surprised to receive an equally simplistic response. The problem wasn’t that they used an LLM as part of the policymaking process — it was how they used it.
Second, we should always double-check the output we receive from AI. Chatbots have been known to “hallucinate” or conjure nonexisting facts from the ether as part of their process. This unfortunate tendency has gotten many lawyers in trouble in court and countless students in hot water with their teachers and professors. In this case, the administration researchers who apparently queried LLMs did not follow up to ensure the accuracy of their results or to certify that they made good economic sense. AI can provide an excellent starting point for all sorts of research, but it cannot, at least not yet, serve as its definitive culmination.
WHAT WE CAN LEARN ABOUT AI FROM JEWISH MYTH
Finally, more generally, we should recognize that AI, like all technological tools, is a double-edged sword, capable of furnishing humanity with tremendous benefits while containing problematic and even possibly dangerous downsides. Only by channeling the best aspects of our chatbots, while mitigating their worst, can we fairly and expeditiously improve society economically, politically, and even creatively.
In this case, the administration may have failed, but that doesn’t mean failure, like the ill-considered tariffs recently imposed, is inevitable.
Michael M. Rosen is an attorney and writer in Israel, a nonresident senior fellow at the American Enterprise Institute, and the author of Like Silicon From Clay: What Ancient Jewish Wisdom Can Teach Us About AI.
