Did an AI chatbot help draft the US tariff policy?

But whether Team Trump used AI or not, asking the likes of ChatGPT and others to come up with the plan does elucidate the situation in some interesting ways.

The road test

I submitted the following prompt to a number of chatbots:

Please come up with a formula that the US government could use to impose tariffs on each nation. The goal is to put the US on an even footing when it comes to trade deficit.

Google’s Gemini immediately cautioned that “designing a formula that is economically sound, fair, and doesn’t trigger harmful retaliations is incredibly complex”, and even though it provided a formula, it additionally worried that it was “highly simplified and potentially problematic”. The formula finds the difference between imports and exports, and expresses it as a percentage of total US imports.

Gemini then delivered a very long and detailed explanation of why the wording of my question was problematic, and why implementing the plan was dangerous. Humorously, it suggested a better strategy would be to focus on US competitiveness by investing in education and infrastructure, while working constructively with other nations to address economic imbalances.

DeepSeek went further, suggesting the same base formula but adding additional penalties for undervalued currencies and for exports that “exploit weak labor/environmental standards”. That way, it said, the nations engaging in unfair practices would be hit hardest but there would be ways for them to reduce the tariff through negotiating. It did warn US consumer prices would rise.

ChatGPT again suggested a similar base formula, with an adjustable level of aggression and a global correction factor “if the overall trade deficit is persistent”. It noted that its formula meant that balanced or surplus countries face would face no tariffs, so to get to Trump’s calculations I would have had to ask for a 10 per cent base level for all nations.

ChatGPT also gave a long list of reasons the formula would not work, explicitly advising that the US would be shooting itself in the foot, and handily summed the implications up as: higher consumer prices, damaged trade relationships, legal blowback, and slower economic growth. I asked for mitigation ideas, it listed an upper limit on tariffs, a gradual phase-in, exemptions for critical goods or those not available in the US, carve-outs for allies and reinvestment of tariff revenue. The Trump administration is not adopting any of those.

So it seems likely that even if the Trump administration did use AI, it took the formulation and ran with it despite the chatbot itself spelling out why that would be such a bad idea. Krishnan wrote that asking language models about governance might not be a bad idea in absolute terms, but that this case pointed to a lack of chatbot literacy; the user asked a bad question filled with wrong assumptions, then ran with the answer ignoring the qualifications.

He called it “vibe governing”, a spin on the recently coined phrase “vibe coding”, in which a user describes a desired output to an AI and lets it do the coding.

Loading

RMIT’s Dr Samar Fatima said that directly using the output of an AI chatbot to craft public policy design or governance could have lethal results, and that the responses from large language models (LLMs) — broad and based on data indifferently scraped from the public internet — were not reliable enough for government use.

“There are so many factors which are contextual, which need that human insight, which have to cover those small nuances of a country’s economy, the geographical position, the political environment, the overall international trade environment,” she said.

“An LLM will not be able to comprehend those unspoken factors, which are there but they are not quite published, or part of the data set.”

So could the Trump administration have taken a chatbot’s word for it and tanked the global economy by accident? It’s impossible to know. And with AI advancing so quickly Fatima said that regulation was unlikely to catch up, but that changes which obliged policymakers to disclose AI use could help mitigate some of the worst impacts.

“In terms of transparency, AI systems are still a black box. And if the output’s used in a system where it is not even disclosed that it was generated by AI, then the black box goes to another level of blackness,” she said.

“Then we cannot even really figure out how the decision was made, while it’s affecting the lives of billions.”

Get news and reviews on technology, gadgets and gaming in our Technology newsletter every Friday. Sign up here.