A Goldman Sachs partner spent months building the bank's AI platform. Then he went on the record to warn it could be quietly wrecking the reasoning skills of the very bankers using it.
Chris Churchman, who leads Goldman's institutional platform Marquee and co-chairs the bank's Global Banking and Markets AI working group, made the case on the firm's own Exchanges podcast.
The Cognitive Atrophy Problem
"There's a huge danger here that in the era of AI, we outsource our reasoning to these models, and we have cognitive atrophy that stops us being able to reason from first principles ourselves," Churchman said.
The concern isn't abstract for Goldman specifically. AI adoption makes the bank more profitable today by handling routine work faster.
That's exactly the work that has traditionally taught junior bankers how to think, structure an argument, and make a call under pressure.
"Reasoning is still important," Churchman said. "You still need to reason about problems and structure it into an argument, and now we're delegating reasoning."
Wall Street has already been testing how far that delegation can go. CNBC reported last year that firms were exploring ways to shrink the ratio of junior bankers to senior staff using AI, a trend that would only accelerate if reasoning skills stop being something junior employees need to develop at all.
How Junior Bankers Actually Learn
Churchman, who ran currency trading at UBS before joining Goldman in 2021, described the apprenticeship model AI now threatens to bypass entirely.
"You learn by doing, and a lot of knowledge is tacit, it was never written down," he said.
Junior traders build real judgment by fielding client pricing requests under the direct supervision of experienced risk takers. Automating that step is technically straightforward.
"We can absolutely automate that," Churchman said, "but then do we get the senior traders that fully understand?"
Even Goldman, with its own dedicated AI working group, hasn't resolved that tension yet. Churchman said the bank still hasn't figured out how to manage the transition, and that systems need to be designed so employees keep making the final call on high-stakes decisions rather than becoming passive operators.
Even Goldman's Own AI Admitted Its Limits
Churchman's most striking comment wasn't about people. It was about the AI system itself.
Marquee's AI features, still limited to Goldman employees rather than clients, are held to a standard consumer chatbots never face: answers need to be verifiably accurate and auditable, not just plausible. During development, Churchman said, the team pushed the system hard enough that it volunteered its own limitation unprompted.
"When we challenged it hard, at least it was honest," Churchman said. "It was like, 'Look, in the end, I'm better at sounding thorough than being thorough.'"
That's essentially the same calibration problem researchers at McGill University are trying to solve with AI systems that can flag their own uncertainty, a story MAIN covered last week, playing out inside one of the world's largest investment banks.
What This Means for Miami
Miami's growing base of financial firms, family offices and fintech companies are adopting AI for research, analysis and client-facing tools at the same pace as Goldman, often with far fewer resources to build the kind of internal guardrails Churchman describes.
Reasoning skills built through years of hands-on work are exactly what junior finance professionals in Miami are supposed to be developing right now. Any firm here handing that work entirely to AI without a deliberate plan is making the same bet Goldman is still trying to figure out how to manage, without Goldman's balance sheet to absorb it if the bet goes wrong.