Here’s a problem if you’re hoping a technology-driven productivity upturn will soon supercharge the American economy: Artificial intelligence continues to attract capital more quickly than actual users. Goldman Sachs’s latest adoption tracker shows corporate use of generative AI crept up to 9.7 percent of US firms in the third quarter, only a modest rise from 9.2 percent three months earlier. Financial and real-estate companies are adding AI fastest, while education has pulled back. “Broadcasting and publishing firms reported the largest expected increase in AI adoption over the next six months,” the bank notes.
Investment tells a different story, however. Since ChatGPT’s debut in late 2022, analysts have boosted their end-2025 revenue outlook for chipmakers by $203 billion, roughly 0.7 percent of US GDP, and for other hardware suppliers by an additional $123 billion, or 0.4 percent of GDP. Moreover, shipments of AI-related components have climbed in the United States, Japan, and Canada, with part of the recent surge tied to tariff front-loading.
(I would add that investment bank UBS reckons companies will splurge $375 billion on AI infrastructure this year, rising to $500 billion in 2026. Software and computing gear accounted for a quarter of America’s economic growth in the second quarter, according to The New York Times. In other words, the largest technology investment cycle since the 1990s internet boom continues apace.)
The result of these two trends is a now-familiar mismatch. Companies are buying the components of AI infrastructure but remain hesitant to weave it into daily operations. GS: “Recent industry surveys suggest that concerns around data security, quality, and availability remain the top barriers to adoption.”
Perhaps that insight helps explain why larger firms, those with 250 or more employees and medium-sized firms with 100 to 249 workers, actually saw adoption rates tick down (but expected future adoption to remain strong). Also hanging over this phenomenon: A recent MIT study found 95 percent of organizations saw no AI investment return.
My speculation for that result: Bigger companies especially might be more risk-averse to large language models generating hallucinations. If so, a new paper from OpenAI, “Why Language Models Hallucinate,” offers a diagnosis and possible treatment for these confident falsehoods. Like students guessing on multiple-choice exams, LLMs are rewarded for bluffing when unsure. Evaluations penalize an “I don’t know” response more than a plausible but wrong answer. That misaligned incentive sustains overconfidence and undermines corporate trust. “This creates an ‘epidemic’ of penalizing uncertainty and abstention, which we argue that a small fraction of hallucination evaluations won’t suffice,” the authors explain.
So what’s the fix? The research team suggests the path forward is less about inventing new tests than about changing the rules of existing ones. These “tests” are the benchmarks and leaderboards—like multiple-choice accuracy exams—that developers use to measure whether a model is improving. Right now, they mostly reward a confident answer and give zero credit for an “I don’t know.”
As the paper puts it, “Simple modifications of mainstream evaluations can realign incentives, rewarding appropriate expressions of uncertainty rather than penalizing them.” If models were credited for saying “I don’t know” when unsure, rather than punished for it, they would have far less incentive to bluff. That shift could help turn AI into a more reliable colleague—and give firms more confidence to deploy the tools they are already buying.
Until then, corporate caution looks rational. Firms can hardly be blamed for waiting until the technology behaves a bit more like a trusted coworker than a smooth-talking guesser. Yet as a believer in the notion that today’s AI is the worst it will ever be, I think that C-suite caution will give way to a more aggressive embrace.
