By Dave DeFusco
A by Pablo Hernandez-Lagos, director of the MBA program in the Sy Syms School of Business, has found that AI companies can sometimes boost profits by keeping their technology less transparent, even if greater openness would increase user adoption.
To understand why, it helps to start with how modern AI works. Systems like chatbots or recommendation engines improve by learning from users. Every interaction鈥攅very prompt, click or correction鈥攂ecomes data. The more people use a system, the more data it collects and the better it becomes. Economists call this a 鈥渇eedback loop鈥: more users lead to more data, which leads to a better product, which attracts even more users.
At first glance, this suggests that companies should want users to fully understand their technology. After all, if people trust and understand a product, they鈥檙e more likely to adopt it. Hernandez-Lagos shows that this intuition can break down, especially for companies that rely heavily on investors.
In his model, firms face a balancing act. On one side are users, who generate data and revenue. On the other are financial investors, who fund the company and shape its valuation. Investors don鈥檛 directly see how good the technology really is. Instead, they infer its value from signals, especially how much the company spends developing it. That鈥檚 where things get complicated.
鈥淲hen you open the newspapers and see companies spending $100 billion on AI, it鈥檚 puzzling,鈥 said Hernandez-Lagos. 鈥淭hese firms are spending more than entire countries on research and development. That made me question what鈥檚 really driving those decisions.鈥
In theory, spending should reflect genuine progress, but, in practice, it also serves another purpose: signaling. Big spending can convince investors that a company鈥檚 technology is valuable, even if its true potential is still uncertain. Hernandez-Lagos calls this effect 鈥渕anipulation power.鈥
鈥淏ecause we don鈥檛 know the value of the technology yet, firms can use their spending as a signal,鈥 he said. 鈥淚f they spend a lot, people might believe the technology is better than it actually is.鈥
This creates a subtle incentive. If users understand the technology very well, they adopt it with more confidence. That increases data and revenue, but it also forces the company to spend even more to meet expectations and maintain its image with investors. In some cases, those extra costs grow faster than the benefits. The result is counterintuitive: better-informed users can actually reduce profits.
鈥淭he key mechanism is that more transparency leads to more adoption,鈥 said Hernandez-Lagos, 鈥渂ut that also amplifies spending to a level that can hurt the bottom line. The extra revenue can be offset, or even outweighed, by the extra costs.鈥
To make this concrete, he offers the simple analogy of a job interview. Candidates often put in enormous effort to impress employers鈥攆ar more than they might sustain on the job itself. Companies, he argues, behave in a similar way with investors: they 鈥渙verperform鈥 through spending to signal quality. That pressure creates what he describes as a kind of arms race not necessarily with competitors, but with expectations.
鈥淐ompanies would be better off if they could coordinate and just spend what鈥檚 needed,鈥 he said. 鈥淭hey don鈥檛, though, because each one wants to look like the leader in the eyes of investors.鈥
This dynamic also helps explain a broader trend in the AI industry: why many powerful systems remain opaque. Despite widespread use, the inner workings of large AI models are often difficult to fully understand even for experts. This isn鈥檛 just a technical challenge, according to Hernandez-Lagos, it鈥檚 an economic choice.
鈥淚f the technology is very uncertain, firms may avoid making it fully transparent,鈥 he said. 鈥淭ransparency can tie them to higher spending, because they have to live up to the expectations it creates.鈥
His model shows that this lack of transparency isn鈥檛 mainly about hiding secrets from competitors or avoiding regulation. Instead, it stems from how companies interact with financial markets. The need to impress investors can distort decisions about how much to explain and how much to spend.
Hernandez-Lagos suggests that if companies rely less on investor money and more on revenue, or on public funding, they may feel less pressure to signal through excessive spending.
鈥淧ublic institutions like universities could help develop and validate the technology,鈥 he said. 鈥淭hat would reduce the need for firms to prove its value through massive expenditures.鈥
Another approach is for companies to rethink their 鈥渃apital structure鈥濃攖he mix of funding from investors versus customers. Finding the right balance, he said, could reduce wasteful spending and encourage greater openness.
鈥淭he stakes are high,鈥 said Hernandez-Lagos. 鈥淎s AI becomes more central to the economy, understanding these incentives matters not just for businesses, but for society. Calls for transparency are growing louder, especially as people worry about the risks of systems they don鈥檛 fully understand.鈥
