Are markets efficient? i.e. do prices reflect all available information? The Grossman–Stiglitz paradox is that prices can only become informative if some traders first pay to acquire information and then trade on it. Through market clearing, those trades help reveal information in prices.

But if prices already reflected all information perfectly, uninformed traders could simply look at the price and infer that information for free. Then nobody would have an incentive to pay for research, data, or analysis.

If nobody becomes informed, nobody trades on information. And if nobody trades on information, prices cannot fully reflect it.

Therefore, markets cannot be perfectly efficient when information is costly. The equilibrium is one of partial efficiency: prices are informative enough that informed traders earn no excess profit after costs, but not so informative that becoming informed has no value.

More formally, let the fraction of informed traders be $\lambda$, and let $G(\lambda)$ is the gross benefit of being informed over uninformed traders, and $C$ is the cost of being informed, which includes data, research, analysis, infrastructure, time, etc.

As $\lambda$ increases, prices become more informative, the benefit of being informed reduced $G(\lambda)$ decreases, and vice versa, therefore equilibrium occurs at $\lambda^{*}$ where the marginal trader is indifferent between being informed and not informed: $G(\lambda^{*}) = C$.

Where does $\alpha$ come from? a given trader can still have lower costs, better data, better skills and models, better risk tolerance, balance sheet, liquidity, better infrastructure and execution, such that the net alpha for trader $i$, $G_i(\lambda^{*}) - C_i > 0$ is positive.

How does AI change things? AI lowers the cost of processing public information, research, and analysis, so more traders become informed, and prices become more informative, which competes away the old information edge. AI creates a new equilibrium: $G(\lambda_{\rm AI}^{*}) = C_{\rm AI}$. The alpha shifts to whatever is still scarce or hard, things not everyone can do even with AI, i.e. as before, better skills, private data, execution, better judgement, risk tolerance, etc.

Original paper: https://www.aeaweb.org/aer/top20/70.3.393-408.pdf