资产管理中的生成式AI：AI是否让市场更高效？

Gen AI in Asset Management: Is AI making markets more or lessefficient? “Is AI going to make markets more efficient or less efficient?”This is often a question,we get from investors. In this note, we show evidence to present both sides of the argumentand find that overall AI is likely to improve micro-efficiency — faster information processing,better execution, broader data coverage; but can worsen macro/market-structureinefficiency through crowding, concentration, and sharper narrative-driven reversals. Webelieve AI would create lower average inefficiency but higher tail inefficiency. Rupal Agarwal+65 6326 7641rupal.agarwal@bernsteinsg.com Cheng Zhang, CFA, CQF+852 2123 2636cheng.zhang@bernsteinsg.com Evidence for AI improving efficiency:The case for AI making markets more efficientrests on its ability to process vast amounts of information faster, more consistently, andat lower cost than humans. Early evidence is broadly supportive: Italy's 2023 temporaryChatGPT ban reduced information production and widened bid-ask spreads, while sinceGenAI's launch, EPS forecast dispersion among many large-cap/well-covered stocks hasfallen, suggesting lower information asymmetry. AI is also accelerating price discovery:IMF shows post-LLM equity price reactions to Fed minutes aligning more quickly withlonger-term price moves, while S&P 500 earnings surprises have steadily declined asanalyst forecasts converge faster toward actual outcomes. AI driven productivity gain(c.40% time save) is allowing broader coverage which is likely to compress the traditional‘neglect premium’. In the last 1yr there has been 28%/23% jump in coverage of EM small-caps/India mid-caps. Gen AI agents (democratized access vs. once available only to quantinvestors) are expected to mitigate behavioral inefficiencies as shown in a recent paper-AIagents made rational decisions 61%–97% of the time vs. 46%–51% for humans. Case for AI increasing inefficiency:AI could make markets more inefficient throughherding behavior, systemic fragility, synthetic information shocks and reduced competition.More Gen AI usage raises the risk of signal homogenization and consensus positioning.The August 2024 Nikkei sell-off, driven in part by the unwinding of crowded carry tradesand systematic strategies, highlights these risks. The IMF's 2024 Global Financial StabilityReport also warned that AI could heighten systemic vulnerabilities through synchronizedtrading, model concentration, and cyber risks. AI-generated misinformation introducesa new source of market inefficiency as seen during the fake Pentagon explosion imagethat temporarily moved US equities in 2023. Recent academic work from HEC Paris andother studies suggest AI trading agents can learn collusive-behavior without explicitcommunication, potentially reducing competition, liquidity and price discovery. The Reflexivity problem—Is AI itself the inefficiency?:As investors increasingly relyon similar AI models, model outputs influence trading activity, which alters prices andsubsequently reinforces future model signals. Rather than simply reducing mispricing, AIcan create self-reinforcing feedback loops that amplify momentum, crowding and valuationextremes. This mirrors George Soros' theory of reflexivity, where perceptions influencereality and reality subsequently reinforces perceptions. This is so evident in markets today-more AI is being used by market participants while AI is also the dominant narrative. Thishas led to unprecedented market concentration and correlation. Momentum trade hasreached extreme valuations and crowding, seeing record high bullish analyst sentiment andproducing 25yr high factor dispersion. (see here, here, here). DETAILS AI is also accelerating price discovery, with studies showing faster incorporation of news into asset prices and decliningearnings surprises as forecasts converge more quickly toward outcomes. At the same time, AI-driven productivity gains—savingroughly 40% of analyst time and expanding coverage of under-researched segments such as EM small-caps and India mid-caps—could compress the traditional neglect premium. Finally, AI agents appear less prone to behavioural biases than humaninvestors, with recent research finding substantially higher rates of rational decision-making, potentially reducing behaviouralinefficiencies and democratizing capabilities once available primarily to quantitative investors. Over the last 12months, we have had multiple conversations with CIOs and PMs on how Gen AI is, and would impact assetmanagement industry. We have covered multiple aspects of this topic in our previous work focusing on key use-cases forinvestors, AI tools for investors, best practices emerging at large asset managers, key concerns and risks, how AI-native assetmanagers would emerge by 2030, expected AI spend by asset managers and the impact of AI on the industry efficiency gains/margin improvement. We have also touched upon some long-term impact of AI on the broader