As I’ve said before, AI excels at organizing tedious, repetitive material that no one would review by hand. Form ADV brochures fit that description. They’re long, dense, and there are thousands of them.

So I ran them through a local model.

I downloaded roughly 16,000 Form ADV filings filed in March, the annual disclosures registered investment advisers file with the SEC describing how they run their businesses. Of those, roughly 15,000 were standalone Part 2A brochures, covering about 12,600 advisory firms across hedge funds, private equity, venture, real estate, and traditional RIAs.

What the Filings Show

I scanned every filing for AI-related keywords. About 5,800, just under 40%, mentioned AI at all. In most cases, that meant boilerplate risk disclosures warning about potential impacts on portfolio companies or financial markets.

More than 1,200 firms described using AI in their operations. Fewer than 320 named a specific product. Of those, more than 450 firms disclosed a formal internal AI policy, and 88 said AI-related costs were being charged directly to clients or fund investors.

The detailed disclosures are concentrated at the top. The largest managers — those with more than $100 billion under management — were meaningfully more likely to describe specific AI use cases and governance frameworks than smaller firms. Building an AI governance framework requires lawyers, compliance staff, and engineers working in concert. Most smaller firms don’t have that infrastructure. Explicit references to AI driving investment decisions are less common, and usually more carefully qualified.

Where firms do describe AI in concrete terms, it’s mostly operational. Concentric Capital Strategies, with $3.1 billion in regulatory assets under management, discloses the use of LLMs, such as ChatGPT, within its investment research and business processes.

Caveat: Form ADV filings capture what firms consider material enough to disclose. Routine or limited AI use may not appear at all.

A Closer Look at the Largest Managers

I selected 100 of the world’s largest and most recognizable money managers — firms spanning traditional asset management, hedge funds, private equity, and venture capital — and scanned their 2026 ADV filings with Gemma.

Seventy-five of the 100 disclosed some form of AI use. Twenty-four named a formal governance policy. Thirteen disclosed that AI-related costs may be passed to investors.

The following is for paid subscribers and examines the firms that named governance policies and what those policies actually say, the private equity pattern of charging AI infrastructure costs to fund investors, and three large managers that went from no AI disclosure in 2025 to named frameworks in 2026.

The Governance Tier

Read more

Companies are using smaller, specialized systems because they’re easier to test, control, and generally, are more consistent.

Some examples:

• Capital One replaced a single LLM with a multi-agent system for call summaries, where agents interpret, reason, cross-check, and document the interaction before finalizing it.

• Daloopa built dozens of narrow models, each trained on structured financial data and focused on one task. “Our models have an IQ of 250 on one task and 2 on something else,” Thomas Li, CEO and co-founder of Daloopa, told me in September.

• At BlackRock, researchers broke stock screening into specialized agents—fundamentals, sentiment, and valuation—that debate and cross-check each other before reaching a final decision.

A new paper from Andrew Ang, a former BlackRock executive, Nazym Azimbayev, a sovereign wealth fund CIO, and Andrey Kim PhD, a Deutsche Bank quant, takes the BlackRock debate architecture further.

The paper, the Self-Driving Portfolio: Agentic Architecture for Institutional Asset Management, asks if autonomous driving is here, why not autonomous investing?

Their answer is a 50-agent pipeline that runs the process and documents each step of its reasoning.

Manage Email Preferences

If you prefer to receive one weekly email with all AI Street content, deselect Research and Interviews.

Manage How Often You Receive AI Street

They view running a strategic asset allocation process at an institutional level as a bandwidth problem as much as an analytical one. A CIO can supervise maybe 10 to 15 investment departments. A research team can realistically cover 20 to 30 asset classes before the process bottlenecks and an investment committee meets quarterly.

To speed this process up, they built a 50 agent pipeline that produced a documented strategic asset allocation with capital market assumptions, portfolio construction, peer review, and a board memo.

Here’s what they built

The pipeline is organized around the Investment Policy Statement (IPS), which governs the whole system the way it would govern human portfolio managers. Every agent reads it; the chief risk officer agent checks compliance for every portfolio candidate; the final output must satisfy it.

To illustrate how the pipeline works in practice, the authors ran it in March 2026 against the following mandate: 18 liquid asset classes (6 equity, 8 fixed income, 4 alternatives), a target real return of CPI +3–4%, a volatility band of 8–12%, a maximum drawdown of −25%, and a tracking error ceiling of 6% relative to a 60/40 benchmark.

• Macro agent: Classifies the current economic regime — expansion, late-cycle, recession, or recovery — using macro data, market indicators, and web searches for real-time readings. Output flows downstream to every other agent.

• Asset class agents: Agents run in parallel, one per asset class. For equity classes, each estimates expected returns using six different methods, then blends them into a seventh composite. An LLM-as-judge step reads all seven alongside the current macro regime and valuations, and selects a final estimate with explicit weights and a written rationale.

• Portfolio construction agents: 20 agents each build a portfolio using a different method, ranging from simple rules of thumb to more sophisticated approaches. A 21st researcher agent scans the academic literature and proposes methods not yet in the pipeline. A separate adversarial diversifier, one of the original 20, deliberately constructs the portfolio most different from the consensus of all the others.

• Strategy review: Each agent reviews two others — one using a similar approach, one using a different one — and all reviews are released simultaneously. Agents then vote, ranking their top five and flagging a bottom pick. Votes are combined with a performance score, and the final shortlist must include methods from at least three of the four broad categories.

• CIO agent: Combines the top candidates using seven different aggregation methods and selects the one best suited to the current environment. Produces a board memo written for non-technical stakeholders.

• Meta-agent: After each rebalancing cycle, compares past forecasts against realized returns, identifies systematic weaknesses, and updates both the code and instructions governing the other agents. All changes are logged.

  ---

Here’s what they found

• The macro agent classified the current environment as late-cycle with stagflationary risk.

• When each asset class agent settled on its return forecast, the pattern was consistent: the more expensive the market, the more the agent discounted historical estimates. US Growth stocks had their forecast cut 2.0 percentage points below the composite; US Large Cap was cut 1.1 points; Emerging Markets were barely adjusted. The agents weren’t pessimistic across the board — they were specifically skeptical of backward-looking estimates for the assets where current prices already implied low future returns.

• The same reasoning surfaced in the portfolio construction vote. In a late-cycle environment where return forecasts are uncertain, the agents collectively favored methods that lean on historical volatility and correlation data rather than return predictions. Maximum Diversification — a method that spreads risk across assets without relying heavily on return forecasts — ranked first. The portfolio that was deliberately constructed to be as different as possible from all the others came last, which was expected: its value is in the final blending step, not as a standalone recommendation.

• The final portfolio came out modestly underweight stocks (44.9% vs. 60% in a standard balanced portfolio), roughly in line on bonds (41.7%), with an 8.1% cash position. Over a backtest from 1996 to 2026, it produced nearly the same return profile as a 60/40 portfolio — but with a peak-to-trough loss of 25.6% versus 34.3%.

To be sure, this is a proof of concept, not an investment strategy. One run producing a sensible-looking portfolio doesn't tell you much given the short-time horizon.

I asked the paper’s authors about the results and received email responses from Azimbayev, who is also CEO of Altbridge, which describes itself as an AI-native hedge fund.

Read more

Manage Email Preferences

If you prefer to receive one weekly email with all AI Street content, turn off Research and Interviews here:

Manage How Often You Receive AI Street

The transformer breakthrough began in text, but researchers are adapting the architecture to other kinds of sequential data. With enough data, transformer models can learn the patterns of “language” in that dataset in ways that traditional models missed.

For example, AlphaFold is a transformer-based system trained on protein data to predict how proteins fold into their 3D shapes from amino acid sequences, which determine how they function. It effectively solved the protein folding problem. (The work later contributed to a Nobel Prize in Chemistry awarded to its creators, including Demis Hassabis, who is not a chemist.)

As I’ve written before, no one knows exactly how these models work. They’re grown rather than built, as the CEO of Anthropic likes to say. We didn’t know how aspirin worked for like 70 years, but we knew it was effective.

This brings us to a new paper from JPMorgan researchers, who trained a transformer model on market data.

The market as a language

Every buy, sell, order submission, or cancellation leaves a trace: what happened, how much size was involved, how far from the market mid-price it was placed, and when it occurred. Multiply that across thousands of stocks and millions of events per day, and you get a massive stream of sequential data.

TradeFM is a 524-million-parameter model trained on 10.7 billion training tokens drawn from more than 9,000 U.S. equities, using data spanning 368 trading days from February 2024 to September 2025.

Instead of predicting the next word in a sentence, their model — called TradeFM — predicts the next event in a sequence: its timing, size, price depth, and direction.

Trading data is messy. Stocks trade at different prices. A $5 move on a $2 stock is massive. A $5 move on one that’s $500 isn’t news.

If you feed those raw numbers into a model, it can’t really compare one stock to another, so it struggles to learn general patterns.

So the researchers adjusted the data before training. They expressed price-related features in relative terms, compressed volumes so large and small trades are easier to compare, and measured time as the gap between events.

That puts different stocks on a common scale, so moves are comparable whether it’s a $2 stock or a $200 stock.

They then discretized each event’s features and combined timing, price depth, volume, side, and action type into a single composite token. The result was a vocabulary of 16,384 trade event tokens.

Related Research

What they found

The researchers tested the model inside a simulated exchange, where it predicts trades in a continuous loop. The resulting data reproduces core patterns seen in real markets, including clustered volatility and large price swings. Across 9 stocks, 3 liquidity tiers, and 9 months of held-out data, it matched those patterns 2 to 3 times more closely than a standard baseline known as a Compound Hawkes process.

What’s most interesting is that TradeFM’s behavior extends beyond the U.S. data it was trained on. JPMorgan tested the model, without any adjustments, on trading data from China and Japan, where market structure differs meaningfully. Japan uses batch auctions at the open. China imposes 10% daily price limits. Spreads in both markets are several times wider than in the U.S. Despite those differences, the model’s performance degraded only moderately. It had never seen these markets, yet it still captured their core dynamics.

The model appears to be learning structure that carries across markets.

Subscribe now

Arman Khaledian, PhD, a former quant at Millennium and now CEO of Zanista AI, said: “That’s not a toy result. It means the model is picking up something real about how markets work at a structural level.”

Read more

The C-suite can choose what numbers to highlight — revenue per user, lifetime value, total addressable market, etc.

These selected metrics support the corporate narrative, but they’re not static. When a number is strong, it gets airtime. When it softens, it tends to quietly disappear from the script, replaced by whatever metric tells a better story that quarter.

Shifting corporate narratives happens so often it has a name: “moving targets.” This tracks the fraction of previously highlighted metrics that go missing from the next comparable earnings call. Research showed that firms with high metric turnover tend to underperform in subsequent months. The more a company reshuffles the numbers it talks about, the worse its stock tends to do.

The challenge is detection. Most approaches rely on keyword matching across transcripts, comparing terms quarter over quarter. You need to be able to link “revenue growth” to “top-line expansion.” Keyword search can’t distinguish “North America cloud revenue” from “revenue.”

At scale, this becomes difficult to track. Following the metrics that appear and disappear across thousands of earnings calls is not feasible to do consistently by hand. This is where LLMs fit: extracting and standardizing how companies describe performance over time. This is tedious work for humans and easy to scale with AI.

A group of researchers at MIT, BlackRock, and J.P. Morgan asked whether LLMs could close this detection gap.

Here’s what they did:

• Instead of scanning transcripts for predefined terms, they use an LLM to extract full phrases with context. Where keyword methods pull “revenue,” the model pulls “North America cloud revenue.” Where it grabs “dividends,” the model also captures “cash flow,” “share repurchases,” and “cash flow from operations.”

• They then compare metrics across quarters using semantic similarity rather than exact matches. Instead of forcing a binary match, they allow for an “ambiguous” range where similarity is scaled.

• They apply this across firms listed in the S&P 100 index from January 2010 to December 2024, yielding 5,615 firm-quarter observations across 64 quarters.

• To test it, they sort stocks by how much their metrics shift and compare returns, then run cross-sectional regressions with standard controls for size, valuation, and prior returns.

Here’s what they found:

Read more

The three leading AI providers, OpenAI, Anthropic and Google’s Gemini, have all released updated capabilities for Excel and are reporting higher accuracy on benchmarks.

For example, OpenAI said this month that performance on its internal investment banking benchmark jumped to 87.3% (GPT-5.4 Thinking) from 43.7% (GPT-5).

While these are material improvements, AI still can’t be relied on without supervision. “Mostly right” is not good enough.

A recent benchmark on AI in Excel points to the same issue: performance drops sharply as tasks get more complex.

FinSheet-Bench tests how models handle real-world private equity workbooks with messy layouts, multiple funds, and non-standard formatting. Across 10 models from OpenAI, Google, and Anthropic, the best result came from Gemini 3.1 Pro at 82.4% accuracy, followed closely by GPT-5.2 with reasoning and Claude Opus 4.6 with thinking, both around 80%.

On simple lookups, top models exceed 90% accuracy.

The gap widens further on large, realistic files. On the most complex workbook tested, with 152 companies across eight funds, average accuracy was worse than a coin flip.

One reason: models don’t actually “see” the spreadsheet. They operate on a text-serialized version that strips out layout, formatting, and visual structure.

So this:

Becomes this

Spokespeople for OpenAI, Anthropic and Google didn’t respond to requests for comment on FinSheet-Bench’s results.

What I’d like to see from model providers, rather than the latest numbers from internal benchmarks, is basic operating data: real error rates, how often outputs need to be corrected, and what level of reliability users should expect.

Subscribe now

ICYMI Research

The problems of AI in Excel echo the broader challenges of AI, namely that if the model has to traverse thousands of pages or dozens of tabs in Excel, accuracy plummets.

Over the last few weeks here, we’ve covered how there’s more of a premium on reliable models rather than the most powerful.

Read more

The way you write a prompt can steer a model toward different answers, even when the underlying question is exactly the same. I wrote in November:

If you asked a (human) financial analyst whether Microsoft or Apple is the better investment, the answer wouldn’t depend on whether you said Microsoft or Apple or Apple or Microsoft. For LLMs, that word order matters, according to new research.

This risk is harder to detect since an answer isn’t necessarily wrong, the model just chooses to highlight a different point.

Individual investors and, I suspect, some institutional ones as well, are likely falling for this risk by asking AI for research ideas and stock-picking guidance.

More and more retail investors are relying on AI tools, and almost three quarters of millennials do so, according to an October eToro survey.

To investigate how widespread this issue is, researchers at Auburn University and the University of Tulsa evaluated 48 large language models across investment-style decision tasks.

They presented identical financial scenarios twice, changing only how the information was framed, such as wording risk as a gain versus a loss, adding a prestigious source, or mentioning prior spending. Many of the same biases have long been documented in human investors. The difference is that AI systems can reproduce them consistently and at scale.

What they did

• Showed each model the same scenario twice: once neutral, once with a subtle wording or context change

• Tested 11 well-known investor errors, including framing, anchoring, herding, narrative appeal, and sunk costs

• Ran 25 scenario pairs per error across all 48 models

• Evaluated mitigation methods such as debiasing instructions and prompt rewriting

Results

• Framing alone moved ratings by 1.62 points on a 10-point scale, enough to flip decisions around common thresholds

• Narrative cues dominated fundamentals: describing founders as fitting a familiar archetype raised ratings by 65% or when attributing the analysis to a Nobel Laureate.

Read more

Ask it to compare two firms’ risk disclosures or track strategy over several years, and performance drops fast. Fin-RATE, a new benchmark from researchers at Yale and Goldman Sachs, measures that gap and identifies where the model breaks.

Most financial benchmarks reduce SEC filings to lookup tasks: find a number in a 10-K and repeat it back accurately. That design misses how analysts actually work. Real due diligence requires synthesizing disclosures across companies, time periods, and filing types simultaneously. A pass/fail system doesn’t tell you whether errors came from retrieval, hallucination, or broken reasoning chains.

Here’s what they did:

• Built a body of 15,311 document segments from 2,472 SEC filings (10-K, 10-Q, 8-K, DEF 14A, and others) covering 43 companies across 36 industries, 2020–2025. Sourced from EDGAR, segmented at official SEC item boundaries, converted to structured Markdown.

• Designed three task types

  • Single-document questions

  • Cross-company comparisons

  • Multi-year analysis within one firm

• Created 7,500 question-answer pairs with numbers manually verified against source filings.

• Evaluated 17 models, including closed-source systems, major open-source models, and finance-tuned variants

• Tested performance with passages provided directly versus retrieved using four RAG methods.

The findings:

Read more

Demis Hassabis is not a chemist, yet he was one of three recipients of the 2024 Nobel Prize in Chemistry. The prize recognized major contributions to the study of protein structures. Hassabis, a computer scientist who runs Google’s AI research lab DeepMind, and his fellow honoree John Jumper, who also works at DeepMind, developed an AI prediction model that the chair of the Nobel committee said fulfilled “a 50-year-old dream: predicting protein structures from their amino acid sequences.” Another committee member called it “one of the really first big scientific breakthroughs of AI.”

For decades, uncovering the shape of a single protein meant spending months, even years, of painstaking lab work and hundreds of thousands of dollars toward research and development with no guarantee of success.

With DeepMind’s deep-learning model AlphaFold2, revealing these structures takes minutes, not months. The DeepMind team trained AlphaFold2 with data from lab-determined protein shapes, along with extra examples it created on its own from patterns found in huge protein-sequence databases. The model examined protein shapes and amino acid sequences to determine the physical and evolutionary constraints dictating protein structure.

The team has since predicted more than 200 million protein structures and made them freely available, creating a global resource for scientific research.

AlphaFold2 is one in a growing list of scientific breakthroughs driven by AI. It also represents a new paradigm in scientific discovery: AI models that achieve breakthroughs in ways their creators can’t fully explain. While traditional science builds understanding through hypotheses we can test and verify, these AI systems are discovering solutions by finding patterns in data that remain opaque to human analysis.

There is currently no easy way to examine what AlphaFold2 learned about protein evolution. Its inner workings, and those of other AI systems making important contributions to science and society, remain hidden.

As these models get better, the gap between their performance and our understanding of them is only widening.

Nonetheless, AI adoption is racing ahead. Modern AI works incredibly well. The latest models can perform tasks that, 10 years ago, sounded like science fiction: generating movie-quality videos from a few lines of text, writing entire codebases for working apps, even driving cars without human input.

These advances have quickly entered our personal and professional lives. But this rapid deployment of black-box systems creates a fundamental tension in our relationship with AI: We’re becoming dependent on tools that have reasoning we can’t verify or build upon.

Even the architects of modern AI admit to being troubled by their lack of insight.

Dario Amodei, a cofounder of the AI lab Anthropic and the company’s CEO, wrote in April 2025: “People outside the field are often surprised and alarmed to learn that we do not understand how our own AI creations work. They are right to be concerned: this lack of understanding is essentially unprecedented in the history of technology.”

This has made interpretability, the science of cracking open AI’s “mind,” a pressing priority, and a new wave of research is taking a novel approach. AI-interpretability research has long been a form of detective work done after an AI system has already been trained and deployed. By then, the AI has already “decided” which data matter most when making its predictions.

Instead of trying to work backward to understand AI models after they’re built, scientists are now using new research frameworks to build interpretability into the training process from the start—a notion considered impossible just a few years ago.

Reverse engineering AI

When researchers try to parse the reasoning of an AI model after it has already been fully developed, they are essentially trying to reverse engineer a system that, in many ways, built itself, and attempting to uncover the internal patterns and definitions it formed along the way.

“People think these things are built systems, but they’re really not built per se,” says Ted Sumers, a researcher at Anthropic. “It’s much more like growing a plant than building a building.”

Understanding how a model “grows” has become a central focus for researchers.

One branch of this work, called mechanistic interpretability, maps which neurons activate when a user asks AI a question, and traces how information flows through the network’s intricate layers.

Anthropic, a rival to OpenAI, has been at the vanguard of this type of approach, dissecting neural networks by studying the roles of individual neurons and circuits.

This has yielded practical results. Teams can, without damaging overall performance, identify and remove specific circuits that lead to biased or unwanted outputs. They can also locate the exact parts of a model that enforce safety rules—like refusal to answer harmful queries—and adjust those directly. Since the techniques go down to the neuron level, they offer a way to audit whether a model is memorizing sensitive data. Together, these advances make models easier to edit, test, and trust as they continue to grow more capable.

Still, it’s like peering into a house through a keyhole.

The struggle to understand AI

Unlike traditional software, which relies on top-down, hard-coded rules, a neural network—a type of artificial-intelligence model that’s often described as resembling the structure of a human brain—learns from the bottom up, ingesting training data and making internal adjustments based on what it observes. Such models learn patterns from massive datasets, some with trillions of data points.

For example, to learn to identify pictures of dogs, a neural network reviews millions of labeled images of the animals rather than relying on a fixed set of definitions.

During training, the model guesses what each image shows and compares its answer to the correct label. If it guesses “not dog” for an image labeled “dog,” it recognizes the mistake and adjusts its internal settings to reduce the error. This process repeats again and again.

After enough examples, it becomes very good at identifying dogs. But it does so in a way that’s fundamentally different from how humans process and recall information. AI relies on statistical analysis to identify patterns, rather than mental imagery.

AI doesn’t “see” the way we do.

It sees the world through numerical representations of data. All types of data that AI works with—whether text, images, or audio—are converted into numbers that the system can mathematically manipulate. For example, the sentence, “AI sees the world through numerical representations of data” is converted into: [17527, 27432, 290, 2375, 1819, 57979, 63700, 328, 1238] according to OpenAI’s tool, which displays how a piece of text might be tokenized by a language model. (Different models tokenize the same words differently.)

Turning data into strings of numbers makes them usable by AI models. Computers may not be able to see or read in the traditional sense, but they can run mathematical operations on numbers. That’s how AI detects patterns, compares inputs, and ultimately learns from data instead of relying on fixed rules.

These numbers aren’t stored in a database or an Excel spreadsheet. They exist in what’s called a high-dimensional space.

We can visualize and understand the difference between two and three dimensions. Schoolchildren are taught that a rectangle has two dimensions—length and width. A cube adds another dimension: depth. It’s much harder for us to grasp a fourth dimension.

But AI can understand hundreds, even thousands, of dimensions.

To navigate these vast high-dimensional spaces, a model learns during training which direction matters most by adjusting its internal weights—think of them as groups of dials that turn up or down to chart a course through this mathematical terrain. As training progresses, turning up weights for “furry” and “four legs” steers it deeper into dog country, while dialing down irrelevant features such as “fire hydrants” prevents the model from wandering into dead ends.

Through training, the model groups together the features that typically appear in pictures of dogs without being told what exactly a dog is. However, there is no index, as you might find in the back of a book, that you can consult to find the exact “dial” or weight corresponding to doglike features; those features are intertwined across the model’s complex architecture. Researchers have to go find them.

This gets at the core challenge of AI interpretability. Researchers know how to build and train these models. But they often can’t see what, exactly, in an image causes the model to adjust one specific dial out of billions.

Understanding how a model makes its predictions can help illuminate how much we can trust it—or, if necessary, how to fix it when its behavior deviates from what we want or expect.

Continue Reading at CBR

Many AI + investing research papers suffer from the same problem: the models were trained on historical internet data that often contains the outcomes they are asked to predict. Ask a model today what happened to a stock in 2022, and it may already know.

This look-ahead bias makes me skeptical of many papers with “big” conclusions. But since the models have become ubiquitous, researchers can test their theories in real time.

Two Peking University researchers, Zefeng Chen and Darcy Pu, did just that. They ran a live, nine-month experiment asking a frontier AI model to pick stocks every night across the Russell 1000.

Here’s what they did:

• Every night from April 2025 through January 2026, they queried a leading U.S. frontier AI model via its web interface with live search enabled, with no pre-selected news or filings fed to the model. The model autonomously searched the web, synthesized what it found, and returned a score (−5 to +5) for each Russell 1000 stock.

• Signals were generated after the 4pm close and before the next open. Portfolios were entered at the opening auction and exited the following open.

• They ranked about 1,000 stocks by the model’s daily score, built a portfolio of the top 20 weighted by market value, and tested its performance using standard factor models.

Here’s what they found:

Read more

I’ve thought of this quote often since interviewing Tharsis Souza, now at Citadel, in a podcast episode in January last year.

Souza presciently highlighted the diminishing costs of expert analysis. (Check out the recording here.)

I’ve seen more and more examples of this, in my own life, like trying to sort out a hairy Italian-U.S. tax issue with the help of ChatGPT.

More empirically, studies are coming out showing how much investors are using AI to broaden their investing universe.

When Italy Banned ChatGPT, Retail Traders Narrowed Their Bets

In March 2023, Italian regulators forced OpenAI to suspend service over data privacy concerns.

This 28-day blackout created a natural experiment, revealing how the sudden loss of a leading AI tool—available to neighboring countries but denied to Italians—altered trading behavior.

The Study

A research team led by Omri Even-Tov (UC Berkeley) analyzed granular, account-level data from the brokerage platform eToro.

• The Scope: 3 million accounts across 100+ countries, covering stocks, crypto, ETFs, and commodities. The paper’s final analysis sample consists of 24,185 investors and 169,295 investor-month observations from Italy and neighboring control countries between January and July 2023.

• The Method: They compared Italian investors (the treatment group) against peers in France, Switzerland, Austria, and Slovenia (the control group).

The Key Finding: Narrowed Horizons

The researchers used the Herfindahl-Hirschman Index (HHI) to measure trade concentration. While the average retail investor typically sticks to 2–3 assets, the ban caused Italian portfolios to shrink even further.

• The Spike: During the ban, Italian trade concentration rose by ~3.1% relative to the control group.

• The Proof: The data showed a “dynamic” shift—the concentration spike appeared only during the month of the ban.

• The Rebound: Once the ban was lifted, Italian investors did not immediately return to their pre-ban behavior, even showing a brief “overcorrection” as they explored new assets to make up for lost time.

The ban also reduced the likelihood that investors would initiate trades in assets they hadn’t previously touched. The odds of adding a new asset declined by roughly 10.3% during the ban. And trading shifted toward popular assets (the top 50, top 100, top 200 by aggregate retail volume) while pulling away from the long tail.

Not Just Behavior — Prices Moved Too

Even-Tov’s study captures what happened at the account level. But a separate team had already used the same Italy ban to measure something different: what happened to stock prices.

A separate study used the same Italy ChatGPT ban to examine market-level effects. The authors find that Italian firms with higher exposure to generative AI underperformed by roughly 9 percent during the ban. Bid-ask spreads widened, and analysts based in Italy issued fewer forecasts.

That’s two studies exploiting the same 28-day shock, measuring different outcomes, and arriving at the same directional conclusion: removing access to GenAI compressed both participation and market quality.

Where the Effect Was Strongest

Researchers sliced the data by asset class, investor characteristics, and firm characteristics.

By asset class, the results were driven by stocks and cryptocurrencies. ETFs, commodities, and currencies showed no significant effect. That’s consistent with the information processing story: stocks and crypto involve large volumes of unstructured, or hard-to-verify information. ETFs and commodities track standardized benchmarks. There’s less for an AI tool to help with.

By investor type, the effect was strongest among investors with lower income or less trading experience (consistent with those investors facing higher information processing costs) and among investors in technology-related occupations or students (consistent with higher likelihood of actually using ChatGPT in the first place).

By firm characteristics, the decline in trading activity was steeper for hard-to-value stocks. Firms with lower profitability or higher volatility saw the largest drop in Italian investor participation during the ban. The effect also intensified around earnings announcements, when investors face concentrated releases of financial data.

The Portfolio Consequences

Higher trade concentration has downstream effects. During the ban, Italian investors’ portfolios exhibited more return comovement across holdings and higher portfolio volatility. Fewer distinct positions, more overlap in the stocks they held, less diversification.

One thing to note: the ban had no detectable effect on risk-adjusted trading performance. Investors weren’t making worse picks during the ban. They were just making fewer of them, and those picks looked more alike. GenAI appears to broaden the set of assets retail investors consider, but there’s no evidence (at least in this 28-day window) that it helps them choose better ones.

A Pattern, Not a One-Off

Read more