AI can already read your portfolio, generate investment ideas and, in some products, initiate transactions. But “AI investing” spans radically different legal and operational models. The real question is not whether software can click Buy. It is who defines the mandate, who checks the model, who carries the fiduciary or best-interest obligation, and how much damage a bad decision can cause before a human stops it.
The Core Question Is Not Whether AI Can Trade
A retail investor can already hand a portfolio to software. Robo-advisers have done versions of this for years: collect a risk profile, map the customer to a portfolio, invest the account and rebalance it algorithmically. The SEC describes robo-advisers as registered investment advisers that use computer algorithms to provide advisory services, often with limited human interaction. In other words, software making portfolio decisions is not legally exotic by itself.
What is new is the arrival of large language models and agentic interfaces that can understand a natural-language instruction, inspect account data, retrieve market information and call transaction tools. Robinhood Cortex is a useful boundary case. Its 2026 agreement says the assistant can interact with an account and execute certain transactions using natural-language commands, but its disclosure simultaneously states that Cortex is not an investment adviser and does not provide investment recommendations. The AI can help operate the account; the customer still owns the investment decision.
OpenAI draws the line even more clearly in its current consumer finance product. ChatGPT can review connected financial accounts and portfolio allocation, but OpenAI says it cannot make trades, move money or act as a registered investment adviser. So the market already contains at least three different things that are casually called “AI investing”: AI that explains, AI that helps execute, and regulated automated advice that actually manages a portfolio.
Figure 1. Dave Finances framework. “AI investing” ranges from information tools to discretionary portfolio management; the legal responsibility does not move automatically with the model.
The Legal Line Sits Around Advice, Discretion and Accountability
The key regulatory point is technology neutrality. A firm does not escape securities law because the recommendation came from a neural network rather than a human adviser. FINRA’s 2026 regulatory oversight report says its rules and the securities laws continue to apply when firms use generative AI, including obligations involving supervision, communications, recordkeeping and fair dealing. The SEC takes a similar approach to automated advice: its Form ADV guidance says automated advice is a means of providing an advisory service, not a separate category of service.
That distinction matters when an AI moves from describing a stock to deciding that a customer should own it. Registered investment advisers owe clients a fiduciary duty under the Investment Advisers Act, including duties of care and loyalty. The SEC’s 2019 interpretation says an adviser must serve the client’s best interest and cannot place its own interests ahead of the client’s. A broker-dealer making a recommendation to a retail customer instead operates under Regulation Best Interest and related obligations. The software may generate the output, but the regulated entity remains responsible for the process around it.
This is why “the AI chose it” is not a complete compliance answer. If a regulated firm delegates parts of research, allocation or trade generation to a model, it still needs a defensible method for determining what the model may do, what data it may use, how its outputs are supervised and how conflicts are handled. The model has no balance sheet, license or fiduciary status of its own.
| Product model | Does AI choose investments? | Can it transact? | Who authorizes the economic decision? | Typical regulatory character |
| General AI chatbot | May generate ideas | No, unless separately integrated | User | Information tool; not necessarily an adviser |
| Broker AI assistant | May analyze/prepare actions | Sometimes | User confirms or directs | Brokerage rules; recommendation rules may matter |
| Traditional robo-adviser | Algorithm selects/allocates | Yes | Client grants advisory mandate | Registered investment adviser |
| Fully discretionary AI adviser | Yes | Yes | Client grants discretion to regulated adviser | Adviser fiduciary regime plus model governance |
Table 1. The same “AI” label can describe very different authority and legal relationships.
The Harder Question: Is the AI Actually Good at Investing?
The empirical answer is uncomfortable for anyone selling “AI portfolio management” as a solved problem: results are mixed, highly sensitive to model design and often difficult to separate from ordinary factor exposure. One 2025 Finance Research Letters study found ChatGPT could construct portfolios that reflected different investor risk appetites. A 2026 Journal of Financial Economics paper found GPT-based news scores could predict some subsequent return drift, particularly in smaller stocks and negative-news settings. Research on systematic investing has also found value when an LLM is used as an additional signal inside an existing quantitative process.
But other work is much less flattering. A 2025 experiment covering 30,000 simulations, 1,522 U.S. companies and 20 years of data found Gemini 1.5 Flash did not consistently outperform a naive portfolio or the S&P 500 on returns and Sharpe ratios. A separate large-scale study of ChatGPT, Gemini and Copilot found that LLM-generated advice systematically increased several kinds of portfolio risk, including concentration, trend chasing and expense exposure; ChatGPT and Gemini did not deliver statistically significant Sharpe-ratio improvements over the benchmark in the reported performance test.
Even apparently strong results deserve a second pass. A retrospective U.S. study found positive abnormal returns from GPT-4 stock picks over long historical samples, but the paper also found that portfolios tilted toward large growth stocks and that abnormal performance was inconsistent by year. Another six-month test found ChatGPT portfolios beat the S&P 500 in raw returns, yet their Fama-French alphas were not statistically significant. Outperformance that comes from taking more growth, sector or market risk is not the same thing as a durable forecasting edge.
| Evidence | What it suggests | What it does not prove |
| LLMs can classify news and sometimes predict short-term drift | Language models may extract signals from unstructured text | That a general chatbot can run a complete portfolio safely |
| LLMs can produce risk-differentiated portfolios | Models can translate risk prompts into different asset mixes | That the chosen weights are optimal or stable across regimes |
| Some backtests show outperformance | There may be exploitable signal or factor exposure | Persistent live alpha after costs, taxes and crowding |
| Large simulations show inconsistent benchmark beating | Model intelligence alone is not enough | That AI has no useful role inside a supervised investment process |
Table 2. The literature supports a hybrid conclusion: useful signal extraction is plausible, but autonomous portfolio superiority is not established.
The Hidden Cost of “Smarter” AI Is Turnover
An AI does not need to hallucinate a company to damage a portfolio. It can simply trade too much. The more frequently a model revises its view, the more alpha it must generate merely to offset spreads, slippage, market impact and, for taxable accounts, potential tax consequences. Commission-free trading makes this less visible because the cost is not necessarily shown as a ticket fee.
Consider a $10,000 portfolio. Suppose an AI strategy turns over the equivalent of the whole portfolio twice, six times or 12 times per year. Using an illustrative all-in friction of 10, 25 or 50 basis points for each 1x of portfolio turnover, annual drag ranges from just 0.2% at 2x turnover and 10 bps to 6% at 12x turnover and 50 bps. At the latter setting, the AI must produce more than $600 of additional annual gross return on a $10,000 account merely to break even on modeled trading friction. That is before any advisory fee or tax drag.
Figure 2. Illustrative Dave Finances calculation: annual drag = portfolio turnover × assumed all-in trading friction. This is a cost hurdle, not a forecast of actual brokerage costs.
| Annual turnover | 10 bps friction | 25 bps friction | 50 bps friction |
| 2x | 0.20% ($20) | 0.50% ($50) | 1.00% ($100) |
| 6x | 0.60% ($60) | 1.50% ($150) | 3.00% ($300) |
| 12x | 1.20% ($120) | 3.00% ($300) | 6.00% ($600) |
Table 3. Illustrative annual friction on a $10,000 portfolio. The model excludes taxes and advisory fees.
Permission Design May Matter More Than Model Intelligence
The safest architecture for AI investing is not “make the model perfect.” It is “make any single model error survivable.” This is the same principle used in payments and cybersecurity: constrain authority, require confirmation for high-risk actions and keep a deterministic control layer outside the probabilistic model.
Imagine an AI mistakenly buys an asset that falls 20% before the error is noticed. If the agent can allocate the entire $10,000 account to one new position, the modeled loss is $2,000. If the execution layer enforces a 20% maximum new-position size, the same market move costs $400. At a 5% cap, it costs $100. The investment thesis has not improved; the system design has limited the blast radius.
Figure 3. Illustrative scenario: a $10,000 account, one erroneous purchase and a 20% adverse price move. Position caps materially reduce maximum modeled damage.
For retail accounts, useful guardrails can include per-order dollar caps, concentration limits, prohibited instruments, leverage limits, minimum cash buffers, confirmation requirements for options or margin, cooldown periods after large losses, and hard restrictions against transferring cash or changing beneficiaries. A model can still be wrong, but it should not be able to transform a bad sentence into an existential account event.
The Bigger Risk Is Not Just Hallucination
Hallucination gets most of the attention because it is easy to understand: the model invents a fact, misreads an earnings release or relies on stale data. But an investment agent creates a broader set of failure modes. It can optimize the wrong objective, overreact to noisy news, repeatedly trade the same theme, misinterpret the investor’s risk tolerance or select investments that are technically consistent with a prompt but unsuitable for the customer’s real financial situation.
There is also a conflict problem. An AI embedded inside a broker or platform may be able to choose between products that create different economics for the firm. The SEC proposed a specific predictive-data-analytics conflicts rule in 2023, then formally withdrew the proposal in June 2025. That withdrawal did not erase existing fiduciary, best-interest, disclosure or anti-fraud obligations. It simply means there is no special standalone federal rule that magically resolves every AI conflict.
The SEC has already shown that even the marketing around AI can be an enforcement issue. In 2024 it charged Delphia and Global Predictions over false or misleading statements about their purported use of AI; the firms agreed to $400,000 in combined civil penalties. For investors, the implication is important: “AI-powered” is not evidence of a superior model, and it is not a substitute for checking who is registered, what the strategy actually does and what legal entity owes the customer duties.
Why the Most Credible Model Is Hybrid, Not Fully Autonomous
The strongest case for AI in investing today is not that a chatbot should replace portfolio construction, risk management, compliance and execution. It is that LLMs can compress information and add an additional signal inside a process whose hard constraints are designed elsewhere. This is consistent with the research: several positive papers use LLMs to classify text, rank securities or narrow an investable universe, then combine that output with conventional portfolio construction or risk controls.
That matters because the jobs are different. Language models are well suited to unstructured information: earnings-call tone, regulatory filings, corporate announcements, news and explanations. Deterministic systems are better suited to exact calculations, exposure limits, order validation, tax rules, portfolio accounting and compliance checks. A regulated adviser or broker can then supervise the combined system and remain accountable to the customer.
Robinhood’s own 2026 technical description of Cortex illustrates the direction of travel. The assistant is grounded in account information and live tools, while input and output guardrails sit around the model. Yet Robinhood’s legal disclosure still says the assistant is not an investment adviser and that customers remain responsible for investment decisions. The architecture is becoming agentic faster than the legal relationship is becoming discretionary.
What Retail Investors Should Check Before Letting AI Near an Account
1. Is this advice or merely information? A personalized interface can feel advisory even when the legal disclosure says otherwise. Check whether the provider is a registered investment adviser, broker-dealer or neither.
2. Who actually has trading discretion? There is a major difference between an AI preparing an order for confirmation and an adviser with authority to rebalance without asking each time.
3. What is the model allowed to buy? A system limited to diversified ETFs presents a different risk profile from one that can trade single stocks, options, leveraged ETFs or crypto.
4. What hard limits sit outside the model? Look for concentration caps, leverage controls, order-size limits, prohibited assets and confirmation thresholds.
5. What happens when data are wrong or stale? The system should identify data timestamps, source provenance and error states rather than silently trading through missing information.
6. How often can it trade? High turnover raises the alpha hurdle even when headline commissions are zero.
7. Who benefits economically from the recommendation? Understand advisory fees, spreads, payment for order flow where relevant, cash-sweep economics, proprietary products and other potential conflicts.
8. Can you override or stop it immediately? Retail automation should have a clear kill switch and a simple path back to manual control.
What Would Prove the Skepticism Wrong?
A stronger case for fully autonomous retail AI investing would require evidence beyond impressive demos and selective backtests. The most convincing proof would be multi-year, live, net-of-cost performance across different market regimes; stable risk-adjusted returns after controlling for conventional factor exposures; transparent turnover and tax impact; documented failure rates; and evidence that the model respects investor-specific constraints during stress rather than only in normal markets.
It would also require governance evidence. A system that can explain why it bought a security but cannot reliably enforce a 5% position limit is not ready for discretion. Conversely, a model that produces only modest incremental alpha but operates inside strong risk, compliance and execution controls may be commercially far more valuable than a spectacular backtest with weak operational safeguards.
Bottom Line
Yes, AI can invest your money—but that statement is less futuristic than it sounds. Algorithms already manage portfolios through regulated robo-advisers, and newer AI assistants can read account data and in some cases initiate transactions. The meaningful dividing line is not whether software touches the order. It is whether the customer or a regulated adviser retains responsibility for the investment decision and what controls sit between a probabilistic model and the brokerage account.
The evidence does not yet justify treating a general-purpose LLM as a superior autonomous portfolio manager. Research shows genuine promise in news interpretation, stock screening and portfolio support, but also inconsistent benchmark beating, factor-driven returns and higher concentration or behavioral risk in some experiments. For retail investors, the safest near-term model is therefore likely to be AI-assisted investing: let the model search, summarize, compare and perhaps prepare actions, while deterministic risk controls and an accountable human or regulated institution retain authority over the money.
The most important question is not “How intelligent is the AI?” It is “What is the AI allowed to do when it is wrong?”
Methodology
This article distinguishes general-purpose generative AI, broker-embedded AI assistants and registered automated investment advisers based on current product disclosures and U.S. securities-regulatory materials available as of October 8, 2026. Research findings are summarized from peer-reviewed or working-paper studies; they are not directly comparable because datasets, models, markets and evaluation windows differ.
The turnover model is an illustrative calculation, not a forecast. Annual trading drag is calculated as annual portfolio turnover multiplied by an assumed all-in friction of 10, 25 or 50 basis points per 1x turnover. The $10,000 permission-cap scenario assumes a single erroneous purchase followed by a 20% adverse price move. It excludes recovery, diversification, taxes, interest, fees and subsequent trading.
Sources
1. SEC — Robo-Advisers: Guidance Update (2017)
2. SEC — Form ADV and IARD FAQs (automated advice as a means of providing advisory services)
3. SEC — Commission Interpretation Regarding Standard of Conduct for Investment Advisers
4. SEC — Regulation Best Interest
5. SEC — Withdrawal of Predictive Data Analytics Conflicts Proposal (June 12, 2025)
6. FINRA — 2026 Annual Regulatory Oversight Report: GenAI
7. SEC / FINRA / NASAA — Artificial Intelligence and Investment Fraud Investor Alert
8. SEC — AI Washing Enforcement: Delphia and Global Predictions
9. OpenAI — Finances in ChatGPT
10. OpenAI — Financial Services Terms (updated September 16, 2026)
11. Robinhood — Cortex Assistant Disclosure
12. Robinhood — Cortex Assistant Agreement
13. Robinhood — Cortex: The Loop Stays Dumb So the Model Can Be Smart (Aug. 25, 2026)
14. Perlin et al. — Can AI Beat a Naive Portfolio? Finance Research Letters (2025)
16. Schneider & Yilmaz — Stock Portfolio Selection Based on Risk Appetite: Evidence from ChatGPT (2025)
17. Anic et al. — ChatGPT in Systematic Investing: Enhancing Risk-Adjusted Returns with LLMs (2025)
19. Biased Echoes — LLMs Reinforce Investment Biases and Increase Portfolio Risks (2025)
Johan Shamshad is a financial markets writer at Dave Finances covering cryptocurrencies, trading platforms, brokers, fintech, financial regulation, and developments across global markets. He previously worked at Gulf News, adding newsroom experience to his coverage of fast-moving financial and digital-asset markets.
His work focuses on identifying market-moving events, company developments, regulatory changes, product launches, and shifts in trading and financial infrastructure.
Johan contributes news and analysis designed to help readers understand not only what happened, but why a development matters and how it may affect the wider financial landscape.