Who carries the fiduciary risk when a hedge fund's analyst is a model?

The supplied Perplexity Discover reference describes a hedge fund manager running his firm with AI. The underlying account, manager's identity, publication date and original statements are not verified here. This essay therefore tests that operating claim against the governing record, rather than attributing unverified practices or results to a particular firm.

The legal answer is less novel than the staffing arrangement. An investment adviser can delegate analytical work. The adviser's fiduciary obligation does not move with the investment memo.

Who is actually the fiduciary?

The SEC's investment adviser interpretation, adopted on June 5, 2019, describes a fiduciary duty comprising care and loyalty. The adviser must act in the client's best interest. The relationship's scope matters, but an adviser cannot simply disclose away the entire obligation.

For a hedge fund adviser, the advisory client is generally the fund. That distinction matters. A fiduciary duty to the fund does not automatically establish an identical advisory relationship with every limited partner.

Investors nevertheless have separate protection. Rule 206(4)-8 prohibits advisers to pooled investment vehicles from making materially false or misleading statements to investors and prospective investors. It also prohibits fraudulent conduct toward them.

Calling software the analyst changes neither relationship. The relevant actors remain the adviser making investment decisions, the fund receiving advice and the investors receiving representations about their money.

What must the manager understand?

The SEC's 2019 interpretation says the duty of care includes a reasonable investigation into an investment. Advice must not rest on materially inaccurate or incomplete information.

That obligation attaches to the recommendation, not to the job title of whoever assembled its supporting research. A fluent memo is not evidence that its cited filing exists, that its calculation is correct or that its conclusion fits the fund's mandate.

NIST's July 2024 Generative Artificial Intelligence Profile identifies confabulation as a risk. It describes confidently presented erroneous or false content. This is a documented system risk, not proof that every model-generated recommendation is defective.

The manager therefore needs a defensible basis for relying on the analysis. Checking a model's output against filings, market data and the investment mandate can supply that basis. Asking the same model whether its answer is correct does not independently establish the underlying facts.

Where does the money go?

Suppose the adviser replaces paid analytical work with model subscriptions. If the advisory agreement leaves the management fee unchanged, lower operating expenses can increase the adviser's margin. That is a contractual consequence, not an established fact about the manager in the supplied reference.

The next question is who pays for the software. An expense absorbed by the adviser reduces that margin. An expense properly charged to the fund reduces fund assets instead. The allocation cannot be inferred from the claim that AI runs the firm.

Form ADV Part 2A requires covered advisers to describe fees and compensation. It also requires disclosure of investment methods, strategies and material risks. The advisory agreement and offering documents establish the particular commercial arrangement.

There is no verified fee schedule or expense allocation for this firm in the record used here. A claim about cheaper research is therefore not evidence that investors receive cheaper investment management.

Who absorbs a bad trade?

A fiduciary duty is not a performance guarantee. The SEC's interpretation requires care and loyalty, not profitable results from every decision.

An ordinary trading loss initially reduces the fund's assets and the value of investors' interests. Recovering that loss from the manager requires an applicable contractual or legal basis. The existence of a model-generated mistake does not, by itself, settle that question.

The reverse is equally important. Describing the mistake as a software failure does not remove the adviser's obligations under Section 206 of the Investment Advisers Act. Responsibility turns on the conduct and applicable duty, not the branding of the research tool.

OpenAI's business terms provide one illustration of the vendor side. They place responsibility on the customer for evaluating output accuracy and appropriateness, including human review where appropriate. Those terms do not establish this fund's vendor or contract. They show why buying analytical capacity is not the same transaction as buying investment accountability.

What would an examination find?

For SEC-registered advisers, Rule 206(4)-7 requires written compliance policies reasonably designed to prevent Advisers Act violations. It also requires an annual review and a designated chief compliance officer.

A manager saying AI runs the firm still needs that compliance structure. The rule does not create an exemption for a small human staff or a large automated workload.

Rule 204-2 supplies a separate records obligation. Among other records, it covers specified written communications concerning recommendations and advice. It does not expressly command the retention of every model prompt. Whether a particular exchange belongs in the required records depends on its content and the applicable provision.

The practical evidentiary question is concrete: can the adviser produce the recommendation, its supporting material and the relevant communications? An assertion that the model handled the analysis is not a substitute for records the rule requires.

Does a smaller firm escape?

Registration status must be established rather than assumed. Rule 203(m)-1 provides a federal registration exemption for qualifying advisers solely to private funds with less than $150 million in private fund assets under management in the United States.

That exemption matters because not every requirement imposed on SEC-registered advisers applies identically to an exempt adviser. It does not make fraud lawful. Rule 206(4)-8's pooled-fund protections are not confined to registered advisers.

Nor does an audit solve the research problem. Under the custody rule, a qualifying annual audit can satisfy specified requirements for a pooled vehicle. A financial-statement audit is not regulatory approval of each investment memo or of the system producing it.

Registration, custody safeguards and investment diligence answer different questions. None supplies a blanket certificate that an AI-run firm exercises adequate judgment.

What can the manager advertise?

The SEC has already enforced the distinction between using AI and making supportable claims about it.

On March 18, 2024, the agency announced settled charges against Delphia and Global Predictions over false or misleading statements about their use of AI. Delphia agreed to a $225,000 civil penalty. Global Predictions agreed to a $175,000 penalty. Neither admitted nor denied the SEC's findings.

Those cases did not establish that AI investment management is inherently unlawful. They established consequences for unsupported descriptions of what advisers were doing.

The investment adviser marketing rule separately prohibits covered advertisements from making material factual claims without a reasonable basis for substantiation. A manager's account of automated research, oversight or capabilities belongs against that standard. The firm's operating records must support the sales language.

Who carries the risk?

The adviser carries the fiduciary obligation, while the fund and its investors carry investment losses unless a contract or legal remedy shifts them. A software provider's separate obligations depend on its contract and conduct. Running research through a model can change staffing costs and increase the manager's margin. It does not make the model responsible to the fund, establish that investors share the savings or excuse the adviser from examining the work.