SEC Marketing Rule and AI Claims: Evidencing ‘Fair and Balanced’ Performance Tech Assertions

The moment an adviser says its strategy is “AI-driven”, it has made a claim the SEC now treats as a performance assertion — one it must be able to substantiate on demand, with documentation the marketing team has almost certainly never seen.

Most writing on “AI washing” frames it as a marketing-copy problem: tone down the adjectives, drop “revolutionary”, add a disclaimer. That misreads what the Commission actually did. When the SEC settled its first two AI-washing cases in March 2024, it did not fine the firms for enthusiastic language. It fined them because the AI they advertised either did not exist in the form claimed or could not be shown to work as described. The fix is not a lighter touch on the adjectives. It is an evidence file that ties every AI claim to what the model does, how it was tested, and where it fails.

What the Marketing Rule actually demands of an AI claim

Rule 206(4)-1 under the Investment Advisers Act — the Marketing Rule, adopted in December 2020 with a compliance date of 4 November 2022 — sets out seven general prohibitions. Two of them do the work here. An adviser may not make a material statement of fact it does not have a reasonable basis for believing it will be able to substantiate on demand by the Commission. And an adviser may not discuss a benefit without fair and balanced treatment of the associated material risks or limitations.

Read those two together against a sentence like “our machine-learning models identify mispriced securities” and the compliance surface becomes obvious. “Substantiate on demand” means that when an examiner asks how the model identifies mispricing, someone has to produce the answer — not a marketing narrative, but the model’s design, its inputs, its validation results. “Fair and balanced treatment of limitations” means the advert cannot present the model’s upside while staying silent on what it cannot do, when it degrades, or how often it is wrong. The claim drags the model’s technical documentation directly into the marketing-compliance file. If that documentation does not exist, the claim is unsubstantiated by definition.

What the first two cases were really about

On 18 March 2024 the SEC settled with two robo-advisers. Delphia (USA) Inc. paid a 225,000-dollar civil penalty; Global Predictions, Inc. paid 175,000 dollars. Both settled without admitting or denying the findings, and both were charged under the Marketing Rule alongside the anti-fraud provisions of Sections 206(2) and 206(4) and the compliance-programme rule, 206(4)-7.

Delphia had told clients and prospects it used their personal data and machine learning to inform its investment process. The Commission found it had not, in fact, incorporated that client data into its algorithms in the way advertised. Global Predictions marketed “expert AI-driven forecasts” and called itself the “first regulated AI financial advisor” — claims the SEC found it could not support. Neither case turned on a subtle question of degree. In both, the gap between the assertion and the evidence was the violation. That is the tell for every firm making AI claims now: the enforcement risk is not the word “AI”, it is the distance between the word and the documentation behind it.

The substantiation file to assemble before you publish

Before an AI capability appears in a pitch deck, a website, a fact sheet or an RFP response, the following should exist and be retrievable. This is the record a supervisor will reconstruct after the fact, so build it before the claim goes out.

• A plain description of what the model does in the investment process. Where it sits, what it consumes, what it outputs, and whether its output drives a decision or merely informs a human who does. “AI-driven” and “AI-assisted” are different claims; the file must say which is true.
• Evidence the inputs are what the marketing says they are. Delphia’s undoing was claiming to use client data it did not use. If the advert says the model learns from client data, the data lineage has to show client data actually reaching the model.
• Validation and testing records. How the model was tested, on what data, against what benchmark, with what results. Backtests carry their own Marketing Rule baggage around hypothetical performance; the testing that supports a capability claim is distinct from performance advertising and needs its own documentation.
• A stated set of limitations. Known failure modes, conditions under which the model degrades, and what it does not do. This is what makes the claim “fair and balanced” rather than one-sided.
• A change and version history. Models are retrained and swapped. The claim made in January may not describe the model running in July. Someone must own the review that keeps the advert and the model in step.

None of this is exotic. It is the same discipline that underpins an assurance-ready data pipeline: a claim is only as good as the lineage and testing you can put behind it when someone asks. The difference is that most firms built that discipline for financial data and never extended it to the model layer.

Where the record has to live

Substantiation is not a one-time exercise. The advert itself is a record the adviser must retain, and the evidence behind it has to survive long enough to answer a question raised in a later examination. That is a recordkeeping problem as much as a compliance one — the advertising copy, the model documentation current at the time it ran, and the testing results all need to be tied together and held immutably, in the same way firms already handle compliant immutable storage for other regulated records. A model-documentation file that has been quietly overwritten with the current version cannot prove what was true when the claim was made.

The organisational fault line

The reason AI washing is a governance problem and not a copy-editing one is that the claim and the evidence live in different departments. Marketing writes “AI-driven”. The data-science team holds the only records that could substantiate it. Compliance owns the Marketing Rule obligation but rarely reads model cards. Under the rule, the adviser needs a reasonable basis to substantiate before the statement is disseminated — which means the sign-off has to reach across those three functions before the claim is published, not after an examiner asks.

The uncomfortable version of the test: pick the boldest AI sentence in your current marketing and ask who, today, could produce the model documentation that proves it. If the honest answer is that nobody could find it inside a day, the claim is already unsubstantiated — the SEC just has not asked yet.