Photo by Austin Hervias on Unsplash
The short version
- "AI" in investing means three different things — a theme (the companies), an engine (the fund's process), and a tool (the investor's workflow). Marketing thrives on conflating them.
- Most AI-branded equity ETFs, when decomposed into factor loadings, look very similar to crowded large-cap U.S. growth. The fee premium is paid for the label, not for differentiated exposure.
- For a long-horizon investor, the most defensible use of AI in 2026 lives in the workflow — reading filings, summarising prospectuses, screening universes — not in fund products that promise alpha from a model you cannot inspect.
The interesting question in 2026 isn't whether AI will reshape investing — it already is, on the margin. The interesting question is which of the things issuers and platforms now label "AI" actually does something different from what a competent rules-based process has done for thirty years. Because if the answer is "not much," then the 40 to 70 basis points of premium fee that the label commands is a fee on the word, not on the work.
This piece is a framework for telling the two apart. It is written for an investor with a multi-decade horizon, who reads ETF fact sheets, and who would rather pay for distinct exposure than for a story.
Three things "AI" can mean in an investing context
Most of the confusion in this category dissolves once you separate three distinct claims. They are sold under the same word, but they are not the same thing — and the standard of evidence each one deserves is different.
| Layer | What is being sold | How to evaluate it |
|---|---|---|
| AI as theme | Equity exposure to companies that build, supply, or deploy AI (chips, cloud, software, robotics) | Factor and sector decomposition vs. plain large-cap tech; concentration; valuation |
| AI as engine | A fund whose security selection or weighting is driven by a machine-learning model | Live track record vs. backtest; capacity; turnover; tracking-error vs. benchmark; transparency of process |
| AI as tool | Software the investor uses — chatbots, screeners, document summarizers, research assistants | Repeatability; verifiability of outputs; whether it accelerates judgement or replaces it |
These three layers move on different time scales. The theme depends on capex cycles and earnings. The engine depends on whether a model trained on past data continues to work when the environment changes. The tool depends on the investor's discipline in using it. Combining them under one word — "AI investing" — is precisely how a small theoretical edge in one layer gets stretched into a marketing claim that spans all three.
The current backdrop matters for how much the label can cost
With the 10-year Treasury at 4.47% and CPI year-on-year still at 3.9% (FRED, as of mid-May 2026), the opportunity cost of paying an extra 60 basis points for "AI selection" is no longer rounding error. A risk-free rate near 4.5% sets a hard floor on what active management has to clear, net of fees, just to be worth the effort. The VIX at roughly 17 says the market does not currently believe a crisis is imminent, which removes one of the easier excuses for active mandates to charge extra.
In other words: the macro backdrop is unfriendly to vague premium pricing. That is a useful filter when reading any "AI-driven" fund pitch.
Layer 1 — AI as a theme: what you actually own
The most populated category. Thematic ETFs branded around artificial intelligence, robotics, automation, or "machine learning" typically hold a few dozen U.S. and international large-caps in semiconductors, hyperscale cloud, enterprise software, and a smaller satellite of robotics and industrial automation names. Many are constructed by a third-party index provider, with the issuer paying a licensing fee and the investor paying it back through the expense ratio.
The honest test is straightforward: if you regress the fund's daily returns on broad U.S. equity plus a sector-tech factor, how much idiosyncratic exposure is left? In most published decompositions of large AI-themed ETFs, the answer is "not much." The funds load heavily on the same handful of mega-cap chip and cloud names that already dominate broad indices, plus a long tail of smaller names that contribute volatility without changing the dominant exposure.
That does not make thematic AI funds useless. It makes them tilted tech, not a distinct asset class. If an investor already holds a broad market or Nasdaq-100 sleeve, adding an AI-themed fund layers a tilt on a tilt — and the tilt being added is the one the market is already most enthusiastic about. The discussion in "Why Factor Investing Still Works" applies directly: tilts pay over decades only when they are persistent, distinct, and not already crowded into the price.
Layer 2 — AI as an engine: what the model is actually doing
This is where the standard of evidence has to be highest, because it is the layer most prone to marketing.
Quantitative funds that use machine-learning techniques are not new. The professional version — what firms like Renaissance, Two Sigma, and D.E. Shaw built — relies on years of clean, granular data, expensive infrastructure, and a research culture organised around invalidating findings rather than promoting them. The case study in "The Full Framework for Scientific Investing" is useful here as a yardstick: what those firms do is recognisably scientific, with explicit attention to data integrity, look-ahead bias, and capacity limits.
A retail-facing "AI-driven" ETF is a different animal. The fund is typically small, the model is undisclosed, the live track record is short, and the prospectus describes the process at a level too vague to falsify. The investor is being asked to pay an active-management fee for a black box whose only quality signal is the same backtest the issuer used to design it.
Most AI-branded equity funds, decomposed, look like crowded large-cap tech with a thesaurus.
The diagnostic questions for any "AI as engine" fund are mechanical:
- Live versus backtest. How many years of live returns exist, and how do they compare to the backtest the issuer published at launch? A gap of more than two or three percentage points annualised is the norm and a red flag.
- Capacity. What is the AUM, and what is the strategy's stated capacity? Machine-learning signals that work at $50 million in AUM can degrade meaningfully at $500 million.
- Turnover and tax cost. A high-turnover ML model in a taxable account is paying a tax-cost ratio that quietly subtracts from any pre-tax alpha. This is the same point made at length in "It's Not What You Make, It's What You Keep."
- Tracking error vs. an honest benchmark. Not "AI vs. cash," but AI-fund vs. the broad large-cap or sector index that captures its underlying exposure.
A fund that cannot answer these clearly is selling the word, not the engine.
Layer 3 — AI as a tool: where the real edge is, for the investor
The most under-discussed layer is also the one where a long-horizon individual investor probably gains the most. Modern large language models, used carefully, do a few things competently:
- Summarise long fund prospectuses, 10-Ks, and rebalancing index methodologies in minutes rather than hours.
- Translate index construction rules into plain English, which makes it easier to spot the small print that drives behaviour (caps, screens, reconstitution dates).
- Generate first-pass screens across a universe of ETFs — by expense ratio, by issuer, by stated factor — that the investor then verifies against primary sources.
- Surface counter-arguments. Asking a model "what is wrong with this thesis" is a cheap discipline against confirmation bias.
None of this replaces judgement. All of it accelerates the parts of research that are otherwise tedious enough to be skipped. The shift is subtle: the tool changes the cost structure of doing the work, so more of the work actually gets done. Over a multi-decade horizon, that compounds.
The discipline that makes it work is the same one a research-trained editor would apply to any new source: every number the model produces has to be verified against a primary document before it influences a portfolio decision. AI tools hallucinate; primary sources do not.
What the evidence — and the silence — tells you
If retail-facing "AI-driven" funds were producing systematic, persistent excess return net of fees, two things would be visible by 2026: their AUM would be much larger, and institutional asset allocators would be loud about it. Neither is true. The largest AI-themed ETFs are growing because investors want the theme, not because the funds have demonstrated durable alpha against an honest benchmark. The "AI as engine" segment remains small in AUM terms relative to the broad ETF market.
This is the single non-obvious point in the entire category: the absence of a story is itself a story. The institutional money that pays Renaissance-style fees and demands Renaissance-style audit trails has not moved into retail AI ETFs, because it cannot get the disclosure and capacity it needs. The same caution should apply at the individual level. The wider context in "The Complete Guide to AI Infrastructure Investing" reinforces this — owning the picks-and-shovels is a different proposition from believing a black-box fund will pick winners for you.
A practical filter before paying extra for "AI"
| Question | If the answer is unclear, the right move is |
|---|---|
| Which of the three layers is this product selling? | Don't buy until you can name it |
| What is the live track record vs. an honest benchmark? | Default to the benchmark |
| What does the fund's factor decomposition actually look like? | Assume "crowded tech tilt" until proven otherwise |
| What does the fee differential cost over 20 years on this position? | Compute it once; the number is usually decisive |
Editor's read
If forced to take a position, the editor's view is this: in 2026, the most reliable place for a long-horizon investor to capture value from AI is on the workflow side — using the tools to do more thorough research at lower personal time cost — and on the broad-tech-exposure side, which most diversified investors already own through cap-weighted indices. The middle layer, AI-as-engine retail funds, is where the marketing gap is widest and the evidence thinnest. That does not mean the category never produces a winner; it means the prior should be skeptical, and the burden of proof belongs to the product, not the buyer.
FAQ
Does an AI-driven ETF actually beat a passive index?
On the available evidence, retail AI-driven ETFs as a category have not demonstrated persistent excess return net of fees against an honest factor-and-sector benchmark. Individual products may outperform in a given year; the question is whether the outperformance is repeatable and worth the fee. To date, the burden of proof has not been met at the category level.
Are "AI" ETFs different from generic tech ETFs?
Usually less different than the marketing implies. Factor and sector decompositions of the largest AI-themed funds tend to show heavy overlap with mega-cap U.S. tech and semiconductors. If the investor already owns a Nasdaq-100 or broad-tech sleeve, adding an AI-themed fund mostly concentrates an existing tilt rather than adding a distinct exposure.
Can I use a chatbot to make better investment decisions?
Yes, with discipline. Large language models are useful for summarising long documents, drafting counter-arguments, and accelerating first-pass research. Every quantitative claim they produce must be verified against primary sources (issuer fact sheets, SEC filings, FRED) before it informs a portfolio decision. The model is a research assistant, not an analyst.
What's the difference between machine learning and traditional quantitative finance?
Traditional quant largely uses linear, explicit factor models — value, momentum, quality, low-volatility — whose construction is auditable. Machine-learning approaches can capture non-linear relationships and interactions between many variables, but they are harder to interpret and more prone to overfitting on historical regimes. Both depend on the assumption that past relationships hold in the future, and that assumption breaks during structural shifts.
Should I just avoid every fund with "AI" in the name?
No, but the label is not a reason to buy. Treat it as neutral: evaluate the fund's actual exposure, costs, capacity, and process the same way you would evaluate any other fund. If those check out on their own, the AI branding is incidental. If they don't, the branding does not rescue them.
Key takeaways
- "AI investing" collapses three distinct claims — theme, engine, tool — into one word. Separating them is the first defense against overpaying.
- AI-themed ETFs are mostly tilted tech. They are not a distinct asset class, and they usually concentrate exposures a diversified investor already has.
- AI-as-engine retail funds carry the heaviest burden of proof and have produced the least visible evidence. The institutional silence on them is informative.
- The most defensible use of AI for a long-horizon investor is as a workflow tool — used with primary-source verification — not as a fund product.
- At 4.47% on the 10-year and 3.9% on CPI, the opportunity cost of premium fees is real. Pay for distinct, persistent exposure or pay for low cost. The middle is where money quietly leaks.
What this framework can and can't tell you
This piece treats "AI in investing" as a category and looks at how the label is used in 2026. It does not evaluate any specific fund's track record in depth, and it does not rule out the possibility that a particular AI-driven product will, over the next decade, demonstrate durable net-of-fee alpha. The framework's job is to set the prior — skeptical, evidence-led, fee-aware — and to give the investor a small set of questions strong enough to filter most of the category without further work. Any specific fund decision still requires reading the prospectus and decomposing the actual exposures.
Methodology
Macro figures from FRED, asof 2026-05-14 (10-year Treasury, VIX) and 2026-04-01 (Fed funds, CPI YoY). Categorical observations about AI-themed and AI-driven ETFs draw on issuer fact sheets and standard sector/factor decomposition techniques described in the academic literature on factor investing. No specific fund AUM or return figures are cited where the editor could not source them directly from issuer or regulatory documents at time of writing.
By the Mulden editor. The editor does not hold any single-theme AI ETF at the time of writing; broad U.S. and global equity exposure held in the core is incidentally exposed to AI-related companies through cap-weighting. This article is for educational purposes and does not constitute personalized financial advice. See Disclaimer.