Is that AI fund's track record real, or am I looking at the survivor?

Open any best-AI-ETFs list this week and the message is the same: the category works, and the track record proves it. The iShares Future AI and Technology fund returned about 91% over the past year. Global X’s AI and Technology fund holds $8.56 billion and is up roughly 51%. Roundhill’s generative-AI fund is up 151% since it launched in May 2023. The 48 funds that now carry an artificial-intelligence label hold around $36.5 billion between them, and they pulled in more than $19 billion over the past twelve months. The belief that follows is reasonable: the average AI fund has beaten the index in real money, so the strategy is doing what it claims. I held a version of that belief too, until I started reading the closure notices.

The kernel of truth

The belief is not built on nothing. The demand is genuine and the flows are not imaginary: quantitative strategies took roughly 70% of the hedge-fund industry’s net inflows last year, and AI-themed funds have been among the fastest asset-gatherers in the exchange-traded world. Some of the headline numbers are real too. A fund that actually returned 91% returned 91%, and nobody is fabricating the survivors.

The kernel becomes a myth at the next step, when we read the average of the survivors as if it described the strategy. That is where the wiring matters, and it is wiring most of us never look at, because the funds that would pull the average down are not on the list we are reading. They have already been closed.

What the data actually says

Start with how fast funds now die. Issuers launched about 370 new exchange-traded funds by early May, against 290 by the same point last year, and roughly 80% of this year’s launches are active strategies rather than plain index trackers. The other side of that funnel is unforgiving.

More than 40 funds were liquidated in the first two months of 2026, against 33 in the same stretch of 2025. And the survivors are getting younger.

We are not talking about funds that ran a decade and faded. We are talking about strategies wound down before their second birthday. On May 29 one issuer, YieldMax, filed to close four of its funds at once, with trading ending June 15 and cash returned around June 18, on the stated grounds that the funds had not reached “the scale or traction needed to best serve shareholders.” Stack the dispersion on top: inside the same AI theme this year, the iShares fund is up around 91% over a trailing year while JPMorgan’s US Tech Leaders fund sits near negative 8% for the year, against a Nasdaq-100 that is itself down about 4%. The category does not have a return. It has a range, and the closures keep trimming the bottom of it.

The mechanism behind the divergence

Here is the mechanism, and it does not require a single fund to be dishonest. A fund family launches a wide set of thematic strategies, sometimes a dozen variations on one idea. Some work, some do not, and the house rules now close the laggards inside twelve to eighteen months. The survivors keep reporting their records. The funds that died stop reporting anything. Average only the survivors and the number drifts upward on its own, mechanically, with no skill involved. Statisticians call it survivorship bias, and the modern ETF launch-and-cull cycle has turned it into an industrial process.

Incubation is the quieter cousin. A manager can run several strategies privately, in-house, and only public-launch the one that performed, sometimes attaching a backtested record for the years before the fund existed. The SEC’s marketing rule treats that backfilled, hypothetical performance as its own category precisely because a backtest is chosen with hindsight, not lived through. The simplest tell is a track record that is longer than the fund’s actual age.

What we do with this

So the test is not “what did this fund return.” It is “what did this fund’s whole cohort return, including the dead ones, and how old is the record I am being shown.” Two quick checks size the edge fairly.

A high closure count is not a scandal. It is the upstream of the average, and it tells us the surviving record was selected, not earned.

First, ask whether the advertised track record predates the fund’s inception. If it does, part of it is a simulation, and simulations are written by people who already know how the period ended. Second, look at how many sibling funds the same issuer has closed in the past two years.

The funds that beat the index are real. What is missing from the list is everything that was quietly removed before the list was printed. I keep coming back to one question: if a record survived a cull I never saw, whose skill am I actually measuring, the manager’s, or the editor’s?