A policy analyst using Anthropic’s coding assistant Claude Code has cast doubt on one of the most widely shared housing statistics of 2026. The Federal Reserve’s finding that 49% of U.S. adults under 30 lived with a parent in 2025—a 12‑point rise since 2019—may be largely the product of an unbalanced survey sample, not a genuine social shift, according to a video posted by lawyer Matt Bruenig.

The challenge matters well beyond a single data point. It demonstrates how AI‑powered tooling can quickly surface problems in high‑profile statistics that influence personal finance decisions, news coverage, and policy debates. For Windows users who encounter splashy survey numbers daily, it’s also a reminder that even official figures deserve scrutiny.

The Disputed Statistic and What Claude Code Found

In its 2025 Survey of Household Economics and Decisionmaking (SHED), the Federal Reserve reported that 49% of adults aged 18 to 29 lived with a parent in 2025, up from 37% in 2019. Publications including The Wall Street Journal used the 12‑point jump to frame a story about young adults embracing multigenerational living as a pragmatic financial strategy.

Bruenig, a socialist lawyer and policy analyst who regularly works with microdata, fed the raw SHED files into Claude Code—Anthropic’s terminal‑based coding agent—and asked it to examine the age distribution within the 18‑to‑29 bracket. The assistant quickly highlighted a compositional issue: not all survey years contain the same proportion of the youngest respondents, those closest to 18.

An 18‑year‑old is much more likely to live with parents than a 29‑year‑old. If one survey year happens to include relatively more 18‑year‑olds and another more 29‑year‑olds, an unweighted top‑line comparison can make the entire group look as though a dramatic change occurred when, in reality, the underlying age‑specific rates barely moved. According to Bruenig’s analysis, the real rise—after accounting for the age mix—may be closer to 3 percentage points rather than 12.

Crucially, this is not an AI “discovery” that happened autonomously. Bruenig used Claude Code as a research assistant: loading files, inspecting variable dictionaries, and running transformation checks. He noticed the original headline looked suspicious and then probed the data with the tool until an anomaly emerged. The AI accelerated the grunt work of data handling and hypothesis testing, but the skepticism and domain knowledge came from the human.

What This Means for You

For everyday Windows users, the episode delivers a practical lesson: when you see a startling statistic about housing, employment, or the economy, the raw survey data often contains quirks that can change the story. You don’t need a PhD to question those numbers, and AI tools are starting to lower the barrier for anyone willing to download a public‑use data file and spend an afternoon inspecting it. That said, you do need to understand what a survey weight is and why age cohorts matter—subjects Bruenig’s video walks through in detail.

Power users and IT professionals who already work with local data sets on Windows stand to gain the most from the workflow Bruenig demonstrates. Claude Code, GitHub Copilot Chat, and VS Code‑based agents can connect directly to folders full of CSV, Parquet, or Stata files. Instead of manually writing Pandas or R scripts from scratch, you describe a question in plain language and let the assistant generate, run, and refine the code. That pattern is instantly transferable to your own projects—whether you’re analyzing server logs, customer feedback, or internal ops data.

But the speed of these tools also creates a danger. A clean chart and a confident explanation do not guarantee correct analysis. In Bruenig’s case, the Fed’s SHED data are weighted to represent the adult population; any serious challenge should replicate the official weighting scheme before proposing an alternative estimate. Bruenig’s video is best understood as a prompt for further investigation, not a settled refutation. For IT teams, that means AI‑assisted analysis must still be paired with reproducible scripts, peer review, and clear documentation of methodology.

How We Got Here

The roots of the 49% figure reach back to the Federal Reserve’s annual SHED, a survey designed to capture the financial well‑being of U.S. households. The 2025 edition reported that 49% of adults under 30 lived with a parent, and the accompanying materials did not proactively caution readers about possible age‑composition effects in the 18‑29 band.

What turned the stat into a headline was The Wall Street Journal’s feature article, which presented the 12‑point jump as evidence of a cultural pivot: young adults were now choosing parental co‑residence as a savvy money move rather than a last resort. Radio hosts and social media amplified the piece, and many listeners accepted the number at face value.

Into that environment stepped Matt Bruenig. Already known for his contrarian embrace of AI among U.S. leftists, he had previously argued that progressives risk ceding technical advantage if they refuse to use powerful new tools. His 51‑minute YouTube video shows a terminal setup that runs tmux inside the Ghostty emulator, with Claude Code loaded in one pane and survey files open in another. The demonstration is slow, deliberate, and heavy on statistical detail—but it resonated because it tackled a concrete, widely‑circulated claim.

For Windows professionals, the toolchain is different but the concept is identical. VS Code with a Copilot extension, Windows Terminal with WSL access to Python or R, and a local folder of data can replicate every step Bruenig shows. The difference is that on Windows, many users are already comfortable with these environments for daily development tasks, making the leap to AI‑assisted data sleuthing smaller.

What to Do Now

If the Federal Reserve’s statistic gave you pause—or if you’re simply interested in verifying claims you encounter—here are actionable steps:

  • Download the source data. The SHED public‑use files, including codebooks and weighting documentation, are freely available from the Federal Reserve’s website. Start with the most recent release and read the methodology notes.
  • Check age buckets. When a survey reports a rate for “adults under 30,” dig into the internal composition. Are 18‑ to 21‑year‑olds proportionally the same across years? Many public surveys publish microdata with single‑year ages, making this check straightforward.
  • Use AI assistants to speed things up—not to replace your thinking. On Windows, you can point GitHub Copilot Chat or an open‑source agent to a local data folder and ask: “Load the survey CSV, calculate the median age of respondents in the 18‑29 group per year, and plot the distribution.” Let the AI generate the code, but verify every step. Double‑check that it’s using the correct weight variable (often named weight or wgt) and that it’s treating missing values appropriately.
  • Reproduce the official numbers first. Before you claim a statistic is wrong, replicate the Fed’s published result exactly, including all weighting and filtering decisions. Once you have that baseline, you can test alternative groupings and see how sensitive the outcome is.
  • Share your scripts, not just your conclusions. A reproducible notebook—whether Jupyter, R Markdown, or a plain script—allows others to spot flaws you missed. Bruenig’s video shows his terminal in real time, which is a rough form of transparency; a published repository is better.

Outlook

Bruenig’s video has not been independently validated at the time of writing, and the Federal Reserve has not issued a response. Whether the real increase in 18‑ to 29‑year‑olds living with parents is 3 points, 8 points, or 12 points remains an open question that will require a full re‑analysis with proper survey weights and sensitivity checks.

What the episode has already done is demonstrate that AI‑powered coding assistants, when wielded by a domain expert, can bring healthy skepticism to widely‑circulated numbers in hours rather than weeks. For Windows users and IT teams, the message is clear: the tools to verify the headlines are already on your desktop. The remaining requirement is the human instinct to ask, “Does this really make sense?”