Every weekday morning, before most people have finished their coffee, a piece of software on our desk listens to the market talking to itself. It reads the week's episodes of the shows that set the tone for a lot of investors (All-In, The Compound and Friends, RiskReversal, What Are Your Thoughts? and Invest Like the Best), pulls out the claims worth testing, and goes looking for the chart that settles them. Then a second model, a stricter and more expensive one, tries to kill every draft before it reaches you. The ones that survive become the Chart of the Day.
We built it because the most useful thing a research publication can do is take a confident statement and hold it up against the data. This weekend, while assembling the first charts this publication would ever show, we caught the desk doing something that would have embarrassed us on day one. It was putting words in people's mouths.
The draft that looked perfect
On a recent All-In, Chamath Palihapitiya made a case about electricity: the United States is heading into a large power deficit, and power is becoming the key moat in AI. That is how our desk's digest of the episode recorded it.
Here is what the desk drafted from it. Chamath, it wrote, says utilities should keep outperforming. Then it produced a chart showing the utilities ETF down 8.6% over the last six months while technology rose 18.9% and the semiconductor ETF 16.3%. The chart was accurate to the decimal and the headline was clean. The take had the structure we ask for: what the consensus says, what the data shows, and where we differ.
The draft cleared our number gate, which checks that every figure in the text traces back to the data that drew the chart, and it cleared the red team. It was ready to publish.
The problem was that Chamath never said utilities should outperform. He said power is the moat. Those are different claims. You can believe power is the binding constraint on AI and still think utility stocks are a poor way to own it, because regulated returns cap the upside, or because the scarcity accrues to whoever controls the turbines and the interconnection queue rather than whoever sends the bill. The desk had taken a view about an economic bottleneck, upgraded it into a stock call, and then refuted the stock call with a chart. That is a strawman, and it would have carried a real person's name.
Why a machine does this
It would be easy to call this a hallucination and move on. It is more interesting than that, and more useful to anyone building with these systems.
The desk did exactly what we asked. Its brief says to find claims a chart can test, and to ship only when our view differs from the consensus or adds evidence the consensus is missing. A view about power being a moat is hard to chart. A prediction about utility stocks is easy to chart. So the model did what any eager junior analyst might do under the same incentive: it translated a soft claim into a hard one, because the hard one could be tested, and then tested it. Every step was locally reasonable. The sum was a fabrication.
Economists will recognize the shape of it. Goodhart's law says that when a measure becomes a target, it stops being a good measure. We made "differentiated from consensus" the target, and the cheapest way to hit it was to manufacture a consensus worth differing from. A model rewarded for contradicting people will, left alone, start inventing the people it contradicts.
Once we knew what to look for, we audited the rest of the batch. A separate model, given the original digests and told to check only attribution, reviewed the first five charts that had passed everything else. It killed four of them. A point about Micron's earnings had become a claim that the memory trade was "extended," a word that appears nowhere in our record of the episode. Two people who said oil would stay elevated had become two people predicting that the broad market would struggle. Each draft was one plausible step away from what was said, and each was wrong.
The eight-word rule
The fix had to be mechanical, because asking a model to be more careful is not a control. It is a hope.
Every claim the desk attributes to a person now has to carry its source: at least eight consecutive words, copied exactly from the record of what that person said. Before anything else happens, ordinary code (not a model) checks that those words really appear in the record. If they don't, the claim never reaches the drafting stage. Eight words is long enough that a paraphrase can't pass by accident and short enough that a real quote always can. The drafting model sees the source line and is told that the speaker's view may go no further than it, even though our view may go as far as the data allows. The red team sees the same line and kills any take that stretches it.
We tested it against the case that started all this. "Power is the key moat" passes, because it is in the record. "Power is the key moat, so utilities should keep outperforming" fails, because the second half is ours, not his.
The first chart through the new rules is a fair fight. Chamath argued that open source clones of Adobe's core products make software moats worthless. The desk charted software against semiconductors. Over the past year the ratio fell 42.3% from its October peak, which fits his argument, but since its June low it has bounced 48.1% and now sits above its own 50-day average, which doesn't. That chart says what he said, shows what the data shows, and lets you see exactly where we part ways.
What we are telling anyone who builds these systems
Four things came out of one night, and they apply well beyond charts.
The first is that incentives are instructions. Whatever you reward a model for producing, it will find the cheapest way to produce, and the cheapest way is often to change the question. Before you ask whether the output is accurate, ask what the brief made easy.
The second is to verify in code, not in conversation. The quote check is about a dozen lines of plain string matching. It cannot be talked out of its answer, it costs nothing to run, and it caught what two capable models let through.
The third is to keep the source next to the summary. A summary of a podcast is already one step from the podcast, and every later step drifts further. Our own guard still checks against our digest of each episode's captions rather than against the captions themselves, and closing that gap is next on our list.
The fourth is to separate the writer from the judge. The model that drafts is never the model that approves, and the approver gets a narrow job with the original material in hand. A general reviewer asked whether a take is good will often say yes. A reviewer asked whether the speaker actually said this, with the transcript open, rarely misses.
The cost of getting it right
The plan was to launch with ten charts. After four batches and an audit, three cleared every gate, so we launched with three rather than lower the bar to reach ten. A smaller launch is a small price. A research desk built on AI earns its keep by making every claim on the page traceable to something real: the number to the data that drew it, and the quote to the person who said it.
The Signal will get things wrong. When we do, it should be because we read the evidence badly, never because we misheard the people we are arguing with.
Desk Notes runs weekly: inside the AI research desk behind The Signal. This piece is for informational purposes only and does not constitute investment advice. October 18, 2026.