Written for Machines: How to Read an Earnings Call Past the Spin

Guides
Moneda Guides: an illustration of the Moneda face

Short answer

  • Machines read company reports first, so companies learned to avoid the words algorithms score as negative. The tone changed. The numbers did not.
  • Three things are hard to polish: whether executives actually answer analysts' questions, what changed in the annual report since last year, and whether profit is backed by cash.
  • Machines read fast, not wisely. If you use AI to read for you, pick a tool that shows where every number came from.

In 2011, two finance professors at the University of Notre Dame published a list of words.

They had noticed that the dictionaries used to measure the mood of company reports were getting finance badly wrong. They treated a word like "liability" as gloomy, when in an annual report a liability is simply money the company owes. Almost three quarters of the words a standard dictionary flagged as negative were actually neutral in a financial context. So they built their own list of the words that really do signal bad news.

Their list became a standard tool for software that reads company reports at scale. Then the predictable happened: the companies being read began to avoid the words on that list wherever they could.

This week, several of the largest US tech companies publish their quarterly results. In the US, listed companies have to report their figures every three months: what they sold, what they earned and what they spent. Most then hold an earnings call, a conference call in which senior executives talk analysts and investors through the numbers. This week's calls will sound confident. Part of that confidence was written with a machine in mind. This guide explains how that came about, and sets out three checks, each grounded in published research, for reading past it.

Who reads company reports first?

Mostly machines, according to the most detailed study to date. US companies publish their results on EDGAR, a free public database run by the Securities and Exchange Commission, the US markets regulator. A team of finance researchers traced who downloaded annual and quarterly reports, known as 10-Ks and 10-Qs, between 2003 and 2016.

Downloads by machines rose from about 361,000 a year to about 165 million. Over the same period their share of all downloads climbed from 39% to 78%. The researchers counted a reader as a machine if it fetched filings from more than 50 companies in a single day, a pace no human reader could sustain. Their study was published in the Review of Financial Studies in 2023.

Why hand the reading to machines? Speed and scale. Software can digest a report seconds after it is published, when a human analyst would need hours. In markets those seconds are worth money, and the same study found that where more machines were reading, the first trade after a filing came measurably sooner. A single program can also follow thousands of companies at once, every day of the year, something no research team could match.

Speed is not the same as judgement, though. Many algorithms judge a report by counting its positive and negative words, very often against the Notre Dame list, and a word count cannot tell a genuine warning from routine legal language. In April 2013, a hacked news account tweeted that two explosions at the White House had injured the US president. Algorithms scanning news feeds sold immediately, and within minutes about $136.5 billion was wiped off the S&P 500. Prices recovered once people established that the tweet was fake.

How did companies change the way they write?

They adjusted to their audience. The same researchers found that firms with more machine readers began avoiding the words on the Notre Dame list once it was published. The shift was sharpest for precisely the words a machine would count against them, rather than words a human reader might simply find gloomy or even ignore altogether.

It went beyond the page. When the researchers ran recordings of earnings calls through software that analyses the human voice, executives at firms with more machine readers sounded more positive and more excited.

The loop has since tightened. A 2026 study in the Journal of Accounting Research detected generative AI in all five types of disclosure it examined, from earnings press releases and the prepared remarks on calls to the risk factors in annual reports and IPO filings. In 2024, up to 4.5% of newly written disclosure text came from AI. Machines now help to write the reports that other machines read first.

Are companies misleading investors?

Not in any legal sense, and the research does not suggest they are. The figures in a report are still bound by accounting rules. What has changed is the language wrapped around them.

It does mean, though, that the part of a report most people meet first, through headlines, summaries and the opening minutes of a call, is the part that has been most carefully tuned. So a better question than "does this sound good?" is "which parts are hardest to polish?" The research points to three.

How to read an earnings call properly: go straight to the questions

An earnings call is a conference call, usually held on the day a company publishes its results. It comes in two halves: a presentation read from a script prepared well in advance, then live questions from analysts, the people paid to study the company.

The second half is where the information is. Research has found the question and answer session more revealing than the presentation, especially for companies that many analysts follow.

Listen for one thing above all: whether the question actually gets answered. Across 6,802 calls by S&P 500 companies, researchers found that dodged questions tended to come before weaker earnings the next quarter, and rambling answers before a lower share price the next day.

Evasion is rarely a flat refusal. More often, executives answer a slightly different question. A 2026 study used AI to measure how far each answer strayed from what was asked. Answers that wandered were judged evasive 67% of the time, against 22% for those that stayed on point.

How to do it. Find the recording or transcript in the investor relations section of the company's website and skip straight to the questions. For each one, note whether the answer meets it head on or slides towards easier ground. Phrases such as "we don't break that out", or a sudden turn to a more flattering number, are the ones to write down.

How to read an annual report: compare it with last year's

A US company's annual report, the 10-K, is long and getting longer. The Wall Street Journal found that the average grew from about 30,000 words in 2000 to about 42,000 in 2013, close to the length of The Great Gatsby. You do not need to read all of it. What matters is what changed.

In "Lazy Prices", a 2020 study in the Journal of Finance, three economists compared every annual and quarterly report filed by US companies between 1995 and 2014 with the one before it. Much of each report is carried over from the last. When companies did change their language, the changes tended to come before weaker earnings and profitability, and in some cases even bankruptcy. The most revealing changes sat in three sections: the risk factors, the description of the executive team, and legal proceedings.

The market hardly reacted when the altered reports appeared. Prices moved later, once the news behind the changes became public. The information had been there all along. Few people were comparing.

How to do it. On EDGAR, open this year's 10-K and last year's, and go to Item 1A, Risk Factors. Look for risks that are new, risks that have quietly disappeared, and risks now described in more specific or urgent terms. A new paragraph about a major customer, a lawsuit or a supplier deserves a second look. Companies outside the US publish annual reports with a similar section, often headed "principal risks", and the same comparison works there.

Why cash tells you more than profit

Because profit can be recorded before the money arrives, and profit without cash behind it is more likely to fade.

A company can count a sale as revenue once it has delivered the goods, even if the customer has not yet paid. Accountants call the gap between reported profit and the cash a business actually takes in "accruals". There is nothing improper about them; they are how accounting works. But they involve judgement, and judgement can be generous.

Research has repeatedly found that profit built on accruals is less durable than profit backed by cash. Markets, meanwhile, tend to take the headline figure at face value until later earnings make the difference plain.

How to do it. Find two lines and compare them over three to five years: net income, which is the profit, and net cash from operating activities, which is the cash the business generated from its day-to-day work. Both appear in the annual report, the second in the cash flow statement. If profit climbs while cash stays flat, find out why. A fast-growing company waiting for its customers to pay is one explanation. There are less comfortable ones.

What these checks can and cannot tell you

They are signals, not verdicts. The studies above describe patterns across thousands of companies, and any single company can break the pattern. An evasive answer may be protecting a genuine commercial secret. A new risk factor may reflect a change in regulation that affects an entire industry. Most of the research also looks at US companies, and some of it relies on data that is now several years old.

Used together, though, the three checks move your attention from how a company sounds to what it has actually disclosed. That is a habit worth having in any earnings season.

Can AI read company reports for you?

Increasingly, yes, and there is a neat irony in that. Machines have been reading company reports for traders for well over a decade, and companies have adjusted to them. Running these three checks by hand still takes hours per company: time most people investing on their own have never had, and help most of them have never been able to afford.

Software can now read filings and call transcripts in seconds. If you use it, look for one quality above all: it shows you where every number and every quote came from, so you can check the work and make the call yourself. Our guide to whether you can trust AI for investment research looks at what that means in practice.

We will be reading this week's Big Tech results closely. Follow Moneda on X to see what sits behind the tone.

The tone is written for the machines. The numbers are still written for you.

Frequently asked questions

What is an earnings call?

A conference call in which a listed company's management presents its latest results and then answers questions from analysts. Calls usually take place on the day results are published, and most companies post the recording and transcript on their investor relations website.

What is a 10-K?

The annual report that US-listed companies file with the Securities and Exchange Commission. It covers the business, its financial statements and the risks it faces. Quarterly reports are called 10-Qs. Both are free to read on the SEC's EDGAR database.

Do machines really read most company reports?

For US filings, the most detailed study says yes. Research published in the Review of Financial Studies found that machines accounted for 78% of downloads of annual and quarterly reports from EDGAR in 2016, up from 39% in 2003.

If I only have ten minutes, which check should I do?

Read the analysts' questions on the latest earnings call and note which ones were not answered directly. It is the quickest of the three, and the research on evasive answers suggests it is one of the more revealing.

Sources

This guide is for information and education only and is not investment advice. The value of investments can fall as well as rise, and you may get back less than you put in.

Share this post