Deceptive language on an earnings call is not a vocabulary list. Peer-reviewed work finds measurable differences in how executives speak when later-restated results are being discussed, and it finds that evasion in the question-and-answer segment is common, classifiable and associated with weaker subsequent outcomes. Linguistic analysis can flag calls for scrutiny. It cannot establish intent, and the sections below separate what the evidence supports from what it does not.
TL;DR
- Evasion is the better-documented phenomenon: managers can escape a hard question without saying anything false, and researchers classify the tactic (refusal, inability, deferral, omission, vagueness, restatement).
- Larcker and Zakolyukina found that models built on CEO and CFO language predicted restatements better than chance, and that top CFO deception scores were followed by negative annualized alpha of 4% to 11%.
- Extreme positive wording, fewer references to shareholder value and more references to general knowledge appear among the markers associated with deceptive calls.
- Language analysis yields probabilistic flags, not proof of intent. Claims must be traced to filings and later delivery before any conclusion is drawn.
Table of Contents
- 1. What deceptive and evasive language means
- 2. Language patterns associated with deception
- 3. The Q&A segment: how executives dodge
- 4. How researchers detect it: models and NLP
- 5. What the market does with this language
- 6. What language analysis cannot establish
- 7. A practical framework for reading a call
- 8. Perspective: why the gap matters more than the word list
- 9. Applying the method with Lacuna Index
- FAQ
- Sources
1. What deceptive and evasive language means
Three concepts are routinely conflated and should be kept apart.
- Positive tone is the share of favorable vocabulary in prepared remarks or answers. It is common, partly cultural and often legitimate.
- Uncertainty or hedging is language that qualifies a claim ("approximately," "we expect," "subject to"). Much of it is a legal convention attached to forward-looking statements. A typical SEC-filed safe harbor statement cautions that forward-looking statements are not guarantees and that actual results may differ materially, which is why hedging alone says little about candor.
- Evasion is a response that does not meet the informational intent of the question. Academic work describes evasive responses as the mechanism that lets managers escape difficult questions without lying.
Deception in the strict sense requires a false or misleading representation made knowingly. The literature mostly measures proxies: language on calls later associated with restatements, or answers that fail a responsiveness test. A call can be evasive and honest, or fluent, specific and wrong.
2. Language patterns associated with deception
2.1 Markers in prepared remarks and answers
The most cited study in this area, Larcker and Zakolyukina (2012) in the Journal of Accounting Research, compared calls labeled deceptive (those tied to later restatements) with the remainder. Deceptive executives used more references to general knowledge, fewer nonextreme positive emotion words and fewer references to shareholder value. Deceptive CEOs also used more extreme positive emotion and fewer anxiety words.
The pattern is counterintuitive for practitioners who equate caution with concealment. In this sample, the signature was less of the ordinary, measured vocabulary and more of the superlative, combined with distance from concrete accountability to owners.
2.2 Vocal and linguistic cues
A 2015 study of more than 1,000 earnings-call utterances found that restatement-related utterances differed on pitch, intensity, tempo, specificity and hedging. Audio features add a channel that transcripts omit, although most commercial and academic pipelines still work from text.
2.3 Temporal framing
A 2025 FinnNLP study of financial Q&A reported that fully evasive answers favored future-oriented or unanchored time framing: 29% of such answers, against 14% of direct answers. A question about last quarter that is answered with next year's ambition is a recognizable displacement.
3. The Q&A segment: how executives dodge
Prepared remarks are scripted and reviewed by counsel. The Q&A segment is where the script ends, and therefore where most evasion research concentrates.
3.1 A taxonomy of non-answers
Larcker, Gow and Zakolyukina, summarized by Stanford Graduate School of Business, sorted evasive executive responses into three main types:
| Type | Form | Example |
|---|---|---|
| Refusal | Declines to provide the information | "We don't disclose those numbers" |
| Inability | Claims not to have the information | "I don't have that information right now" |
| Deferral | Moves the topic elsewhere | "Let's take that discussion offline" |
The 2025 FinnNLP study proposed a complementary set of strategies: omission, vagueness, non-sequitur and restatement of the question. It found that approximately 30% of answers in its financial Q&A data were evasive.
3.2 What provokes evasion
The same Stanford summary reports that analyst questions with a negative tone, high complexity (a four-part question, for instance) and references to past failures are more likely to draw evasive non-answers. The implication is methodological: a non-answer cannot be read in isolation from the question. A compound question that is answered in part is a different signal from a simple question that is ignored.
3.3 Prevalence is high
In research by Chen, Rui and Whinston covering 6,802 earnings calls from 2006 to 2018, managerial responses showed on average high levels of evasiveness (60% of the maximum on their scale) and incoherence (62%). Evasion is the baseline condition of the genre, so the analytical question is relative: which issuer, which topic, and how much worse than its own history and its peers.
4. How researchers detect it: models and NLP
4.1 Classification against a ground truth
Larcker and Zakolyukina built linguistic classification models from CEO and CFO narratives. Their out-of-sample performance exceeded random guessing by 6 to 16 percent when predicting financial restatements. That is meaningful and also modest: the models are screening tools with substantial false positives.
4.2 Responsiveness rather than sentiment
The FinnNLP researchers argue that evasiveness must be measured by whether the answer meets the core informational intent of the question, not merely by sentiment or factuality. Their lightweight baseline targets omission, vagueness and reframing. This is the direction most relevant to practitioners, because it tests the relationship between question and answer, which sentiment dictionaries cannot see.
4.3 Machine reading at scale
Chen, Rui and Whinston proposed that machine learning can quantify evasiveness and incoherence in strategic conversations and use them to predict future company performance. As Whinston notes in the McCombs write-up, machines are systematic and process call content quickly, which supports consistent coverage across a large universe of issuers.
5. What the market does with this language
5.1 Outcomes associated with deceptive language
- Firms in the portfolio with the highest CFO-narrative deception scores produced an annualized alpha of negative 4% to negative 11% in the Larcker and Zakolyukina analysis.
- In the Chen, Rui and Whinston sample, a 40 percentage-point increase in managerial incoherence corresponded to a 0.74% drop in one-day stock returns and to lower next-quarter earnings of 5.6 cents per share per unit of seasonal earnings change volatility.
5.2 Vagueness is a persistent trait
A 2017 NBER working paper by Dzieliński, Wagner and Zeckhauser found that markets perceive vague talkers as less informative, producing slower and smaller price responses. The fixed component of a speaker's vagueness, rather than its time-varying strategic component, most strongly affected market reactions. A manager's baseline style is therefore priced, and a deviation from that baseline is the more informative event for an analyst, even if the market does not always register it.
5.3 Suspicion changes interpretation
In a 2025 Management Science article, Leonardo Barcellos showed that investors who adopt a suspicious state of mind, with suspended judgment and consideration of rival possibilities, are better at distinguishing evasive tactics from legitimate responses. Method, not temperament alone, improves detection: a reader who asks "what would a direct answer have contained?" performs better than one who reads for tone.
6. What language analysis cannot establish
The limitations are structural, and a credible report states them.
- Intent is not observable. Language markers correlate with outcomes such as restatements. They do not show that a speaker knew a statement was false. Evasion also has benign causes: pending litigation, competitive sensitivity, disclosure rules and counsel's advice.
- Context dependence. Hedging is partly a legal convention, and the safe-harbor boilerplate attached to forward-looking statements inflates uncertainty vocabulary for reasons unrelated to candor.
- Adversarial adaptation. Word-frequency and sentiment methods are, according to a recent financial-fraud text-analysis paper, vulnerable to deliberate word choice manipulation, and they overlook context, negation and tone shifts.
- Label noise. Restatements are an imperfect proxy for deception: some are honest errors, and some deceptive calls never produce a restatement.
- Question effects. Hostile or compound questions elicit evasion from honest managers, so analyst style confounds the signal.
- Moderate accuracy. A 6 to 16 percent improvement over chance supports prioritization, not adjudication.
The disciplined conclusion is that linguistic signals identify where to look. Verification requires external evidence: filings, subsequent results and the record of what management said it would do.
7. A practical framework for reading a call
The following sequence converts the research into a repeatable procedure.
- Log the claims. Record each quantifiable or checkable commitment in prepared remarks (targets, timelines, margin paths, capital allocation).
- Log the questions. Note which analyst questions were direct, which were compound and which referenced past misses.
- Classify each answer. Use the taxonomy: direct, refusal, inability, deferral, omission, vagueness, non-sequitur, restatement.
- Check temporal anchoring. Flag answers to backward-looking questions that pivot to unanchored future framing.
- Compare with baseline. Contrast the executive's evasion and vagueness with prior calls and with peers in the same sector, since vagueness is partly a fixed trait.
- Test against delivery. Return to later filings and results to determine whether stated commitments were met.
Published methodology on related steps is available in the Lacuna Index discussions of hedging language detection, Q&A quality scores traced to public records and sentiment versus performance testing in SEC filings.
8. Perspective: why the gap matters more than the word list
Our view is that deception vocabulary, taken alone, is the weakest layer of evidence on an earnings call. The research above supports that position: effect sizes are moderate, the markers are contested, and honest executives can display them. What we regard as more durable is the aspiration-to-execution gap, the measurable distance between what a company said it would do and what its public record later shows it did.
That gap can be audited. A commitment made on a call can be located by quote, matched to a later filing and scored against an explicit standard. Our approach applies the same method to every covered issuer, uses public records only, claims no insider access and publishes governance logs for scoring, so a reader can trace a score to the sentence behind it. We treat linguistic signals such as evasion in Q&A as inputs to scrutiny, never as findings of intent. A pattern of unanswered questions becomes analytically useful when it coincides with commitments that were later not delivered, and it is then a documented observation instead of an inference about character.
For related reasoning, see our forensic audit framework for earnings call credibility and our analysis of why companies overpromise on earnings calls.
— Glen
9. Applying the method with Lacuna Index
Institutional investors, governance professionals, proxy advisors and financial journalists who need the evidence trail rather than a sentiment score can use Lacuna Index to compare narrative claims with delivery. Sector benchmarks are free to read, including Financials, Health Care and Industrials. Company-level forensic reports are available by subscription. The methodology walkthrough documents the scoring process, and the plans page describes Cohort Pulse and Sector Sweep.
For deeper reading on how narrative scores connect to quotes, see earnings call narrative analysis, and for the valuation consequences, narrative engineering in corporate reporting.
FAQ
Can software prove an executive lied on an earnings call?
No. Models can flag language associated with later restatements, and the best published classifiers beat random guessing by 6 to 16 percent. That supports prioritizing calls for review. Establishing a lie requires evidence of what the speaker knew, which transcripts do not contain.
Is a very positive earnings call a warning sign?
Not by itself. In Larcker and Zakolyukina's data, deceptive CEOs used more extreme positive emotion, but positive tone is also normal for firms performing well. Extreme wording is a reason to check claims against filings and later results, not a conclusion.
What is the difference between evasion and deception?
Evasion is failing to answer the question asked, through refusal, inability, deferral, omission, vagueness or restatement. Deception requires a knowingly false or misleading representation. Evasion can be legitimate, for instance when disclosure is restricted, and it can occur without any false statement.
Do analysts' questions affect how evasive answers are?
Yes. Research summarized by Stanford GSB finds that negative, complex and past-failure-referencing questions are more likely to draw evasive answers. Any evasion score should therefore be read alongside the question that prompted it.
Sources
- Larcker and Zakolyukina, Detecting Deceptive Discussions in Conference Calls
- Barcellos, Overcoming Deceptive Evasions in Earnings Calls (Management Science)
- Stanford GSB, What Executives Are Really Saying When They Dodge Questions
- McCombs, To Detect Evasive CEOs, Use AI
- Detecting Evasive Answers in Financial Q&A (FinnNLP 2025)
- Which Spoken Language Markers Identify Deception in Earnings Calls
- Chen, Rui and Whinston, Conversation Analytics
- Dzieliński, Wagner and Zeckhauser, Straight Talkers and Vague Talkers (NBER)
- Financial statement fraud detection using topic-driven sentiment analysis
- Example SEC-filed safe harbor cautionary statement