Skip to main content

Evidence Standards for Policy Claims

The short answer

Grade the claim, not the confidence of the sentence carrying it

A claim's evidence grade (measured, modeled projection, interested estimate, unsupported assertion) depends on how the underlying figure was produced, never on how declaratively it is stated.

What you will be able to do

  • Grade any policy claim you encounter on a four-level evidence ladder (measured, modeled projection, interested estimate, unsupported assertion), based on how the claim was actually produced, not on how confidently it is stated.
  • Trace a policy statistic from a headline or secondary retelling back to its primary source, distinguishing the original researcher's own words from every layer of paraphrase added on the way to you.
  • Check a claim's date and its population before restating it: when the underlying measurement was taken, whether it has since been revised or superseded, and exactly who or what it describes, not the broader group a headline implies.
  • Distinguish the verbs a primary source actually uses (measured, found, estimates, models, could, projects, may) from the stronger verbs a secondary retelling substitutes (will, is, causes), because the verb is where a claim's evidence grade quietly changes.
  • Restate a graded claim at the exact strength its evidence supports, producing a sentence a skeptical reader could not fairly accuse of overclaiming.
  • Assemble an evidence ledger: a working document that records, for every claim in a policy argument, its source, its grade, and its honestly restated form, ready to defend line by line.
  • Apply the grading method evenly, to claims that support a preferred conclusion and to claims that complicate it alike, and set a named review trigger so a graded claim's currency is checked rather than assumed.

The ideas, one by one

The primary source is the actual document that produced the finding, not the retelling closest to it

Follow every citation back to its origin; a secondary source citing a primary one is not itself the primary source.

Date and population are checks, not decoration

A checked date tells you whether a figure is current; a checked population tells you who or what was actually measured, almost always narrower than a claim's plain-language framing implies.

The verb carries the grade

"Measured" and "found" signal Grade M; "estimates," "models," "could," and "projects" signal Grade P; a stated interest signals Grade I regardless of verb; no traceable methodology signals Grade U. Watch specifically for a hedge quietly deleted between the primary source and the version in front of you.

An interested estimate is not automatically wrong, and it is not automatically trustworthy

Grade it Grade I, weigh it against whatever independent corroboration exists, and never let a stake-holder's own confident framing substitute for your own graded restatement.

Restate every claim at the strength its evidence supports

The correct restatement is usually longer and more qualified than the compressed version that first reached you; it is also the only version you can defend, unprompted, to a reader who independently checked the same source.

The evidence ledger makes grading auditable and reusable

Five columns, one row per claim: the original wording, the primary source, the checked date and population, the grade with justification, and the restated claim, ready to defend line by line.

An honest restatement can be a stronger argument, not just a safer one

As the immersive scenario shows, the correctly graded version of a claim sometimes makes a more precise, more persuasive case than the overstated original, because it survives scrutiny the overstated version invites and fails.

A claim's grade is not permanent

A Grade M count gets revised, a Grade P projection gets superseded, a study can face a later credible critique. Every ledger row needs a named review trigger, the same discipline this module has taught for a legal citation and a crosswalk row.

This skill is the prerequisite every other evidence discipline in this program assumes

The conformity file, the crosswalk, the evidence annex all assume the underlying facts feeding them are already solid; grading a claim before it enters any of those documents is what makes that assumption true.

Apply the method evenly, and build a standing ledger once claims start repeating across documents

Grading a favorable claim as rigorously as an unfavorable one is what keeps the discipline a credibility asset rather than a selectively applied performance of rigor, and reusing a graded claim's trace across documents concentrates fresh effort on genuinely new claims.

Listen to it

The same lesson, talked through by two hosts as episode 51 of the podcast.

Real cases

These examples show the grading method applied to figures a policy professional is likely to encounter directly. Each is cited to its primary or reputable source.

Example 1: The compression chain itself (IMF, January 2024). The gap between IMF Staff Discussion Note SDN/2024/001's carefully hedged, income-group-broken-out finding and the flattened "AI will destroy 40 percent of jobs" version that circulated within days is not a story about any one outlet behaving badly; CNBC, CNN, and Euronews each accurately reported Georgieva's own framing. It is a story about what happens to a Grade P claim after three or four further rounds of informal retelling with no one checking it back against the primary source. This is the anchor case Section 3D traces in full, and it is worth returning to here as the reminder that the compression happens gradually, not in one dramatic misquote.

Example 2: A long-range projection whose range collapses to its upper bound (PwC, 2017). PwC's "Sizing the Prize" report estimated that AI could contribute up to $15.7 trillion to global GDP by 2030, a 14 percent GDP boost, built on an explicit regional and sectoral model (PwC, "Sizing the Prize," June 2017). The figure is Grade P, a transparent long-range economic model, not a measured near-term outcome, yet it is frequently repeated today, nearly a decade after publication and only a few years from its own 2030 horizon, without noting either its age or that it was always a multi-year cumulative projection rather than an annual or already-realized figure.

Example 3: A range that gets quoted as a single number (McKinsey Global Institute, 2023). McKinsey Global Institute's June 2023 analysis of generative AI's economic potential modeled an annual value of $2.6 trillion to $4.4 trillion, built from 63 specific use cases (McKinsey Global Institute, "The economic potential of generative AI," June 2023). Because a range is harder to headline than a single number, secondary coverage routinely quotes only the top of the range, $4.4 trillion, discarding both the lower bound and the explicit use-case methodology that produced it, a Step 3 and Step 4 failure in one move.

Example 4: A projection whose population is a specific employer survey, not "the world" (World Economic Forum, 2023). The Future of Jobs Report 2023's net figure of 14 million jobs lost by 2027 (69 million created, 83 million eliminated) comes from a survey of 803 companies representing 673 million jobs (World Economic Forum, Future of Jobs Report 2023, May 2023). It is frequently restated as what "the global workforce" or "workers worldwide" can expect, when the actual population is a specific, named set of surveyed employers reporting their own hiring expectations, a Step 3 population check every restatement of this figure should carry.

Example 5: A genuinely measured count, for contrast (Stanford HAI, 2025). Not every governance-relevant figure is a projection. Stanford HAI's AI Index Report 2025 counted 59 AI-related regulations introduced by 42 United States federal agencies in 2024, more than double the 25 regulations from 21 agencies the year before (Stanford HAI, AI Index Report 2025, Chapter 6: Policy and Governance, April 2025). This is Grade M: a direct tally of enacted regulatory activity, not a model of future activity, and it is worth holding up beside Examples 2 through 4 specifically because the contrast in verb and confidence is instructive: a measured count earns a flatter, more direct restatement than any projection in this list does.

Example 6: An interested estimate that still cites its own methodology (industry survey practice). A well-run interested estimate discloses its sample size, its selection method, and its sponsor plainly, which is what separates a usable Grade I citation from an unusable one. A vendor or association survey that states "based on responses from 400 self-selected customers surveyed in Q2 2026" gives a policy analyst enough information to grade it honestly and cite it with an accurate population and sponsor disclosed; a vendor claim with no disclosed sample, date, or methodology at all cannot be graded past Grade U no matter how specific the number sounds, because there is nothing behind the number to check.

Example 7: A finding whose income-group breakdown is the whole point (IMF, January 2024, revisited). Return once more to the anchor case, this time for what its own authors treated as the headline finding, not the global average. The IMF's own paper leads with the fact that the 40 percent global figure masks a sharp divide: roughly 60 percent exposure in advanced economies against roughly 26 percent in low-income countries, a gap the paper's authors present as the more policy-relevant number than the flattened global average, because it is the divide, not the average, that drives their call for differentiated policy responses by country income level. A restatement that keeps "40 percent worldwide" but drops the divide has kept the least policy-relevant part of the finding and discarded the part its own authors considered most important.

Example 8: A widely cited figure whose original methodology later shifted (illustrating the currency discipline). Both the World Economic Forum's Future of Jobs Report and Stanford HAI's AI Index Report are actively maintained, regularly updated series, each new edition built on a refined methodology and fresh survey or measurement data rather than a simple rerun of the prior edition's model, but they do not share a cadence: Stanford HAI's AI Index Report publishes every spring, while the Future of Jobs Report has appeared in 2018, 2020, 2023, and 2025, a slower and less regular rhythm. A ledger row built against the 2023 Future of Jobs figures or a prior AI Index edition should carry an explicit note of which edition it cites and a review trigger tied to that specific source's own publishing pattern, not a generic annual assumption, precisely because these are exactly the kind of actively maintained sources Section 3K's currency discipline is built to track, not one-off studies that never get revisited by their own authors.

Example 9: A single-source claim treated as if it were an industry consensus (a pattern, not one incident). A recurring failure across policy submissions in this field is a single individual's public remarks, an executive's conference-stage prediction, a founder's interview answer, cited as if it were a peer-reviewed finding or an industry-wide consensus, when the actual claim traces to one person's stated expectation with no dataset, sample, or published methodology behind it at all. Grading such a claim correctly does not mean the individual's view is worthless; a credible practitioner's stated expectation can be genuinely informative context. It means the restated claim must name exactly what it is, one individual's stated expectation, not a study's finding, which is the distinction Question 11 in Section 10 tests directly.

Where people go wrong

  • "If a respected organization is cited, the claim is safe to repeat as stated." The IMF is a credible, careful source; the claim that "AI will destroy 40 percent of jobs" is not what the IMF's own paper says. A respected source's name attached to a claim tells you the primary source is worth tracing; it does not tell you the version of the claim in front of you matches what that source actually found.
  • "A range is just a more precise single number, so quoting the top of it is fine." McKinsey's $2.6 trillion to $4.4 trillion range is not the same claim as "$4.4 trillion." A range communicates genuine uncertainty in the underlying model; collapsing it to its highest figure overstates the finding's precision and its confidence.
  • "The population doesn't matter if the headline number is accurate." A figure can be numerically correct and still misleading if its actual population (803 surveyed employers, 75 Singapore respondents within a larger global sample, a specific income-group breakdown) gets replaced with a broader implied population ("the world's workers," "businesses everywhere") in the retelling.
  • "Softening the verb is just being overly cautious." Changing "could affect" to "will affect" is not caution removed; it is a grade change, from a modeled projection to an unqualified assertion the underlying evidence does not support. The hedge is not decoration; it is the exact word carrying the claim's true evidence grade.
  • "An interested party's claim is either fully trustworthy or worthless, no middle ground." Both extremes are wrong. An interested estimate deserves neither blind acceptance nor blanket dismissal; it deserves the same five-step trace as any other claim, graded honestly as Grade I, with extra weight given to whatever independent corroboration exists.
  • "If I can't find the primary source quickly, it's safe to assume the secondary source got it right." An untraceable primary source is a finding, not a shortcut. The correct response is to cite the secondary source explicitly as secondary, soften the claim's grade toward Grade U, or keep searching, never to silently borrow the primary source's authority for a claim you were not actually able to verify yourself.
  • "A number that has been repeated everywhere for years must be solid by now." Wide repetition is a social fact about how a claim has traveled, not evidence about how it was originally produced. A widely repeated figure still needs its own trace; repetition can just as easily mean a Grade U assertion has achieved consensus-sounding status without ever having been checked.
  • "Grading evidence down is always the safer, more rigorous choice." Refusing to trace a claim that happens to support your position, and grading it down out of an excess of caution rather than an actual failed trace, is its own failure of the method. The grade must reflect what the evidence trail actually shows, in either direction, not a thumb on the scale toward whichever conclusion feels more defensible to have reached.
  • "Once a claim is graded and restated, it never needs to be revisited." A claim's grade can change: a projection gets superseded by a newer model, a measured count gets revised with better data, a methodology later faces a credible critique. An evidence ledger without review triggers is a snapshot that looks current long after it has stopped being current. (see Topic 6.6) (see Topic 6.10)
  • "Restating a claim honestly always makes the argument weaker." As Geraldine's scenario shows, a correctly graded and restated claim can make an argument more precise and, often, more persuasive to a skeptical or well-informed reader, because it survives the exact scrutiny an overstated version invites and fails.
  • "I only need to apply this method to claims I am suspicious of." The method applies evenly to every specific, checkable claim in a document, including the ones that already feel obviously true, because a claim's grade depends on how it was produced, not on how confident it feels to the person restating it. A claim that "everyone already knows" is exactly the kind of claim citation compression has had the most opportunity to quietly degrade.
  • "A claim that supports my organization's position deserves less scrutiny than one that complicates it." Applying the five-step method unevenly, generous with convenient claims and rigorous with inconvenient ones, defeats the discipline's entire purpose, which is a credibility standard that holds regardless of which direction a correctly graded claim happens to point.
  • "One individual's public remarks are as citable as a published study, provided the individual is a recognized expert." A credible individual's stated expectation is worth recording, but it must be restated as exactly that, one person's view, not upgraded into "research shows" or "studies confirm" language the underlying evidence does not support; conflating personal credibility with published methodology is a common way an unsupported assertion picks up borrowed authority it did not earn.
  • "A claim my own organization produced internally is automatically Grade M, since we generated it ourselves." An internally produced figure earns its grade from how it was actually produced, a direct count, a model, or an unverified guess passed along as fact, exactly like any external claim; being the source of a number does not exempt it from being traced.

Questions people ask

What is evidence ledger?
A working document that records, for every specific, checkable claim in a policy argument, its original wording, its primary source, its checked date and population, its evidence grade, and its restated form, ready to defend line by line. It is the artifact this topic teaches and the third payoff artifact of Module 6.
What is evidence grade?
One of four levels (Measured, Modeled Projection, Interested Estimate, Unsupported Assertion) assigned to a claim based on how the underlying figure was actually produced, never on how confidently the claim is stated.
What is measured (Grade M)?
A figure counted or directly observed, such as a tally of enacted regulations or a recorded enforcement action, reported by the party that did the counting. Still needs its population and date checked, but is not itself a forecast or a model's output.
What is modeled projection (Grade P)?
A figure produced by applying an explicit, published methodology to available data to estimate something that has not yet happened, such as an occupational exposure index or an economic-impact model. Legitimate evidence when the methodology is transparent, but always a model's output, restated with a hedging verb.
What is interested estimate (Grade I)?
A figure produced, funded, or prominently amplified by a party with a direct financial or reputational stake in the conclusion it supports, such as a vendor's own ROI case study. Not automatically wrong, but requires disclosed methodology and independent corroboration before it is cited without heavy hedging.

Keep going

Watch the lesson and prove it

Proving it happens inside the program: watch, explain it in your own words, and it goes on a record employers can check.

Watch the lesson and prove it →