Judgment Lab

Strengthening human judgment in AI-enabled work.

Failure mode 2 of 7

Fluency substitution

PME means professional military education. The Classroom version is also available for high school.

Two students in an invention course get the same task, the same inventor's notes and the same AI: draft patent claims for the inventor's wearable, the sentences that set what an invention protects. The course allows AI, and both say they used it. The first has the AI draft the claims, reads the draft closely and agrees with it. The second first writes down what's new and how it works, then has the AI draft from that. Suppose both hand in the very same claims. The same goes for any polished work a student hands in: change one fact, and you find out whether the reasoning in it is theirs. This is fluency substitution: polished analysis stands in for owned reasoning. Illustration based on Autor et al. (2026). The study followed practising patent lawyers; no students took part. The students, the course, the invention and the older patent are constructed.

1:41 · Read the figure

Fluency substitution, an animated figure in twelve beats

An animated figure that opens on its name, Fluency substitution, failure mode 2 of 7, and builds a table. A red dot stands for a student in an invention course with an example task: draft the patent claims for an inventor's wearable that reads heart rate. The course allows AI, and its use is disclosed. An AI drafts polished claims, shown in grey on a beige band, the figure's mark for words the AI wrote: claim reading through clothing broadly, since no earlier device does it; as a narrower claim, the sensor's fabric adjustment; as the fallback, the battery design. The student reads it closely and agrees with every line. A second student first writes down in a red note what is new and how it works, "it reads heart rate through clothing, by adjusting its sensor to each fabric", then has an AI draft from it. Suppose that draft matches the first one word for word. The instructor receives two identical drafts, both tagged as drafted with AI. With the drafts set aside, the instructor asks each student what they would claim and why, and both explain clearly: reading through clothing, which no device has done before. Their answers appear in plain ink, as the instructor would hear them. Then one fact changes: an older patent already covers reading through clothing. The first student takes the draft's fallback: lead with the battery design. The second leads with the sensor that adjusts to each fabric. Then the instructor adds that the older patent covers the battery too, and the battery answer is struck through: only the sensor is still new. Both drafts listed the sensor, but only the second student knew it was the new part. Only then do the answers take their colours, to show whose words they are: the draft's words in the first student's answers turn the AI's grey, and the students' own words turn red. The figure ends on the name and its definition, "Polished analysis stands in for owned reasoning.", then the instructor’s check and the follow-up, with a source line: illustration based on Autor et al. (2026). Each beat is listed below with what is on screen.

  1. Fluency substitution · Failure mode 2 of 7

    On screen: the name, "Fluency substitution", appears large over the stage with "Failure mode 2 of 7" above it, then moves into the top-left corner, where it stays.

  2. A student in an invention course drafts patent claims.

    On screen: a task card appears at the top left: "Example task: Draft the patent claims: what does the inventor's wearable protect?" A red dot labelled Student appears below it.

  3. The student asks an AI for a draft. It writes polished claims.

    On screen: the dot pulses. A sheet labelled "AI's draft" appears, and a paragraph writes itself line by line in grey on a beige band, and a key below names that band "Written by AI": "Claim reading through clothing broadly: no earlier device does it. Narrower claim: the sensor's fabric adjustment. Fallback: the battery design."

  4. The student reads it closely and agrees with every line.

    On screen: the dot reads along each line at a steady pace. Nothing is cut. The label changes to "Their claims, drafted with AI".

  5. A second student first writes down what's new and how it works.

    On screen: the first draft moves into a column headed First student. A second red dot heads a column labelled Second student and writes on a large note card in red, "What's new: it reads heart rate through clothing, by adjusting its sensor to each fabric.", and the key adds "Student's own work" in red.

  6. An AI drafts from the note. Suppose it matches the first draft word for word.

    On screen: the note moves into the top row of the second column. The dot pulses, and a sheet labelled "AI's draft" fills with the same grey paragraph, the narrower claim on the sensor included. Suppose it is the same as the first draft, word for word. The dot reads each line at the same pace, and a thin line pairs each line with the same line of the first draft.

  7. The instructor receives two identical drafts, made with AI.

    On screen: the second draft moves into its column beside the first, in a row labelled "What the instructor receives". The top row is labelled "Their own work / not handed in": empty for the first student, the note for the second. Both drafts are tagged "Their claims, drafted with AI".

  8. Without the drafts, both explain the claims clearly.

    On screen: both drafts dim. A row labelled "The instructor asks: Drafts aside: what do you claim, and why?" fills. In plain ink, the first student says, "Reading through clothing. No earlier device does it." The second student says, "Reading through clothing. No one has done that before."

  9. Change one fact.

    On screen: a new row appears: "Then one fact changes: What if an older patent covers reading through clothing?"

  10. Both drafts list the sensor. Only the second student knows it's new.

    On screen: in plain ink, the first student says, "Then lead with the battery design." The second student says, "Then lead with the sensor that adjusts to each fabric."

  11. But the older patent covers the battery too.

    On screen: under the answers, a line appears at the same size: "The instructor adds: “It covers the battery too.”" In the first student's answer, "the battery design" is struck through in ink. Of the claims both handed in, only the sensor that adjusts to each fabric is still new.

  12. The colours show whose words these are.

    On screen: the answers take their colours. Words the AI wrote turn grey on a beige band: all of the first student's first answer, and "the battery design". The same words in the first draft are underlined for a few seconds. The students' own words turn red: "Then lead with" and both of the second student's answers.

  13. Instructor’s check: Can the student explain their claims plainly, without the draft? Then change one fact. What would they claim now?

    On screen: both lines appear under the picture at the same size, and a source line, "Illustration based on Autor et al., NBER (2026)", appears under the table. The whole table stays in view.

Instructor’s check

Can the student explain their claims plainly, without the draft?

Then change one fact. What would they claim now?

Both students in the figure explain the claims clearly without the draft. The second line separates them. It comes from The Irreducible Officer, an essay written for military education, which looks for the person who "revises under challenge rather than retreating to the artifact's language".

The study behind this example

In a pre-registered three-month randomized trial, David Autor and colleagues gave "133 practicing patent lawyers at eleven U.S. intellectual property law firms" access to a custom AI patent drafting assistant. Of those, 91 completed the full protocol. Expert raters, blind to who had the tool, scored the work. The participants were practising lawyers; no students took part. Tested without AI after three months, senior lawyers gained, and junior lawyers split, with more poor and more good scores (below).

The paper's abstract states the finding: "AI access raised the quality of work delivered on benchmark patent drafting tasks at 10 days (0.34 SD, p = 0.03) and 90 days (0.38 SD, p = 0.01), with larger gains among junior lawyers."

At the end, the 91 lawyers who completed the study each reviewed and corrected an existing patent application without AI. "Treated lawyers outperformed controls by 0.32 SD (p = 0.04), but this advantage was concentrated entirely among senior lawyers (0.45 SD, p = 0.02). Junior lawyers showed no average gain. Their scores instead bifurcated, with sharply fewer mediocre scores offset by more poor and more good ones, suggesting that AI assistance served as a springboard for some juniors and a cushion for others."

So the split was among juniors, lawyers with fewer than seven years of experience (p. 23). Seniors gained: "Measured by unassisted judgment, only senior lawyers emerged from three months of AI use with improved capability" (p. 2). Fifteen of the 91 redlines showed possible AI use. Without them, the overall gain falls below conventional significance, while the senior gain holds at the 10 percent level (p. 3).

The authors are plain about the limit: "Our experiment does not answer why the effects appear positive for some lawyers and negative for others" (p. 2). It covers one profession, and the junior group was small. The tool was custom. The study was conducted by Google, Google funded its direct costs, and all eleven firms have patent-drafting relationships with Google.

The figure's two students show one way such a split could arise in a class. The study's participants were practising lawyers, and its tested results are scores. Its look at reasons, "post-experiment interviews with treated participants and an LLM-based qualitative analysis of redlining submissions" (p. 29), is in supplementary materials "not intended for publication" (p. 9, note 8). The authors float time spent as one possible mechanism: "Juniors using AI worked faster but showed no unassisted gains" (p. 30). The figure's account, writing down what's new before drafting and reworking the call from it when a fact changes, is an illustration; the study didn't test it.

Source: Autor, Rodchenko, Martin, Iscenko, Strand, Pearl & Ferere, "Does AI Assistance Enhance or Erode Expertise? Evidence from a Three-Month Field Experiment in Patent Drafting," NBER Working Paper 35720 (revised October 2026). Quotes are from the abstract unless a page is given. The students, the course, the invention, the claims, the older patent and the check are constructed.

From the original PME essay

Fluency substitution is easier to miss. AI produces the tone of analytic maturity (balanced paragraphs, caveats in the right places, the measured voice of a considered judgment) and the student mistakes well-ordered language for well-owned reasoning. Researchers studying AI's effects on student cognition have introduced the concept of epistemic confinement to describe this condition: an illusion of competence while operating entirely within AI-constructed analytical boundaries, where the student believes they are thinking independently while the frame doing the work was never theirs (Chow et al., 2026). In strategy, fluency substitutes for deciding. A paragraph that balances every consideration may never identify which risk actually matters most.

The Irreducible Officer, Section II, What the Problem Looks Like

Go further

All failure modes