The simple story about Amy Wax (b. 1953) is that she became more extreme, year by year, as the Internet rewarded her for it.
The coding does not support the year-by-year part.
I scored 173 substantive claims. The public corpus holds 129 claims from 21 appearances between August 2017 and March 2026. The scholarly comparison holds 44 claims from four works published in 2017, 2018 and 2020 on subjects that overlap the public conversations. Each claim carries scores for certainty, scope, qualification, evidence, engagement with counterarguments and concession, on the scales in the codebook. Each appearance carries scores for interviewer pressure and audience alignment. Public models use OLS with subject fixed effects and standard errors clustered by appearance.
Four processes come out of it, and one of them was not in my first reading of my numbers.
Scope broadens gradually across the decade. Certainty and qualification do not. They move as a step, and the step lands between August 2022 and March 2024. Medium predicts evidentiary friction. The interviewer predicts permeability.
That last group of findings makes Wax a demanding test case for Elias Aboujaoude (b. 1971) whose Virtually You: The Dangerous Powers of the E-Personality appeared in 2011. Aboujaoude argues that digital environments produce a bolder and less restrained personality, and that traits developed online migrate back into ordinary life. He credits the positive side. The virtual self can say the unpopular thing that offline inhibition suppresses. Wax tests him hard because the Internet did not supply her beliefs. It supplied the conditions under which she states them, and something other than the Internet appears to have supplied the timing.
The framework arrives in 1996
Any e-personality account has to survive her early publication record, and most such accounts do not.
Her first listed publication is a 1996 review essay in the University of Chicago Law Review on Robert Wright’s The Moral Animal. It reads as a manifesto. Evolved psychology sets boundary conditions on culture. Sexual asymmetry produces average differences in mating strategy, status competition and risk tolerance. Moral systems mobilize one evolved tendency against another. Communities enforce norms through shame, gossip, stigma and the withdrawal of regard rather than through formal punishment. Social policy dissolves inherited restraints more easily than it builds them. Every theme she pursues for thirty years sits in that essay.
The 1996 article on selective subsidies for two-parent homes and the 1998 article on marital bargaining carry the same anthropology into family law. The 1999 article on unconscious bias applies it to Title VII. By 2011, in “Disparate Impact Realism,” measured group differences in cognitive ability enter as an actuarial premise with causation bracketed.
None of that is a digital artifact. It predates her national controversy by two decades, and she kept the 1996 essay at the top of her CV after Penn stripped her chair.
The continuity runs at the level of framework and stops short of the conclusions. In 2009, in Race, Wrongs, and Remedies, Wax walks through the exogenous explanations for the Black-White test score gap, finds them wanting, and declines the hereditarian step. She rests the argument on culture and behavior. That refusal is a fact about the 2009 Wax that the later Wax abandons. So the honest statement has two halves. The framework is old. The application to racial cognition, to Asian immigration, to Jews, to women and to sexual minorities arrives later, and it arrives in speech before it arrives on paper.
Scope broadens on a slope
In the public corpus, each additional year adds about .121 points of claim scope on the five-point scale, with a clustered standard error of .031 and a p value near .0001. Over nine years that is roughly a full scale point.
The corpus shows what the coefficient means. Of the thirty-three public claims that score at maximum certainty and maximum scope together, thirty-one come from 2021 or later. Two come from the whole of 2017 through 2020. The early maximal claims concern immigrant assimilation and bohemian marginality. The late ones cover the sex distribution of civilizational achievement, the proper social position of homosexuality, the desirability of one Anglo-Protestant or pan-European dominant culture, and the loyalties of Asian migrants. The number of populations her claims sort keeps rising.
Broadening is where the digital-audience story does its best work. Scope does not differ between her scholarly and public claims once year and subject are controlled. In the 2017-2020 overlap window the public claims run slightly narrower. So the widening belongs to time rather than to venue, and time here means a decade of invitations to speak about race, sex, immigration and the universities and nothing else.
Certainty and qualification move as a step
The first pass through these data produced a tidy sentence: each year adds .118 points of certainty and removes .150 points of qualification. Both coefficients are strong. Both are also wrong about the process that generates them.
Add an indicator for 2024 and later to the certainty model and the indicator takes .51 points (p = .01) while the annual trend falls to .043 and loses significance. Do the same for qualification and the indicator takes 1.06 points (p = .014) while the annual trend drops to .005, which is nothing at all. Indicators placed at 2021, 2022 and 2023 do nothing. An indicator placed at 2025 does nothing. The discontinuity sits at 2024 and the linear trend was smoothing across it.
The step is not an artifact of who was interviewing her. Inside the eleven Loury conversations alone, the 2024 indicator takes .58 points of certainty (p = .001) and 1.04 points of qualification (p = .033), and the annual trend goes marginal or null in both. Dropping any single appearance and refitting leaves the certainty step between .35 and .64, significant in nineteen of twenty-one refits.
The Loury series shows it without any model at all. Mean qualification per conversation, in order, runs 2.17, 2.00, 2.00, 1.20, 2.00, 1.60, 1.67, 2.40. Then it runs 0.50, 0.29, 0.50. There is no downward slope across the first eight conversations, which span August 2017 to August 2022. There is a cliff between the August 26, 2022 conversation and the March 29, 2024 conversation, and the level never recovers.
What sits in that window is not a technology. It is Penn. The New York Times covered the dean’s complaint in March 2023. The hearing process ran through 2023. The faculty senate report leaked in early 2024, weeks before the March conversation. Penn suspended her at half pay and stripped the Mundheim chair in September 2024. She sued in January 2025.
That timing supports an explanation the e-personality frame does not contain. A person under formal charge cannot qualify her claims without conceding the charge. Caveats are admissions. The lawsuit, filed January 16, 2025, dismissed August 27, 2025, and appealed to the Third Circuit on January 15, 2026, makes retreat a litigation cost on top of a coalition cost. On the evidence here, the sanctions process is a better candidate for the discontinuity than the accumulated effect of a decade of podcasting.
I cannot separate the two cleanly. Her friendly-platform appearances also cluster in 2024. But the step survives inside the Loury series, where the platform is constant and the host is the most persistent challenger in the corpus, and that is the strongest available test.
Counterargument engagement rises
One result cuts against the hardening story.
Engagement with counterarguments rises .089 points per year in the public corpus, with a clustered standard error of .033 and p = .007, after controlling for interviewer pressure. The later Wax addresses opposing arguments more often, not less. Outside the Loury conversations the annual rise is .335 points.
Set beside the certainty step, that gives a specific picture. Wax states her conclusions more categorically and with fewer caveats than she once did, and she spends more of her time answering objections. Those two things travel together in a person who has been arguing the same case under attack for six years. She has heard the objections. She has answers ready. The answers do not narrow the claim.
The medium result
At the appearance level it looked as though Wax became less evidentiary over time. Once subject and interview environment enter the model, that disappears. The annual coefficient on evidence is -.022 with p = .46. There is no chronological decline in evidentiary support.
The gap belongs to medium, and it is the largest effect in the study. Restricting to 2017 through 2020, where scholarly and public Wax can be observed at once, and controlling for year and subject, public claims run .38 points more certain, carry 1.04 fewer points of qualification, carry 1.47 fewer points of engagement with counterarguments, and carry 2.96 fewer points of evidentiary support on the five-point scale. Raw means across the full corpus: evidence 3.82 in the scholarly claims against 0.74 in the public ones.
The academic Wax already made large claims. What the scholarly venue demanded was an argument built around the claim.
One 2020 comparison shows the split inside a single year. On July 31 she tells Loury she would bet that innate factors are at work, which reports where she puts her own probability. Months later, in Pursuing Diversity, she lists racism, poverty, culture and innate group differences as competing explanations and declines to apportion causal weight. A decade of radicalization cannot explain a difference that occurs within twelve months. Podcast Wax tells you where her money sits. Law-review Wax tells you which explanations remain live. Neither version is counterfeit. The media activate different standards for converting belief into assertion.
The demand side
Aboujaoude supplies the half the coding cannot reach.
His chapter on the illusion of knowledge collects the research on how people read online. The 2008 British Library study of five years of digital trails found readers skimming, bouncing and abandoning: four minutes per e-book, sixty percent of academics viewing three pages or fewer of a journal article, sixty-five percent never returning. Jakob Nielsen’s eye-tracking work found that online readers scan in an F pattern and forage rather than read. Aboujaoude’s summary of the British Library finding is that people seem to go online to avoid reading.
Put that beside an evidence score of 0.74 and the account closes. Wax’s citations vanish in conversational media partly because she need not supply them and partly because the audience will not follow them. The listener who hears an unnamed supporting study does not go looking. A ninety-minute podcast still obeys the logic of compression, because compression happens downstream, in the hands of whoever cuts the clip, and the sentence that escapes is never the qualified one.
That is the defensible version of Aboujaoude’s thesis applied here. The network makes certain versions of Amy Wax cheaper to produce, circulate and repeat.
A distributed avatar
Wax is not a social-media intellectual, which is part of why she tests the book well.
Her e-personality is assembled by other people: Bloggingheads, YouTube, podcasts, newspaper interviews, conference video, university statements, student campaigns, reaction clips. Others upload her, clip her, title the episode and tell the audience what kind of person is coming.
Several Wax avatars circulate at once. The courageous dissenter. The racist professor. The race realist. The free-speech martyr. The civilizational conservative. The scholar punished for saying what others fear to say. The faculty member whose conduct Penn judged incompatible with professional standards. Each draws on something real. None matches the person. The network promotes whichever version serves the community circulating it, and the representation acquires causal weight of its own. Aboujaoude calls the fusion of the two selves virtualism, and by September 2024 the mediated Wax and the employed professor had stopped being separable in any institutional sense.
What the interviewer does
My original hypothesis was that the interlocutor predicts Wax’s rhetorical behavior better than time does. That was wrong. With year and subject controlled, interviewer pressure shows no relationship with certainty (-.044, p = .53), scope (.050, p = .36) or qualification (.025, p = .86).
Pressure predicts one thing, strongly. A one-point increase adds .648 points of engagement with counterarguments (p well under .001). That is the largest host coefficient in the study. I scored substantive challenge rather than hostility, because Loury is warm toward Wax personally and remains her most persistent critic.
Four scenes carry it better than the coefficient.
On July 31, 2020, Wax tells Loury she would bet on innate factors. Loury answers within a minute. He would bet that Massachusetts outscores Mississippi, and he says that fact is inadmissible as political discourse because Americans share a country. Wax moves to the politicization of Eric Turkheimer and never returns. She concedes nothing. The listener hears the objection at full strength.
On January 24, 2022, she tells Gad Saad (b. 1964) that the Ashkenazi advantage in tested intelligence is partly innate. Saad says he feels marginalized by the comment. The joke is the entire resistance. She extends the claim to Asian-White gaps and moves on. That conversation carries a mean qualification score of 0.20 and a mean counterargument score of 0.00, the only zero in the corpus.
On August 29, 2022, Richard Hanania (b. 1985) asks why men permitted the feminization of institutions. Wax answers that you can only impose something on people who allow it, and says that part she has never understood. She names an unsolved problem inside her own framework. She does not do that with Saad, and she does not do it with Louise Perry (b. 1992).
On August 14, 2024, Perry asks whether women should return to five percent of the student body. Wax says ideally, then routes the answer back through meritocratic procedure. No resistance follows the ideal.
Same speaker, same framework, different rooms, different products.
Locally corrigible, globally stable
Of the 129 public claims, concession scores zero in the great majority. Fourteen show factual or local acknowledgment. Six show local narrowing. One shows movement in a supporting conclusion. Nothing in the corpus shows Wax abandoning the principal conclusion around which an appearance is organized. Interviewer pressure predicts concession only weakly (.109, p = .078).
The study cannot show that Wax never changes her mind. It shows that in these rooms, challenge produces local revision. Loury can get her to grant a fact, accept an exception, acknowledge uncertainty, or say what follows if a premise weakens. The next step rarely arrives: yes, that is true, and therefore the conclusion I brought into this room has to change.
That is why hard interviews keep their value when nobody switches sides. A good interlocutor may fail to move her and still improve what the audience receives.
What is wrong with this study
The reliability audit is the most useful thing in the workbook, and it is worse news than it first looks.
Re-coding sixty randomly selected claims with date, host, medium, subject and first-pass scores hidden produced quadratic weighted kappas of about .75 for certainty, .65 for scope and .78 for qualification. Those are the variables carrying the hardening results, and they hold up. Evidence came in at .38 and counterargument at .43.
The disagreement is not random. Comparing the two passes, the blinded scores run systematically lower: evidence by 1.10 points on average, counterargument by 0.83, scope by 0.45, certainty by 0.15. Knowing the source inflates the score. That threatens the medium comparison directly, since that comparison rests on absolute levels of evidence, and a coder who knows a claim comes from a law review is primed to find the argument around it. The 2.96-point medium gap survives even if the entire 1.10-point drift is attributed to venue cues, but it shrinks, and the honest range is roughly two to three points rather than three.
This is also a same-model re-code rather than an independent human coder. It measures the stability of the instrument, not agreement between judges.
Second, pressure and counterargument engagement were scored from the same transcripts in the same pass. A coder who hears Wax address an objection is likelier to score the host as challenging. The .648 coefficient is inflated by an unknown amount. The fix is cheap: score host pressure from host turns alone, with Wax’s turns removed, before scoring hers.
Third, pressure and alignment are nearly collinear with host identity. Loury appearances score 2 or 3 on alignment; Saad, Alexander Riley, Nathan Cofnas, Freddy Gray and Hanania all score 4. The correlation between pressure and alignment across appearances is -.69. A model controlling both is close to a model controlling for whether the host is Loury, and there is little independent variation left to identify either effect.
Fourth, the instrument saturates. In 2024, eighty-three percent of certainty scores sit at the ceiling of 4, sixty-nine percent of scope scores at the ceiling of 5, and sixty-six percent of qualification scores at the floor of 0. After 2024 the scales cannot register further hardening even if it occurs. Ordered logit reproduces the year effect on certainty, so the finding is not an artifact of treating ordinal scores as continuous, but every coefficient in the late period is attenuated.
Fifth, subject composition drifts hard. Race and achievement supply nine of twenty claims in 2017 and none of the gender claims; gender and family supply nine claims in 2022 and eight in 2024 and none before 2022. Education and poverty vanish after 2019. Several year-by-subject cells are empty. Subject fixed effects are absorbing variance that may belong to the process under study, because the topics she is invited to discuss are downstream of the same forces that produce the hardening.
Sixth, the scholarly corpus has three distinct time points and ends in 2020. If Wax published no comparable scholarship on these subjects afterward, then medium and late period are partly collinear across the full corpus, and the exit from scholarly venues is a finding rather than a nuisance. Aboujaoude’s migration thesis predicts that exit.
So the findings rank. The 2024 discontinuity is the most secure result and the one I did not expect. Scope broadening is secure. The medium gap is large enough to survive the coding drift, with its magnitude uncertain. The permeability result needs host-only coding by an independent human coder before its coefficient deserves any weight.
The study that should come next
The strongest available control sits in the transcripts already collected.
Loury (b. 1948) appears in eleven of these conversations as interviewer. He is also a public intellectual of the same generation, on the same platforms, across the same decade, discussing the same subjects, and he faced no disciplinary proceeding. Score his claims with the same instrument.
The predictions separate the hypotheses. If Loury shows the same 2024 step, the finding is a period effect and belongs to the political moment. If his qualification holds flat across 2022 to 2024 while hers collapses, the step belongs to the sanctions and the litigation, and the e-personality frame is the wrong frame for the timing even if it remains right about the medium. If his evidentiary support falls in podcasts and holds in his written work, the medium result generalizes past one speaker, which is what would turn it into a claim about public intellectual life rather than a biography.
The third result is the one worth chasing.
The better Amy Wax story
The Internet did not turn a conventional scholar into a person with different beliefs. Her framework is on paper in 1996 and has not moved much.
Two things changed, and they changed differently. The range of the claims widened steadily across the decade, one group at a time, in step with the invitations. The restraint around the claims did not erode gradually. It broke, in the window between August 2022 and March 2024, while Penn’s disciplinary machinery was running toward its conclusion.
Medium explains where the evidence went. Time explains the widening. The sanctions process is the best available explanation for the timing of the collapse in qualification, and the coding cannot rule out that a person who has been formally charged simply stops offering caveats because caveats are concessions.
Academics do not hold opinions and then distribute identical copies through different channels. The medium takes part in producing the opinion as a public object. A proposition in a law review arrives with citations, boundaries and named adversaries. The same proposition in a podcast arrives as a declaration. Repeat the declaration, reward it with invitations, attacks, clips, followers, a hearing board and a federal docket, and it stops being something the person believes and becomes part of who she publicly is. At that point retreat costs more than an argument. It costs a character, and after January 2025 it costs a case.
Aboujaoude asks the right closing question himself, near the end of Virtually You. He wonders whether the online self is in some sense more real than the real thing, whether the disinhibited version returns us to what we were before the social contract taught us to hold our tongues. He leaves it open, and so do these data.
Wax’s e-personality is Amy Wax under a different set of restraints and rewards. The record shows when the restraints came off.
