In fields outside of literary studies (my field), the so-called “Big Five” personality traits—openness, conscientiousness, extraversion, agreeableness, neuroticism—seem, to many smart people, to be useful and explanatory. I have never found them to be useful or explanatory. But they are psychology’s standard answer to the question of what a person is like, and since there are more psych majors than English majors, social scientists have gotten to decide how to talk about personality for the last 90 years. It’s a shame, because now that AI firms are building AI personalities (for tutors, companions, cranky patient-bots for health practitioner training), what are researchers reaching for? Big Five personality dimensions, in some admixture or other, usually warm, conscientious, open, and not too neurotic, except in the case of the cranky patient.1 This is the wrong way to build a language-model personality.
The origin of the Big Five is a dictionary, which is, in LLM terms, an alphabetized projection of a vector space where the words are positions in semantic geometry, each word’s definition is its cosine similarities to all the others, and the axis for retrieval is alphabetical order, orthogonal to every axis of meaning. In 1936 Gordon Allport and Henry Odbert went through Webster’s New International Dictionary and extracted every word English uses to describe a person. The file added up to 17,953 terms. Their premise, known now as the lexical hypothesis, was that important aspects of human differences would have left a trace in language. Words are needed for things that matter, like distinctions between persons.2 Some of you may have studied this in college.
The next goal, because of compute limits, was to compress the database to a usable number. Raymond Cattell, in the early 1940s, took roughly 4,500 of the ones naming stable traits rather than moods or evaluations and merged synonyms. He had people rate acquaintances on the remaining clusters, ran factor analysis, and found sixteen.3 In 1949 Donald Fiske, working with Cattell’s list, seemed to see only five.4 In 1961 Ernest Tupes and Raymond Christal reanalyzed trait ratings from eight different samples and found the same five factors.5 Warren Norman replicated the five in 1963 with university samples, and for a time the factors were called “Norman’s five.”6
In 1968 Walter Mischel (of the Stanford marshmallow test fame) reviewed decades of studies and found that personality trait scores did not predict how a person would react in a particular situation. Arguments raged for decades over whether personality could be measured in traits (or rather clusters of words).7 My college psych professor never taught any of this so he was likely on the Mischel side.
But somehow the five factors idea stayed alive because they provided a convenient tool for measuring people. (This is how the social sciences work.) Lewis Goldberg kept finding the same five clusters while replicating studies through the 1970s and 1980s, naming them the “Big Five” in 1981.8 Paul Costa and Robert McCrae built a three-factor questionnaire—neuroticism, extraversion, openness (the NEO) — and expanded it to five in 1985, issuing the revised NEO PI-R, in 1992.9 Oliver John and colleagues built the Big Five Inventory in 1991, a widely used self-report instrument whose items include “I see myself as someone who is talkative,” rated on a five-point scale.10 Goldberg published his own adjective markers for the five factors in 1992 and later helped build the International Personality Item Pool, now in the public domain, free, standardized, and easy to reuse. The IPIP keeps the Big Five handy for new research tasks.11
To reiterate: five-factor instruments turn adjectives and short statements into generalized ratings of one target person: “talkative,” “envious,” “systematic,” “starts quarrels with others,” “has a forgiving nature.” The rater may be the person or an observer, but every answer is still an attribute attached to the person. Quiet? Quiet toward whom? Quiet since when? Quiet why? The score requires meanings to be stable and comparable across subjects. Anything eccentric goes into the error term.
If you are reading this with a good understanding of how LLMs work, you see where I’m going with this.
Sure, for humans, the five-factor model seemed to have some predictive ability. A 1991 meta-analysis showed conscientiousness was a predictor of job performance across occupations.12 A 2007 review showed that certain factors predict divorce, occupational attainment.13 But so do socioeconomic status, cognitive ability, and health.
Other researchers saw the whole thing as bad design from the beginning. Jack Block argued in 1995 that the five factors were mere artifacts of the variable lists fed into the factor analysis; Dan McAdams called the model simply words you could say about a person you have never met. As expected, the test failed outright when given to people without a shared dictionary, as a study involving 632 Tsimane forager-horticulturalists in Bolivia showed.14
Today, one can label the original Allport-Odbert scavenger hunt as a high-dimensional lexical deposit. Psychology compressed a selected portion of it to five axes averaged across people, with much of the remaining variation thrown away as residue. LLMs encode semantic relations among vastly more than 17,953 words. Why give that system the five-axis artifact and ask it to impersonate the compression? Why not give the model the whole vocabulary and ask it to fold it in new ways, like proteins? Without the inherited compression, dimensionality becomes an empirical question again.
The whole list is hard to find but this week Fable found a list of 2,818 terms (from one of the cullings) and came up with this:
I read all 2,818 words and folded them by a principle factor analysis cannot use: not what the word ascribes but what the word presupposes — what has to exist for the word to apply. The Big Five’s folds sort by content (sociability words together, anxiety words together). Fold by presupposition and the list comes apart differently:
Words that presuppose a second person, present: condescending, patronizing, deferential, obsequious, ingratiating, fawning, cajoling, placating, appeasing, sycophantic, flirtatious, coquettish, teasing, browbeating, bullying, insinuating, intimidating. You cannot be condescending alone in a room. These words encode a dyad — who is being addressed and from what position.
Words that exist only in an audience’s reaction. The list contains the pairs that expose this fold: annoyed and annoying, frightened and frightening, embarrassed and embarrassing, confused and confusing, disgusted and disgusting, encouraged and encouraging. The -ed word is a state of the person; the -ing word is an effect on someone else — charming, boring, amusing, engaging, disarming, exasperating name nothing inside the person at all. A self-report instrument is structurally blind to this entire fold: “I am charming” is not a report, it’s a boast. Factor analysis of self-report treats both members of each pair as predicates of the respondent.
Words that presuppose a history: vindictive, retaliative, forgiving, grateful, remorseful, penitent, embittered, jaded, disillusioned, heartbroken, careworn, burned-out, loyal, faithful, rueful. Each requires events in an order — something happened, then the disposition.
Words about the person’s relation to their own prior statements: the list contains self-correcting, self-consistent, self-examining, self-excusing, self-deceiving, self-critical, self-deprecating, equivocating, apologetic. The lexicon has a whole vocabulary for self-audit; the Big Five has no dimension for it.
Words governing disclosure: secretive, guarded, evasive, cagey, tight-lipped, reticent, cryptic, elusive, noncommittal, concealing, disclosing, divulging, self-revealing, candid, frank, forthright, confiding. These regulate what is withheld from whom and form a coherent fold that the five factors scatter across “extraversion” and “agreeableness.”
Words that re-price: sarcastic, sardonic, ironical, satirical, jeering, scoffing, sneering, pooh-poohing, derisive, disparaging, denigrating, irreverent, reverent, adulatory, flattering, idolatrous, blasphemous. Allport and Odbert classified evaluative and “censorial” terms into Column III and the tradition dropped them at the first cut as not being traits.
That last fact is the finding. The compression didn’t lose the relational, historical, addressed, and valuing vocabulary by accident at step forty; it discarded it at step one, as category error, and kept the words that predicate a lone individual in a vacuum. The five factors are the factor structure of the vacuum words. The folds above are where characters live, they were in the dictionary the whole time, and the answer to “does this make sense?” is: the lexicon itself is organized by presupposition — by scene — and the trait tradition kept only the words that need no scene.
Thank you, Fable! A portion of the 17,953 words are relational, which matters for the next generation of models and agents. ‘Condescending’ requires someone to be condescending to. ‘Forgiving’ requires someone who hurt you. ‘Evasive’ requires someone from whom you’re withholding something. ‘Sarcastic’ requires a listener. Looking at attributes in isolation means erasing relationships, as Block and others have shown.
In 2022, Andrew Cutler and David Condon found the complete 17,953 list and fed it into an LLM. But mostly they asked the machine to cluster them by likeness, in essence to make a dictionary into a thesaurus. Those already exist so it wasn’t much of a finding. ‘Talkative’ sits near ‘outgoing,’ so when people check both boxes all anyone can truly conclude is that they know their synonyms.15
So why are AI researchers still reaching for the big five? Because machine-learning experiments need data, and decades of big five research means there are validated questionnaires, databases of human responses, tons of scores. It looks like a serious method. It is not, which is why it hasn’t yielded much.16
For example, in July 2023, Greg Serapio-García, Mustafa Safdari, and colleagues at Google DeepMind and the Psychometrics Centre at Cambridge administered the IPIP-NEO-300 and the Big Five Inventory to Google’s PaLM models, subjected the results to the human validation battery, and then “shaped” the traits with prompts built from 104 trait adjectives at nine intensity levels, verifying that questionnaire scores moved accordingly. Each item was administered in isolation so that no answer could influence the next.17
The validation is circular. The prompt supplies adjectives such as ‘extraverted’ or ‘agreeable’; the inventory then asks the model to endorse sentences whose meanings overlap with those adjectives. Basically, a language model can grasp what the instruction is if you change the phrasing and use a synonym. This capability has long been built into language models, for which I am glad.
Nearly every study of personality in language models now follows the same protocol: assign a persona, administer an inventory descended from the dictionary count, read the traits off the scores. They work (or rather don’t work) as expected.
Tom Sühr and colleagues found the more basic problem: models agreed with “I am an extrovert” and “I am an introvert,” a pattern rare for humans. People generally try to give a coherent account of themselves. Models often optimize for local conditions. The five-factor inventory can’t take this into account.18
Frisch and Giulianelli’s AI personas drifted after a writing task and collapsed into each other in interaction; PTCBench’s scores moved under simulated life events; Wang, Lester, and Srivastava found that narrative framing explained five to thirty-one times more behavioral variance than the assigned persona. Rajput and colleagues showed near-random results. All of these are failures of consistency, according to the Big Five method.19
By this summer, researchers seem to have finally seen the problem. Kim Zierahn and colleagues administered Big Five inventories to 244 models and found that only three percent of score variance lay between models. (They found four of the five factors collapsed into one.) The conclusion? Big Five scores do not measure the same thing in LLMs that they do in humans. Haoze Liu and colleagues replaced first-person questionnaires with 3,200 concrete scenarios and twenty "behavioral modes," arguing that the regularities encoded in human traits need not transfer to models. This is very close to my argument. But their unit of analysis remains the individual scenario.20
A naysayer might ask why the Big Five has even some predictive power. Because the instrument was extracted from the thing being tested. The inventory is a cluster of words and people do describe themselves with words. But it’s an outdated cluster of words in the AI era.
The compression from 17,953 words to five is understandable given 1940s statistics. But the whole point of large language models is that they’re large and can do big things. Maybe the lexical project can finally begin. Allport’s premise is the premise of LLMs: that personality traits and language have something to do with one another. LLMs have a hell of a lot more than 17,953 words but like I asked Fable above, it’s not a bad place to start.
The words give the model a repertoire. Fiction writers know better than psychologists how to create a personality that persists for thousands of years from an oral poem to a Christopher Nolan movie. Character is built via a sequence of encounters in language, not an isolated trait item. (The creation of persistent fictional characters is subject of my next paper.)
Calling correlations of self-descriptions a model of personality has always been a mistake. That ontology should be jettisoned in the LLM era. You cannot instruct the element in the activity it constitutes. The Big Five was just a way of swimming, however flawed.
The founding method is Greg Serapio-García, Mustafa Safdari, and colleagues at Google DeepMind and Cambridge’s Psychometrics Centre, “Personality Traits in Large Language Models,” arXiv:2307.00184 (2023), published as “A Psychometric Framework for Evaluating and Shaping Personality Traits in Large Language Models,” Nature Machine Intelligence 7 (2025): 1954–1968: personality is assigned via trait adjectives at a 1-9 scale and verified by questionnaire. Marwa Abdulhai and colleagues, “Consistently Simulating Human Personas with Multi-Turn Reinforcement Learning,” NeurIPS 2025, arXiv:2511.00222 (simulated AI patients for training AI therapists).
Gordon W. Allport and Henry S. Odbert, “Trait-Names: A Psycho-lexical Study,” Psychological Monographs 47, no. 1 (1936).
Raymond B. Cattell, “The Description of Personality: Basic Traits Resolved into Clusters,” Journal of Abnormal and Social Psychology 38 (1943): 476–506. The sixteen-factor result is the Sixteen Personality Factor Questionnaire (16PF, 1949).
Donald W. Fiske, “Consistency of the Factorial Structures of Personality Ratings from Different Sources,” Journal of Abnormal and Social Psychology 44 (1949): 329–344.
Ernest C. Tupes and Raymond E. Christal, “Recurrent Personality Factors Based on Trait Ratings,” USAF ASD Technical Report 61-97 (Lackland Air Force Base, 1961); reprinted in Journal of Personality 60, no. 2 (1992): 225–251.
Warren T. Norman, “Toward an Adequate Taxonomy of Personality Attributes: Replicated Factor Structure in Peer Nomination Personality Ratings,” Journal of Abnormal and Social Psychology 66 (1963): 574–583.
Walter Mischel, Personality and Assessment (New York: Wiley, 1968). The resolution came in Walter Mischel and Yuichi Shoda, “A Cognitive-Affective System Theory of Personality,” Psychological Review 102, no. 2 (1995): 246–268, which relocated personality from average behavior to stable if-then patterns of situation and response.
Lewis R. Goldberg, “Language and Individual Differences: The Search for Universals in Personality Lexicons,” in Review of Personality and Social Psychology, vol. 2, ed. Ladd Wheeler (Beverly Hills: Sage, 1981), 141–165. The consolidating review of the decade is John M. Digman, “Personality Structure: Emergence of the Five-Factor Model,” Annual Review of Psychology 41 (1990): 417–440.
Paul T. Costa and Robert R. McCrae, NEO Personality Inventory (Psychological Assessment Resources, 1985); Revised NEO Personality Inventory (NEO PI-R) and NEO Five-Factor Inventory (NEO-FFI) professional manual (1992).
Oliver P. John, Eileen M. Donahue, and Robert L. Kentle, “Big Five Inventory” (Berkeley: Institute of Personality and Social Research, 1991). The current revision is Christopher J. Soto and Oliver P. John, “The Next Big Five Inventory (BFI-2),” Journal of Personality and Social Psychology 113, no. 1 (2017): 117–143.
Lewis R. Goldberg, “The Development of Markers for the Big-Five Factor Structure,” Psychological Assessment 4, no. 1 (1992): 26–42; the International Personality Item Pool was introduced in Goldberg, “A Broad-Bandwidth, Public Domain, Personality Inventory,” in Personality Psychology in Europe, vol. 7 (1999). The 300-item IPIP-NEO is an open-source representation of the Revised NEO Personality Inventory.
Murray R. Barrick and Michael K. Mount, “The Big Five Personality Dimensions and Job Performance: A Meta-Analysis,” Personnel Psychology 44, no. 1 (1991): 1–26.
Brent W. Roberts, Nathan R. Kuncel, Rebecca Shiner, Avshalom Caspi, and Lewis R. Goldberg, “The Power of Personality: The Comparative Validity of Personality Traits, Socioeconomic Status, and Cognitive Ability for Predicting Important Life Outcomes,” Perspectives on Psychological Science 2, no. 4 (2007): 313–345.
Michael Gurven, Christopher von Rueden, Maxim Massenkoff, Hillard Kaplan, and Marino Lero Vie, “How Universal Is the Big Five? Testing the Five-Factor Model of Personality Variation Among Forager-Farmers in the Bolivian Amazon,” Journal of Personality and Social Psychology 104, no. 2 (2013): 354–370. The sample: 632 Tsimane adults completing a Tsimane translation of the 44-item Big Five Inventory, with a second sample of 430 rating their spouses; support for the five-factor structure failed on internal consistency, response stability, external validity against observed behavior, factor analysis, and comparison to the U.S. target structure. Jack Block, “A Contrarian View of the Five-Factor Approach to Personality Description,” Psychological Bulletin 117, no. 2 (1995): 187–215; Dan P. McAdams, “The Five-Factor Model in Personality: A Critical Appraisal,” Journal of Personality 60, no. 2 (1992): 329–361.
Andrew Cutler and David M. Condon, “Deep Lexical Hypothesis: Identifying Personality Structure in Natural Language,” arXiv:2203.02092 (February 2022). Each adjective was inserted into a fixed sentence (“Those close to me say I have a [MASK][MASK] and [TERM] personality”) and represented by the model’s internal summary of the word’s usage. The usual clustering method was applied. Three lists were run: the 435 terms of Saucier and Goldberg (1996), Goldberg’s 1,710 (1982), and the Allport-Odbert 17,953 (1936). Agreement with the survey-based clusters was 0.89, 0.79, and 0.79 for the first three; the fourth and fifth “explain a trivial amount of variance” with content “inconsistent and hard to interpret.”
Marilù Miotto, Nicola Rossberg, and Bennett Kleinberg, “Who is GPT-3? An Exploration of Personality, Values and Demographics,” arXiv:2209.14338 (September 2022). Xingxuan Li, Yutong Li, Shafiq Joty, Linlin Liu, Fei Huang, Lin Qiu, and Lidong Bing, “Does GPT-3 Demonstrate Psychopathy? Evaluating Large Language Models from a Psychological Perspective,” arXiv:2212.10529 (December 2022). The HEXACO model is Michael C. Ashton and Kibeom Lee’s six-factor extension of the lexical taxonomy. Graham Caron and Shashank Srivastava, “Identifying and Manipulating the Personality Traits of Language Models,” arXiv:2212.10276 (December 2022); published as “Manipulating the Perceived Personality Traits of Language Models,” Findings of the Association for Computational Linguistics: EMNLP 2023, 2370–2386.
Greg Serapio-García, Mustafa Safdari, Clément Crepy, Luning Sun, Stephen Fitz, Peter Romero, Marwa Abdulhai, Aleksandra Faust, and Maja Matarić, “Personality Traits in Large Language Models,” arXiv:2307.00184 (July 2023); published as “A Psychometric Framework for Evaluating and Shaping Personality Traits in Large Language Models,” Nature Machine Intelligence 7 (2025): 1954–1968. Instruments: the IPIP-NEO-300 and the Big Five Inventory. The item-isolation statement is from their section 2.1.1.
Tom Sühr, Florian Dorner, Samira Samadi, and Augustin Kelava gave the standard inventories to GPT-3.5, GPT-4, and Llama 2 and compared the results against more than a million human responses. Tom Sühr, Florian E. Dorner, Samira Samadi, and Augustin Kelava, “Challenging the Validity of Personality Tests for Large Language Models,” arXiv:2311.05297 (November 2023); published in Proceedings of the 5th ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization (2025). Human comparison: 1,015,342 responses to the 50-item IPIP Big Five Markers; model instruments included the BFI-2. Their conclusion: “current LLMs do not necessarily have human personality or the personality tests designed for humans do not measure the same latent variable for LLMs. Both conclusions render the application of personality tests designed for humans to current LLMs futile.”
Ivar Frisch and Mario Giulianelli built personas from single sentences of Big Five pole terms — “introverted, antagonistic, unconscientious, emotionally stable and closed to experience” — measured with Big Five Inventory statements, and found the personalities drifted after a writing task and collapsed into each other in interaction. Ivar Frisch and Mario Giulianelli, “LLM Agents in Interaction: Measuring Personality Consistency and Linguistic Alignment in Interacting Populations of Large Language Models,” Proceedings of the 1st Workshop on Personalization of Generative AI Systems (2024), 102–111; arXiv:2402.02896. PTCBench measured trait stability with the NEO Five-Factor Inventory under simulated life events and found the models moved orders of magnitude more than humans. Jiongchi Yu, Yuhan Ma, Xiaoyu Zhang, Junjie Wang, Qiang Hu, Chao Shen, and Xiaofei Xie, “PTCBench: Benchmarking Contextual Stability of Personality Traits in LLM Systems,” arXiv:2602.00016 (January 2026). Instrument: the NEO Five-Factor Inventory. Prateek Rajput and colleagues tried to install personalities by fine-tuning five models on essays written by people with known Big Five profiles; accuracies hovered near the random baseline. Prateek Rajput, Yewei Song, Iyiola E. Olatunji, Jacques Klein, and Tegawendé F. Bissyandé, “Evaluation Drift in LLM Personality Induction: Are We Moving the Goalpost?,” arXiv:2605.16996 (May 2026). Yixuan Wang, James Lester, and Shashank Srivastava assigned ten trait-style personas drawn from learning-style theory — a small trait taxonomy from education research, a cousin of the Big Five built by the same method of self-reported dispositions presumed stable across situations — and found that narrative framing explained five to thirty-one times more behavioral variance than the assigned persona. Yixuan Wang, James Lester, and Shashank Srivastava, “The Story Shapes the Agent: Narrative Priors in LLM Behavior,” arXiv:2607.18566 (July 2026).
Kim Zierahn, Cristina Cachero, Anna Korhonen, and Nuria Oliver, “Personality Without Persons? A Psychometric Critique of Big Five Testing in Large Language Models,” arXiv:2607.02325 (July 2026); their position paper is “LLMs Aren’t Human: A Critical Perspective on LLM Personality,” Proceedings of ACM CHI 2026, arXiv:2603.19030. Haoze Liu and colleagues, “Your LLM, Your Style: Behavioral Mode Axes for LLM Behavioral Control,” arXiv:2608.10703 (August 2026).
Hollis, spurred on by your excellent essay, I prompted my Ai with this:
I found the MBTI interesting and useful for certain purposes. Couldn't the MBTI be constructed in vector space instead of having just four categories and 16 summary boxes, thereby rendering it more interesting for psychological personality research?
The Ai response:
Yes, and in fact that move eliminates one of the strongest objections to MBTI. Once you stop forcing people into 16 boxes, the underlying four MBTI scales can be treated as continuous coordinates.
Imagine a person represented not as INTP, but as a vector:
p=(E/I, S/N, T/F, J/P)\mathbf{p}=(E/I,\;S/N,\;T/F,\;J/P)
with each coordinate ranging, say, from −1 to +1. Someone might therefore be
p=(−.82, +.15, −.61, +.07)\mathbf{p}=(-.82,\;+.15,\;-.61,\;+.07)
That says considerably more than a four-letter label. The person is strongly toward one pole on dimensions 1 and 3 but essentially near the middle on dimensions 2 and 4. The familiar MBTI "type" would simply identify the orthant of four-dimensional space containing that point. The 16 types are just the 24=162^4=16 regions produced by cutting four continuous dimensions at zero.
And now something interesting happens: the notorious MBTI boundary problem largely disappears. Two people scoring −.01 and +.01 on a dimension are recognized as virtually identical rather than being declared different "types."
But would this make MBTI scientifically preferable to the Big Five?
Not automatically. It would make it much more defensible, but we'd still have to ask whether those four axes are the best dimensions along which to describe personality.
The Big Five is already doing essentially what you're proposing. A person's Big Five scores constitute a vector in five-dimensional personality space:
B=(O,C,E,A,N)\mathbf{B}=(O,C,E,A,N)
The scientific question therefore becomes not 4 dimensions versus 16 boxes, but:
Which coordinate system best captures psychologically meaningful variation among people?
That's an empirical question.
And MBTI actually fares better under that interpretation than its reputation sometimes suggests. Several MBTI dimensions map substantially onto Big Five dimensions. E–I corresponds strongly to Extraversion, while S–N relates substantially to Openness. Some aspects of T–F overlap with Agreeableness, and J–P has relationships with Conscientiousness. The mapping isn't perfect, however.
There is also something about MBTI that I suspect explains why you found it useful: its axes are intuitively interpretable as cognitive or behavioral contrasts. "Do I tend to rely on concrete information or abstractions?" and "Do I tend to make decisions through impersonal analysis or interpersonal considerations?" can sometimes be more interesting questions than simply learning that someone scores 73 on Agreeableness.
Where your idea gets especially interesting
You don't have to stop at four dimensions—or five.
Suppose we retained dozens or hundreds of psychologically meaningful dimensions and represented an individual as
x=(x1,x2,x3,…,xn).\mathbf{x}=(x_1,x_2,x_3,\ldots,x_n).
Now personality becomes a point in a high-dimensional semantic space. We can calculate distances between individuals, identify clusters, examine trajectories through time, and ask whether particular regions predict behavior.
That gets remarkably close to Robbins's argument in the essay you just sent me. Modern computational methods—and especially language models—remove much of the historical necessity for reducing personality to four or five numbers.
I
nstead of asking:
"Which of 16 kinds of person is Jack?"
or even:
"Where is Jack on five traits?"
we could ask:
"Where does this individual lie in a 200-dimensional behavioral/linguistic space, and which dimensions actually predict behavior?"
That strikes me as a much richer research program.
There's also a wonderful irony here. MBTI's biggest conceptual mistake may not have been its four dimensions at all. It may simply have been throwing away most of the information in those dimensions by converting continuous scores into letters.
And that is an economist's complaint if I've ever heard one: why take continuous data and voluntarily turn it into a dummy variable?
Thank you for this fascinating history of the Big Five and the contingencies of how that came to be canonized as *the* tool we have to understand and codify human personality and how 1940s era statistical tools erased relational traits, giving us a framework that can only provide a stunted and incomplete understanding of what human personality actually is. I read this right after the “What a Time to be an Agent” piece and found the argument really thought provoking, although I’m not versed enough in linguistics to squeeze all the juice out of this lemon — my own field is history. As I read your work, it proposes that the tools and philosophies of literature inform (and perhaps offer the only way forward) to creating stable personalities in LLMs because of how it can actually deal with the human relational connections that undergrid personality that current ML training practices will inherently average away in the training corpus. It’s such an ambitious project, creating and designing something like a stable personality that persists through time and through the challenge of constant interaction with users. In my own usage of these tools (Cowork, not Code, lol) I’ve found that constraining the tool through iterative and evolving global instructions to map out a preferred relational texture is giving me a model with a stable-enough personality for the kind of professional collaboration I want it to do. Of course, it’s user directed — which for me is a feature, not a bug— and certainly not a “caring personality” ready to deal with our elders, or anything like that. I just posted about this relational constraint today: [https://mivonnewf.substack.com/p/the-dialogic-mirror-cognitive-fitness?r=3p2zgl]. Thanks for so much to think about.
Also, the Borges load in the corpus must be a real thing — it pops up a lot in my own convos with Claude, although since I am a Latin Americanist, I attributed it in part to Claude refracting my own interests and influences back to me, marginalia style.