To understand psychology, it is necessary to understand how academia works. Research at universities is organized around paradigms (Kuhn, 1962). A paradigm is a set of shared ideas, assumptions, methods, and practices—something like a scientific culture.
The overarching paradigm in academic psychology is that scientific claims should be supported by empirical evidence, often obtained through quantitative analysis. In some areas of psychology, data are relatively inexpensive to collect, making it easy to conduct studies and produce results.
The harder part is getting the results published in a peer-reviewed journal. This is where belonging to an established research paradigm becomes advantageous. Researchers working within a paradigm share assumptions about which questions are important, which methods are appropriate, and which findings constitute progress. They also serve as reviewers and editors for journals in their field.
Consequently, research that conforms to an established paradigm has a natural audience and an established route to publication. Research that challenges its fundamental assumptions faces greater obstacles. Reviewers may be less receptive to findings that undermine the theoretical foundations of their own research programs.
These dynamics can allow paradigms to survive long after their scientific usefulness has diminished. In psychology, several competing research programs may coexist for decades, investigating similar phenomena with different theoretical interpretations while paying remarkably little attention to contradictory findings.
The resulting publications generate citations, support academic careers, and demonstrate research productivity. However, the production of publications is not the same as the production of knowledge. A research program can remain productive in terms of publications without making much progress toward answering its central scientific questions.
Kuhn (1962) described how established paradigms resist challenges and how scientific revolutions can involve generational changes in the scientific community. However, the conditions under which science operates are changing.
The internet has made it possible to distribute scientific findings and criticism without relying exclusively on peer-reviewed journals. Artificial intelligence has further accelerated access to information, allowing researchers to examine publications, compare competing claims, and identify inconsistencies more efficiently.
In my scientific utopia, weak paradigms would lose their influence more quickly because their shortcomings could be exposed and discussed almost immediately after publication. Researchers would have less opportunity to sustain an unproductive research program by repeatedly publishing variations on the same empirical finding.
This brings me to the subject of this blog post: the General Factor of Personality (GFP) paradigm.
A phenomenon discovered over 100 years ago
The General Factor of Personality is a relatively recent name for a phenomenon recognized more than a century ago.
Thorndike (1920) observed that when people evaluate other individuals, their ratings of different desirable characteristics tend to correlate more strongly than expected. People judged favorably on one characteristic also tend to receive favorable ratings on others.
For example, individuals rated as intelligent may also be rated as attractive, conscientious, and likable. Conversely, individuals who receive unfavorable ratings on one characteristic tend to receive unfavorable ratings on others.
Thorndike interpreted these correlations as evidence of a halo effect: a general impression of a person influences judgments of specific attributes.
The phenomenon is robust and highly replicable. Nobody seriously needs another demonstration that favorable personality ratings tend to correlate.
The scientific question is what produces these correlations.
Two competing explanations
The first explanation is evaluative bias.
When we like someone, we tend to exaggerate their positive characteristics and minimize their negative ones. When we dislike someone, we tend to do the opposite.
This produces correlations between ratings of attributes that may be only weakly related—or even unrelated—in reality.
For example, cognitive ability, physical attractiveness, and happiness are distinct characteristics. Yet people’s subjective evaluations of these attributes may be positively correlated partly because they share a general tendency toward favorable or unfavorable self-evaluation.
The second explanation is that desirable attributes genuinely cluster together in individuals.
According to proponents of the GFP, some people tend to possess more desirable characteristics across multiple personality domains. This tendency may reflect a broad psychological disposition, sometimes interpreted as social effectiveness. Some researchers have also proposed evolutionary or genetic explanations for the GFP.
These hypotheses are fundamentally different.
The halo hypothesis explains correlations among ratings.
The substantive GFP hypothesis proposes a common source of variation in people’s actual characteristics.
To illustrate the distinction, consider intelligence and attractiveness. If the two attributes were independent, approximately half of the people above average in intelligence would also be above average in attractiveness, simply by chance.
If the correlation were (r = .30), approximately 60% of those above the median on one attribute would also be above the median on the other, assuming normally distributed attributes.
Thus, even a moderate positive correlation would leave considerable room for people who are highly intelligent but unattractive, or attractive but less intelligent.
The existence of these combinations is not the central issue. The question is whether there is a genuine common factor that causes desirable attributes to covary.
What have we learned in 100 years?
After a century of research, one might reasonably expect scientists to have made substantial progress in distinguishing these explanations.
Fortunately, there are ways to do so.
One approach is to compare subjective ratings with more objective indicators. If correlations among favorable ratings are substantially stronger than correlations among independently measured attributes, this suggests that evaluative bias contributes to the observed covariance.
Another, particularly informative approach is to compare ratings obtained from different informants.
If a substantive personality trait causes people to possess multiple desirable characteristics, the common factor should produce convergent evidence across independent observers.
If the factor primarily reflects evaluative halo, much of its variance should be specific to individual raters.
Biesanz and West (2004) found that Big Five ratings were correlated when obtained from the same informant, whereas the structure was much closer to orthogonal when different informants were used.
Anusic et al. (2009) extended this approach using multitrait–multimethod models. The results showed that the general evaluative halo was largely specific to individual raters, whereas broader substantive personality factors demonstrated greater cross-rater convergence.
Moreover, the halo factor was strongly related to independent measures of evaluative biases in self-perception.
These findings do not imply that all correlations among personality traits are artifacts. Nor do they rule out genuine higher-order personality factors.
However, they provide direct evidence that much of the general evaluative covariance in personality ratings reflects the way people evaluate themselves and others rather than a single underlying personality disposition.
This is precisely the kind of evidence that should inform attempts to interpret the GFP as a substantive trait.
One might think that such findings would have shifted the research agenda away from repeatedly demonstrating the existence of a general factor toward investigating its remaining substantive variance and incremental validity.
But the GFP paradigm continues to produce studies demonstrating that the factor can be extracted from different datasets.
A new study using artificial intelligence
In 2025, Dimitri van der Linden and colleagues published a paper titled The General Factor of Personality (GFP) in Natural Language: A Deep Learning Approach in the Journal of Research in Personality.
The authors used a large language model to examine associations among personality-descriptive adjectives in natural language.
They found that the first principal component distinguished desirable from undesirable personality attributes. Moreover, loadings on this component correlated highly ((r = .86)) with loadings from an earlier study based on human personality ratings.
The authors interpreted this finding as further evidence for the generality of the GFP.
However, why should this finding be surprising?
It has long been recognized that personality adjectives have both descriptive and evaluative meanings.
The descriptive, or denotative, meaning concerns what a word says about a person’s characteristics or behavior. The evaluative, or connotative, meaning concerns whether the attribute is considered good or bad.
Consider the difference between describing someone as relaxed and describing the same person as lazy.
Both terms may refer to a tendency toward low activity, but they convey very different evaluations.
Similarly, intelligent and kind describe different attributes, but both have positive evaluative meanings.
A language model trained on human language can learn these evaluative associations. Finding a general evaluative dimension in its representation of personality adjectives therefore tells us something about how personality descriptions are organized in language.
It does not establish that desirable attributes genuinely cluster together in individuals because of a common personality trait.
Indeed, the emergence of a general evaluative factor in linguistic representations is entirely compatible with the halo hypothesis.
The study reports that its first component accounted for 22.7% of the variance in the language-model data, compared with 7.11% in the original human-rating data. Although these percentages are not directly comparable because the datasets have different structures, the findings clearly demonstrate the prominence of evaluation in linguistic representations.
But the paper offers no independent behavioral evidence that this factor reflects social effectiveness or another substantive personality disposition.
What did the new study teach us?
The authors’ discussion acknowledges the unresolved question of what the general factor represents.
They consider the possibility that it reflects methodological artifacts or evaluative biases, as well as the alternative that it reflects a substantive psychological phenomenon.
That is essentially where the debate stood before the study.
The paper does acknowledge halo bias and cites Anusic et al. (2009). However, it does not provide a new empirical test capable of distinguishing the bias explanation from the substantive GFP explanation.
This is the central problem.
The study replicates the phenomenon that needs explaining but contributes little toward explaining it.
After more than 100 years of research on evaluative correlations in personality judgments, another demonstration of a general evaluative factor is not a major theoretical advance.
The more interesting questions concern how much of the covariance is genuine, how much is produced by evaluative bias, and whether any remaining substantive factor has meaningful explanatory or predictive value beyond specific personality traits.
These are questions that require appropriate measurement models, independent informants, and external validation.
Extracting a first principal component from another dataset does not answer them.
When does a research paradigm stop producing knowledge?
The GFP illustrates a broader problem in academic psychology.
A research program can continue to generate publications because its central empirical phenomenon is easy to reproduce.
The repeated appearance of the same finding creates an impression of cumulative scientific progress. Yet the underlying theoretical question may remain unanswered.
The GFP paradigm has repeatedly demonstrated that favorable personality ratings tend to correlate. The phenomenon is real. The disagreement concerns its interpretation.
The evidence from multitrait–multimethod studies indicates that a substantial part of this general evaluative covariance is attributable to rater-specific influences.
This creates a serious challenge for theories that interpret the first general factor in personality ratings as a broad substantive disposition.
It is still possible that some residual covariance reflects genuine common causes. However, proponents need to demonstrate that such a factor can be measured independently of evaluative biases, identify the psychological processes that produce it, and establish its incremental validity.
Merely extracting another general factor does not satisfy these requirements.
Even if a small substantive GFP exists, it would not follow that it is more useful than the specific attributes it is supposed to explain.
A person’s intelligence, honesty, conscientiousness, sociability, and emotional stability can have important consequences. A weak general tendency toward desirable characteristics would not eliminate the need to measure these attributes separately.
The purpose of personality science should be to explain individual differences, not merely to identify increasingly broad statistical summaries of personality ratings.
The van der Linden et al. (2025) study demonstrates that a general evaluative factor can be recovered from natural language using artificial intelligence.
That is a methodological replication of a familiar result.
What it does not demonstrate is that the General Factor of Personality is a meaningful psychological trait.
A more important application of research on evaluative bias concerns the growing use of artificial intelligence to evaluate people (Kim et al., 2025). Recent studies show that AI models reproduce human halo effects. For example, models infer intelligence, competence, and trustworthiness from facial appearance, even though these judgments lack adequate validity. These biases can also influence simulated hiring decisions (Kim et al., 2025; Lehr et al., 2026). Thus, while researchers continue to debate whether the General Factor of Personality reflects a substantive disposition, artificial intelligence is already reproducing the evaluative biases that contribute to this factor.
Conclusion
Rather than demonstrating once again that desirable attributes are associated in language, personality researchers could make a more useful contribution by investigating how these associations create biased evaluations of people and how such biases can be prevented.
References
Anusic, I., Schimmack, U., Pinkus, R. T., & Lockwood, P. (2009). The nature and structure of correlations among Big Five ratings: The halo–alpha–beta model. Journal of Personality and Social Psychology, 97, 1142–1156. https://doi.org/10.1037/a0017159
Biesanz, J. C., & West, S. G. (2004). Towards understanding assessments of the Big Five: Multitrait–multimethod analyses of convergent and discriminant validity across measurement occasion and type of observer. Journal of Personality, 72, 845–876. https://doi.org/10.1111/j.0022-3506.2004.00282.x
DeYoung, C. G. (2006). Higher-order factors of the Big Five in a multi-informant sample. Journal of Personality and Social Psychology, 91, 1138–1151. https://doi.org/10.1037/0022-3514.91.6.1138
Kim, K., Ryu, J., Jeon, H., & Suh, B. (2025). Blinded by context: Unveiling the halo effect of MLLM in AI hiring. Findings of the Association for Computational Linguistics: ACL 2025, 26067–26113. https://doi.org/10.18653/v1/2025.findings-acl.1338
Kuhn, T. S. (1962). The structure of scientific revolutions. University of Chicago Press.
Thorndike, E. L. (1920). A constant error in psychological ratings. Journal of Applied Psychology, 4, 25–29. https://doi.org/10.1037/h0071663
van der Linden, D., Cutler, A., van der Linden, P. A., & Dunkel, C. S. (2025). The general factor of personality (GFP) in natural language: A deep learning approach. Journal of Research in Personality, 117, 104635. https://doi.org/10.1016/j.jrp.2025.104635