Why you need to know this: you are being observed, categorised and described, at work and outside it, by people and by systems, and almost none of those descriptions and perceptions will ever be shared with you.
Somebody who knows you well has a perception of you and an opinion about you. So does a colleague, an old friend, a former partner, a recruiter’s software, your employer’s monitoring tools, and whatever your phone and social media platforms have inferred. Those descriptions do not agree; they will not carry equal weight, and only some of them can be questioned.
Start with the surprising part. Pooling 24 studies and 7,022 people, descriptions of narcissism, Machiavellianism and psychopathy given by somebody who knew the person agreed with that person’s own account at a level the personality profiling field treats as substantial, and higher than the same comparison usually runs for ordinary personality traits (Rico-Bordera, Pineda, Galán & Piqueras, 2025). The traits carrying the heaviest social penalty are not the hidden ones. They’re the ones that get noticed, discussed, and drive bias formation.
Then the part that changes who you should ask. Across 263 samples and 44,178 people, seeing more of somebody improved a rater’s accuracy only modestly, while genuine closeness was needed for substantial gains in understanding (Connelly & Ones, 2010). Family and close friends produced the most accurate descriptions, and colleagues produced weaker ones, which is awkward, because colleagues are the group every employer asks, and the ones most closely associated with professional and financial success.
That’s the first problem with how this personality assessment science is used. It was built almost entirely on partners, parents, friends, roommates, and classmates. Employers then borrowed it and applied it to the one relationship identified as the weakest. Only two of the twenty-four studies above used work.
The second problem is the report itself. You take a test, you get a number, and somebody explains what it means. That format leaves out what makes a trait score interpretable: the situational environment. Narcissism and leadership are the standard case: the association rises to a point and then falls, so the same score means different things at different levels, and no single-score report can tell you which one you’re looking at.
Now the newer observer. A model reading Facebook likes matched self-reported personality more closely than the person’s own Facebook friends did, across 86,220 volunteers (Youyou, Kosinski & Stillwell, 2015). Pooling 21 studies of smartphone data, only extraversion showed a moderate association; the rest were weaker, and the authors concluded that use for accurate individual assessment is currently limited (Marengo, Elhai & Montag, 2023). A recent study built models of all three dark traits from Facebook posts and reported prediction error rather than accuracy, which did not stop the coverage announcing that artificial intelligence can identify a psychopath (Leberecht, Nedderhoff, Zitzmann & Hecht, 2026).
Three things follow. Machines have less access to your circumstances than any colleague does, not more. Regulators have already drawn lines: inferring a worker’s emotions is prohibited in the European Union, and in February 2024 the Information Commissioner’s Office stopped an employer scanning the faces of more than two thousand staff to record attendance. Whatever is held about you in the United Kingdom is accessible through a subject access request.