What is the human signal?

The human signal is the trace of a mind inside a piece of work. It is what someone looks for in the first seconds, before taste arrives: a sign that a real person made decisions that mattered. Rahim Hirji named it and set three tests for it on 30 November 2025. This page keeps that argument separate from the measurements, because everything that has actually been measured concerns the signal going missing and how badly people detect it.

Three things a reader checks without being asked#

The argument starts from an observation about judgement rather than about technology. People search a photograph, a paragraph or a piece of music for evidence of intention before they decide whether they like it. In Box of Amazing on 30 November 2025, Hirji sets out what that search consists of: people "ask whether a real decision has been made", they "look for signs that something difficult has been carried with care, whether that difficulty is grief, doubt, time or reputation", and they listen "for the risk of a personal point of view".

Three tests, then. A decision that could have gone the other way. A cost the maker was willing to carry. A view somebody can be held to. Where all three are present the work reads as human. Where none is, it reads as what the essay calls frictionless sameness, however technically strong it looks.

The three are a claim about what people attend to. No study on this estate operationalises them, and nobody has scored a group of readers on whether they can find a mind in a text. That absence matters more than it first appears, and the rest of this page is about why.

The measured finding is convergence, which is narrower#

Two peer-reviewed experiments come close to the argument without testing it. Doshi and Hauser ran 293 writers producing short fiction with and without AI assistance, judged by 600 evaluators. The assisted stories were rated more creative, better written and more enjoyable, with the largest gains going to the writers rated least creative on their own. They were also markedly more similar to one another. Individual quality and collective range moved in opposite directions. Graded entry.

Hohenstein and colleagues get closer to the mechanism. In two preregistered randomised experiments on live text chat, greater use of algorithmic reply suggestions by one partner led the other person to write with more positive sentiment (b=0.178, p=0.045), and the effect held when the suggested messages were taken out of the sentiment score altogether (b=0.208, p=0.031). It appeared in sentences the person composed themselves. Having suggestions available without using them moved nothing (p=0.1801), which locates the effect in the act of adopting somebody else's phrasing. Graded entry.

Both results describe tone becoming more alike. Neither asks whether a mind was present. The estate holds a good deal of evidence for homogenisation and none at all for detection of intention, and those are different claims that the phrase "you can tell" tends to run together. The fuller treatment sits at does AI make everyone think alike.

People are confident detectors and poor ones#

The awkward result is in the same Hohenstein paper. Participants who actually used the reply suggestions were rated by their partner as more cooperative (b=15.66, p=0.018) and more affiliative (b=21.79, p=0.007). Participants merely suspected of using them were rated less cooperative and less affiliative, both at p<0.0001, after controlling for actual use. And suspicion tracked reality weakly: the correlation between suspected and actual use was 0.22.

So the penalty in that experiment attached to being suspected, and suspicion was close to uninformed. Software detectors perform no better on the question people most want answered: they misclassified 61 per cent of essays by non-native English speakers as machine-written while scoring near-perfectly on US eighth-grade essays. That evidence is set out at does AI detection work.

This does not dispose of the human signal. It does mean the signal cannot be treated as a property a reader reliably measures. A person's confidence that they can feel whether a mind was present is not evidence that they can, and acting on that confidence has a measurable cost for the people wrongly suspected.

The one part with a number behind it is about the maker#

Joshi and Vogel gave participants the same short-story task across conditions running from a three-word prompt to writing unaided. Psychological ownership rose steadily with how much of themselves went in, from a mean of 1.80 to 6.29, and no assisted condition reached the ownership of writing alone. The gain plateaued once the prompt reached roughly the length of the target text. Graded entry.

That is a measurement of what the author feels, not of what the audience detects. It supports the human signal as a description of how work gets made and leaves the perception half of the claim untested. On the reading this estate takes, that is the useful half anyway: the three tests are better used on your own drafts than on somebody else's.

What is still missing from the case#

Nobody has run the obvious study. Give readers matched pieces, one made with real decisions and one assembled, and score them. Until someone does, the claim that people recognise a mind stands on introspection and examples.

The two convergence studies are also narrower than the argument they are being asked to support. Doshi and Hauser is one short creative task with one form of assistance and the authors say so. Hohenstein studies a smart-reply suggester offering short canned options, between strangers, over a few minutes, and measures style as sentiment, which is one dimension of a voice and not the range of what a person might have said. Neither involves a model writing paragraphs of professional work.

And the essay's sharpest prediction has no test at all. It holds that models will learn to produce a convincing echo of the signal, small hesitations included, until genuine and imitated cannot be separated. Nothing on this estate measures that. It is carried here as an argument with a date on it and not as a finding.

Keeping the signal in your own work#

Three moves follow from the evidence above. Put the decision in before the draft, since the ownership result tracks how much of the shape came from you and the availability null suggests the loss happens at the moment you adopt someone else's phrasing. Keep the difficult part difficult, because the material a reader responds to is the part that cost something. And say what you actually think, at your own risk, since that is the element no suggestion system supplies.

A fourth follows from the detection evidence and runs the other way. Do not accuse people. Suspicion is weakly correlated with use, it is punished in the person suspected, and detectors are unsafe on the very writers most likely to be wrongly flagged. Ask what somebody decided and why. That question works whether or not a machine was involved, and asks the same discipline as the source rule.

Key sources

• Hirji, R. (2025). The Human Signal. Box of Amazing, 30 November 2025. The dated first publication of the term and the three tests.
• Doshi, A. R. and Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28). Graded entry.
• Hohenstein, J., Kizilcec, R. F., DiFranzo, D. et al. (2023). Artificial intelligence in communication impacts language and social relationships. Scientific Reports, 13, 5487. Graded entry.
• Joshi, N. and Vogel, D. (2025). Writing with AI Lowers Psychological Ownership, but Longer Prompts Can Help. ACM Conversational User Interfaces 2025. Graded entry.

Related SuperSkills research#

The convergence evidence in full sits at does AI make everyone think alike, and the practical version at how do I keep my own voice when using AI. On whether anyone can tell, does AI detection work. On what the signal is a signal of, what stays human and what is judgement. The cost of never putting the decision in is capability debt.

About this research#

Written by Rahim Hirji, author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. "The human signal" is his term, first published in Box of Amazing on 30 November 2025, which is the dated publication the estate requires before crediting a coinage. The findings above are attributed to the researchers who produced them and kept separate from the interpretation.

How this research works  ·  Reviewed quarterly  ·  Found an error? Tell me and it is corrected on the page.