Inglourious Signs: HUMINT for Detecting AI-Written Texts on LinkedIn
Signs, typographic frictions, expressive calques, and other lightweight markers that often say more than a long semantic assessment.
“Am I chatting with AI?” Have you ever had the feeling that you were speaking more with patterns than with people? On LinkedIn and beyond, that question stopped being naïve a while ago: not because using AI should be treated as a fault, but because a broader grey zone has opened up in which it is no longer immediate to understand where the author’s hand ends and the interface’s direction begins.
I am not interested here in moralising about “who uses AI”, nor in playing the petty inquisitor of automated detectors. I am interested in a more serious question: how do you detect, at a reasonable cognitive cost, that a text is heavily governed by a generative pipeline or by LLM-shaped post-editing? The most obvious route runs through semantics, syntax, discursive patterns, constructions, relaunches, and symmetries. It is a useful route, at times an excellent one. It also has a cost: time, attention, experience, context, and comparative memory. There is, however, a lighter and often faster layer to activate: the layer of signs.
That is why I am taking you through a transdisciplinary path across social media, cinema, HUMINT, and AI. AI would perhaps accuse this piece of not displaying a single numerical datum capable of instantly increasing its technical credibility without damaging its strategic understatement. I will live with that.
This is the level at which what I call Inglourious Signs begins to emerge, with cinephile irony but serious intent. And yes: Inglourious, not Inglorious. The citation is deliberate, because the Tarantino reference is not a flourish; it is the architecture of the argument.
1. Semantics remains decisive, yet often enters too late
Semantic analysis remains a formidable resource for anyone who truly knows how to read. It allows one to catch inconsistencies, inferential fragilities, overly well-behaved postures, simulated depth, polished moralism, and passages that flow with a suspicious degree of composure. All of that matters, and it matters greatly. Yet, especially within the social-media stream, this kind of reading demands a lot: one must pause, remember, compare, keep the present text together with the author’s past writing, and distinguish the author’s real breathing pattern from the one a machine has laid over it.
For that reason, in practice, it often makes sense to reverse the sequence. First, you look at whether the text carries visible traces of its own pipeline; only afterwards, if needed, do you enter the full merit of what it is saying. This is not a surrender of semantics; it is, on the contrary, a better administration of it. In a saturated, noisy, accelerated environment like LinkedIn, starting immediately with the deepest reading often means spending too much too early.
2. Tarantino offers two detectives, and therefore two economies of detection
In Inglourious Basterds, two different ways of intercepting incongruity coexist. One works through depth, accumulation of signals, culture, context, and patience. The other works through the extreme pertinence of the marker, through speed, through the choice of the minimum sign which, precisely because it is minimal, cuts deeper than a long reading. Landa and Hellstrom, in this sense, are not just characters: they are two cognitive models.
2.1 Hans Landa: high-intensity cognitive detection
Hans Landa embodies the detective of high cognitive intensity: a polyglot, razor-sharp, able to make minimal differences of register work together with psychological details, context, posture, and language. His strength lies precisely there: he does not hunt the single error, nor does he live off the isolated slip; he holds many levels together and, by holding them together, reads a system. It is powerful detection because it is culturally costly; it rests on a broad base, on a sophisticated armamentarium, and on a rare capacity for adaptation.
In the opening farmhouse scene in France, his method is unmistakable: he does not merely understand words; he weighs their weight, their temperature, their function, and the moment at which they emerge. Landa does not use language as a simple vehicle. He uses it as a diagnostic weapon. In contemporary terms, it is compute-intensive detection: it requires time, memory, elasticity, a comparative base, and, above all, a quantity of context that we do not always have available in real and digital life.
Brought onto the terrain of digital texts, the Landa model corresponds to the reader who knows how to recognise suspicious argumentative scaffolding, simulated depth, well-tempered moralism, overly smooth fluency, overly happy symmetries, and caution that is too perfectly dosed. Such a reader is precious. Yet it is not the most economical reader to activate first.
2.2 Dieter Hellstrom: the low-compute marker
In the tavern scene, Hellstrom moves from a sophisticated reading of accents and provenances to something even more effective: a gesture, a minimal detail, the number three expressed with the fingers in the “wrong” way for that context. This is the useful point for us: the most learned, deepest, most burdensome level does not always win. Sometimes the winning move is the right sign, in the right context, read at the right moment.
Hellstrom does not give up finesse. He condenses it. He shifts detection from a costly architecture of attribution to a silent, nearly aphonic indicator that carries an extraordinary disproportion between the simplicity of the sign and the diagnostic force of the result. That disproportion is precisely what makes his gesture such a useful metaphor: the marker does not explain everything, but it betrays enough to change the whole picture.
2.3 Why the parallel also holds on LinkedIn
Brought onto the terrain of digital texts, that difference is precious. There is a “Landa” level, made of semantic reading, knowledge of registers, and sensitivity to real or simulated thinking. There is also a “Hellstrom” level, made of mute, small, nearly aphonic markers that do not explain everything, yet betray a great deal. The question is not which school to choose as if they were rivals; the question is which one to activate first, at what cost, and for what purpose.
On LinkedIn, the lightweight marker has an immediate advantage: it lets you raise your threshold of attention without investing in a full reading from the start. First, I look at whether the text carries traces of its own pipeline; only afterwards do I decide whether to enter the more learned assessment.
3. The sign, then, costs less than a full assessment
The signs that interest me are not the coarse clichés that are often used to “unmask AI”: bullet points, emojis, a polite tone, clarity, order. All of those are too widespread, too easily human, too easily imitable. What interests me are less banal and more pertinent signs: typographic frictions, graphic automatisms, recurring rhythms, hinge-formulas, micro-expressive calques, and ways of relaunching and correcting that, once taken together, begin to tell a story of provenance.
Take the em dash. In ordinary Italian usage, it is not as comfortable or spontaneous a sign as the comma, the colon, parentheses, or the short dash. On a standard Italian QWERTY keyboard it is not available as a direct key; on desktop it can be worked around through procedures known to those who know them, but not immediate ones, and that is precisely why its repeated appearance starts to matter. A single occurrence means little; a few isolated repetitions may still mean little; but once the sign recurs with serial ease in Italian prose that does not have an evident typographic tradition behind it, the reader already has a reasonable clue. The sign is not impossible; it is simply not as natural as its frequency would suggest. On mobile touchscreens, the matter varies from environment to environment: shortcuts, automatic substitutions, and smart punctuation exist. Here too, the isolated case matters less; what matters are repetition and ease.
The same applies to overly uniform quotation marks, to always-neatly-groomed parenthetical insertions, and to serial recourse to punctuation that seems to arise more from interface or editorial normalisation than from a living gesture of writing. None of those elements, taken alone, constitutes proof. Yet once assembled as a constellation, the landscape already changes.
4. Epanorthosis, bridge-questions, relaunches: the figure is not guilty, seriality is
At this point, one dangerous shortcut should be avoided. Rhetorical figures are not signs of guilt. Epanorthosis, self-questions, relaunches, suspensions, corrective returns, reprises, double “no” constructions, “What do I mean?”, “So?”, “Why? Because…” all belong fully to the human toolbox and, when there is a recognisable voice behind them, they can be signature, rhythm, gesture. It is worth saying this clearly because it is a decisive point: the figure, taken in itself, proves nothing.
And if, while reading these lines, someone were to detect a slight self-reflexive fold, they would have caught a non-secondary point: when a figure can show itself while being described, we are already in a zone of ironic metalepsis, that is, in a fully human gesture that is difficult to reduce to a template.
So where does the problem begin? It begins when the gesture stops being a gesture and becomes a template. If a text relaunches itself three, four, five times through the same corrective mechanism, if each passage seems modelled on a cadence such as “it is not only… it is…”, if every paragraph finds itself a bridge-question, if the connections are all too smooth, then we are no longer dealing with the individual rhetorical figure: we are dealing with a pipeline regularity.
A simple example. A human author may very well write: “It is not a detail. It is an imprint.” Or: “It is not only style; there is a provenance embedded in it.” Fine. But when a text builds an entire progression through parallel relaunches, all in the same shape, all polished in the same way, the eye begins to sense the serial production of an effect. That is when attention rises.
The same applies to some hinge-formulas that models use with great comfort and that in Italian are now beginning to be heard too often: “the point is not…”, “what emerges here…”, “this brings us to…”, “ultimately…”, “at the level of…”, “in terms of…”. All legitimate, all possible, none incriminating in itself. They become eloquent when their distribution feels more automatic than necessary.
5. Very often, the real marker is not the lexicon: it is the breathing pattern
Many look for telltale words, suspicious phrases, signature syntagms. The temptation is understandable, yet that is rarely where the game is really decided. More often, the revelation arrives from the overall breathing pattern of the text.
What does one see, then? One sees a tendency toward the optimised sentence: neither too short nor too long, always well-composed, always fully functional. One sees a quasi-regular cadence in the number of lines, a certain homogeneity in the weight of paragraphs, prose that avoids not so much error as roughness. In other words, one sees writing that wants to be readable, linear, well-turned, and that, precisely for that reason, risks losing the small human deviation that breaks the polish.
A human text can, of course, be extremely clean. Yet when it is truly alive, it tends to carry some slight asymmetry: a sentence that runs longer than expected because the thought insists there; a parenthesis less disciplined but more true; a semantic twist not fully optimised for readability and precisely because of that more personal. Today’s models are very good at avoiding discomfort. The human writer, unless training themselves to look like a machine, tends to keep some useful discomfort.
6. Differences between LLM families exist, yet they should not be mythologised
The risk here is twofold: either to trivialise everything or to over-attribute. Anyone claiming to recognise with certainty which model wrote a text simply by reading that text, as though it were ballistic evidence, is usually overstating the case. Imprints exist, but they are not pure imprints.
In my experience, and with the necessary caution because versions, prompts, intermediate tools, and levels of post-editing can alter the picture considerably, some tendencies do emerge. GPT models often produce prose that is highly connected, persuasive, and capable of distributing the weight of passages while maintaining an almost professional fluidity. Claude often tends toward a more composed, reflective, disciplined argumentative order, with a pace that can feel elegant but also a little too domesticated. Gemini, in many cases, leans toward greater explicitness, greater didactic clarity, at times almost like an excellent educational populariser. Open-weight models and hybrid pipelines, by contrast, fluctuate more: at times rougher, at times more imitative, at times strongly dependent on who is guiding and cleaning them.
Even here, though, the model alone is never enough. When a text does betray something, it betrays a combination: model, interface, language, layout, workflow, intensity of human revision, publication platform, and authorial intent. Looking for the model name as though it were the sole culprit means losing the wider frame.
7. Disclosure and weaponization: the dilemma is real
At this point it becomes clear why turning this whole discussion into a fully explicit checklist would make little sense. Every effective criterion, once exposed all the way through, also teaches those who would like to neutralise it. This is the old problem of the weaponization of knowledge and, in terms closer to our own operational lexicon, its dual-use character: you raise the awareness of the good-faith reader while simultaneously increasing the evasive capacity of the less transparent writer, or simply of the more opportunistic one.
The useful threshold, then, does not coincide with total silence, nor does it coincide with the full recipe. It makes sense to show the principle, the families of signs, the limits, the false positives, even some examples; it also makes sense to hold back the excessively precise mechanics when those mechanics can immediately be trained against the detector.
8. Conclusion
The near future, in truth already partly present, will not neatly separate human texts from synthetic texts as though we were facing two pure blocks. It will separate governed texts from delegated texts. On one side there will be authors who use AI as a tool, a technical companion, an accelerator, or a sparring partner, while still retaining responsibility for form, thought, cut, and tone. On the other side there will be texts that are perfectly correct, perhaps even effective, yet still carry the signature of a delegation that has grown too wide, too lazy, too unaware.
That is the space in which Inglourious Signs become useful. Not as a moral bludgeon, not as a game of unmasking, not as a liturgy of suspicion, but as a threshold of attention. Is a glance enough? Sometimes yes, provided that the eye knows how to distinguish clue from verdict, convergence from obsession, and a pertinent sign from an interpretive tic.
If you have already recognised some of these signals, or if you have intuitively grasped other useful markers, feel free to write to me in DM: the interesting boundary here is not the act of unmasking in itself, but the frontier between governed use of AI and unconscious delegation. That frontier is also part of the work we do within MINDS Lab and our innovation advisory.
With all due respect to Hans Landa, the most learned and costly assessment in the world is not always required. Sometimes Hellstrom is enough. Sometimes the gesture is enough. Sometimes the right sign, read in the right context, is enough. Long live the Inglourious Signs.