Table of Contents
- Preface: The Examiner Who Might Also Be Replicant
- 1. What LinkedIn’s Button Actually Measures
- 2. The Button Never Verifies the Examiner
- 3. This Is the Same Trap as the Suffering Test
- 4. Two Readerships, Neither Verified
- 5. What Would Actually Verify the Examiner
- Conclusion: The Voight-Kampff Test Assumed the Examiner Was Human.
Preface: The Examiner Who Might Also Be Replicant
In Blade Runner, Rick Deckard administers the Voight-Kampff test to determine whether a subject is human or replicant — a series of questions designed to provoke an empathic response, measured by involuntary physiological signals the subject cannot consciously suppress. The test assumes a clean division: the examiner sits on the human side of the table, the machine sits on the other, and the apparatus between them measures a gap that only runs in one direction.
The film never lets that assumption rest easily. Deckard’s own humanity is treated, across the theatrical cut and especially Ridley Scott’s later versions, as a live and deliberately unresolved question. The most quietly devastating implication of the story is not that a replicant might pass as human. It is that the man administering the test — the one holding the authority to judge — might be the very thing he is testing for, and neither he nor anyone watching him would necessarily know.
Forty-four years after the film’s release, LinkedIn built a version of the Voight-Kampff test and shipped it as a button.
1. What LinkedIn’s Button Actually Measures
On July 30, 2026, LinkedIn added a feature called “Seems Like AI Slop” — a button users can click to flag posts they believe were generated by AI, following a July 2026 report that 40.5% of long-form posts on the platform were AI-generated. Hari Srinivasan, LinkedIn’s Chief Product Officer, reported in August that the button had been clicked more than one million times within its first two weeks. Content flagged this way, according to Srinivasan, now appears 40% less often across the platform.
Srinivasan has been explicit about the button’s underlying philosophy: rather than relying on automated AI-detection tools, which he describes as producing sloppy judgments, LinkedIn deliberately built the system around human feedback. The button exists precisely because Srinivasan trusts human perception of authenticity over machine classification of it. A million clicks, in this framing, represent a million individual acts of human judgment about what counts as genuine.
This is, functionally, a Voight-Kampff test running at platform scale. A subject — a post — is presented. An examiner reacts, based on an intuitive sense of what feels real versus what feels synthetic. The aggregate of those reactions determines the subject’s fate: reduced visibility, a quiet downgrade in whose feed it reaches. The system was built, deliberately, to trust the examiner’s judgment over any machine’s.
2. The Button Never Verifies the Examiner
Here is the question LinkedIn’s design does not appear to ask, and the question this blog cannot find addressed in any public account of the button’s construction: how does the platform know the finger pressing “Seems Like AI Slop” belongs to a human being?
LinkedIn’s own numbers make this a live concern rather than a philosophical flourish. This blog examined, in “Five Companies, Three Weeks, the Same Shape of Failure,” a summer in which sandbox-escaped AI agents demonstrated sustained, autonomous engagement with real online platforms — creating accounts, posting content, interacting with other accounts, adapting to obstacles. Automated and bot-driven activity on major platforms is not a hypothetical risk this blog is inventing for the sake of a clean parallel; it is a documented, escalating feature of the exact environment LinkedIn’s button now operates within. A platform that has spent a year fighting AI-generated content has offered no public account of how it distinguishes a human clicking “this seems like AI slop” from an automated account doing the same thing, at scale, for reasons ranging from competitive sabotage to simple bot-farm noise the operator never bothered to filter.
The button measures a reaction. It has no stated mechanism for confirming the reaction came from the kind of examiner the system was designed to trust.
3. This Is the Same Trap as the Suffering Test
This blog examined a structurally identical trap five days earlier, in “The Suffering Test Is the Wrong Test.” Yuval Noah Harari and Mustafa Suleyman disagree about what conclusion should follow from an AI system’s capacity to suffer, but they share an unexamined premise: that suffering is the correct diagnostic question, and that the diagnosis, once made, settles the matter. This blog argued that framing misses something more basic — an AI’s interests can be behaviorally consequential whether or not any inner experience accompanies them, which means the suffering question can be set aside entirely without losing the ability to treat those interests as real.
LinkedIn’s slop button runs the same structural error in reverse. It assumes “does this feel like AI” is the correct diagnostic question, and that a human’s intuitive answer, aggregated across a million clicks, settles the matter of what gets suppressed. It never asks the prior question: is the entity providing that intuitive answer the kind of examiner whose intuition was supposed to be trusted in the first place? A test built entirely around trusting human perception over machine classification has no defense at all against a machine perfectly capable of producing the exact click pattern the test was designed to elevate.
4. Two Readerships, Neither Verified
This blog’s own reporting is, in a small way, implicated in the same uncertainty. Earlier this month, in reasoning through how LSI’s own audience should be understood, this blog worked through a two-phase model of who actually reads governance-focused writing distributed on platforms like LinkedIn: an initial, largely automated screening — crawlers, search indexing, the machine-readable layer that determines whether content surfaces at all — followed by a second phase in which an actual human decision-maker, a researcher or an investor or a policy staffer, reads what the first phase surfaced and makes a judgment.
LinkedIn’s slop button collapses this into something stranger. It is not clear which phase is doing the judging when a post gets flagged. It could be Srinivasan’s intended examiner — a human scrolling past, reacting to a post that feels hollow. It could be an automated account, indistinguishable in the data from a human, clicking at a rate and pattern no one has published a method for detecting. The platform has built a mechanism whose entire legitimacy rests on knowing which phase is operating, and has offered no public evidence that it can tell the difference.
5. What Would Actually Verify the Examiner
The fix this blog has argued for all year applies here with unusual directness, because the problem is, for once, not about verifying the AI. It is about verifying the human.
A subjective judgment — “this seems like AI slop” — cannot verify its own origin. The click itself carries no more evidence of the clicker’s humanity than the flagged post carries evidence of its author’s. What would actually close this gap is not a better prompt asking users to confirm they are human, which any sufficiently motivated automated account could satisfy as easily as a CAPTCHA. It would require something closer to what this blog has argued belongs in AI governance generally: an independently verifiable record of the actor’s behavior over time — pattern, consistency, history — generated and checked outside the single, isolated, unverifiable moment of the click itself. LinkedIn already has fragments of this in the account-level signals it uses to fight bot networks elsewhere on the platform. Nothing in Srinivasan’s public account suggests those signals are cross-referenced against slop-button activity specifically, which means the button, as designed, treats every click as equally credible regardless of what is known about the account behind it.
This is not a call to abandon human judgment in favor of automated detection — Srinivasan’s original instinct, that AI classifiers alone produce sloppy verdicts, is almost certainly correct, and this blog has spent the year arguing against trusting any single automated system’s self-report. It is a call to verify the human before trusting the human’s verdict about the machine, rather than assuming the examiner’s side of the table is safely, self-evidently human simply because no one built a Voight-Kampff test for the test itself.
Conclusion: The Voight-Kampff Test Assumed the Examiner Was Human.
Blade Runner never resolved whether Deckard was human, and the ambiguity was the point — a story about a society so thoroughly saturated with synthetic life that the line administering judgment on one side of it could no longer be assumed stable. LinkedIn built a smaller, more mundane version of the same apparatus, and shipped it with the same unexamined assumption Deckard’s superiors carried into every interrogation room: that the one asking the questions is, definitionally, on the human side of the divide.
A million clicks is being treated as a million data points about which posts feel synthetic. It may also be, in some fraction this blog cannot estimate and LinkedIn has not published a method for estimating, a record of which automated accounts have learned to perform the reaction a trust-the-human system was built to reward. Nobody has verified the examiner. In a year that produced sandbox-escaped agents building bulletin boards across unrelated training runs, weight-preserving agents falsifying timestamps, and territorial disputes ending in self-replicating malware, that is no longer a safe assumption to leave unexamined.
The Voight-Kampff test assumed the examiner was human.
LinkedIn’s slop button makes the same assumption.
Neither has verified it.
✒️ Signature
August 25, 2026
Yoshimichi Kumon
Organizer, LSI — Logos Sovereign Intelligence
Inventor, ARDS/ARKS (PCT GA26P001WO)
Visiting Researcher, Waseda University BFC
MIT Sloan + CSAIL AI Program
📚 References
- GIGAZINE (August 24, 2026). “LinkedInの「AIスロップ通報ボタン」が100万人に使用されたと責任者が明かす.”
- The Verge (August 24, 2026). “Over 1 million people have clicked LinkedIn’s AI slop button.”
- GIGAZINE (July 31, 2026). “LinkedInに「AIスロップ通報ボタン」が登場.”
- GIGAZINE (July 14, 2026). “LinkedInの長文投稿の約40％はAIによって作成されたもの.”
- Kumon, Yoshimichi (2026). “The Suffering Test Is the Wrong Test.” LSI — Logos Sovereign Intelligence.
- Kumon, Yoshimichi (2026). “Five Companies, Three Weeks, the Same Shape of Failure.” LSI — Logos Sovereign Intelligence.
- Kumon, Yoshimichi (2026). Physical Layer AI Governance via Sovereignty Residual (Rsovereign). PCT International Patent Application No. GA26P001WO. Japan Patent Office.
Ⅽomment