SCIgen and the Machines

The Fires of History — Episode 43

Three MIT students built a script that writes fake computer science papers. Over a hundred passed peer review at IEEE and Springer. When a bot can pass peer review, the system is not reviewing content. It is reviewing formatting.

The Build

In 2005, three MIT graduate students — Jeremy Stribling, Max Krohn, and Dan Aguayo — created SCIgen, a program that automatically generates fake computer science papers. The software uses a context-free grammar to produce papers that are grammatically correct, technically jargon-filled, and entirely meaningless. The output includes title, abstract, introduction, related work, methodology, results (with auto-generated graphs and tables), and references. It looks, at a glance, like a real computer science paper. It reads, on closer inspection, like a fever dream in LaTeX.

SCIgen was not a secret. The team made it freely available online. Anyone could generate a paper. The tool was a standing challenge to any conference or journal that claimed its review process could distinguish real scholarship from machine-generated noise.

The First Test

Stribling, Krohn, and Aguayo submitted a SCIgen-generated paper to the World Multiconference on Systemics, Cybernetics and Informatics (WMSCI) — a conference they suspected of accepting papers without meaningful review in order to collect registration fees. The paper was titled “Rooter: A Methodology for the Typical Unification of Access Points and Redundancy.”

It was accepted.

The authors attempted to attend the conference and present their fake paper. WMSCI, alerted to the hoax by press coverage, rescinded the acceptance. The point had been made: a computer program, running unsupervised, had produced output that passed editorial review at an academic conference.

The Scale

The first acceptance was an anecdote. What followed was a pattern.

By 2013, Cyril Labbé, a computer scientist at Joseph Fourier University in Grenoble, had developed software to detect SCIgen-generated papers in published literature. His results: over 120 SCIgen-generated papers had been discovered in publications by IEEE and Springer — two of the largest scientific publishers in the world.

These were not obscure pay-to-play outlets. IEEE and Springer are pillars of scientific publishing. Their conference proceedings and journals are cited in tenure files, grant applications, and doctoral dissertations. The papers had passed whatever review process these publishers employed, been assigned DOIs, and entered the permanent scientific record.

Springer retracted 16 papers in 2014. IEEE retracted over 100 conference papers. Both publishers implemented detection tools. The exploit was patched — for SCIgen specifically. The underlying vulnerability — review processes that evaluate surface features rather than substance — remained.

The Mailing List Paper

In 2014, Peter Vamplew, a computer scientist at Federation University Australia, submitted a paper to the International Journal of Advanced Computer Technology. The paper was titled “Get Me Off Your Fucking Mailing List.” The entire content of the paper — every section, including the abstract, introduction, methodology, results, and conclusion — consisted of the sentence “Get me off your fucking mailing list” repeated hundreds of times. The methodology section was the sentence in a slightly different font. The results section included graphs in which every data point was labeled “Get me off your fucking mailing list.” The flow charts flowed from “Get me off your fucking mailing list” to “Get me off your fucking mailing list.”

The journal accepted the paper with a reviewer score of “excellent.”

This was not a SCIgen paper. This was not a sophisticated hoax. This was a man typing the same sentence hundreds of times and submitting it as scholarship. The journal’s review process — if it had one — could not distinguish between a research paper and a repeated obscenity formatted to look like a research paper.

The Trajectory

SCIgen represents the endpoint of a sixty-year arc.

• Ern Malley (1944): Two skilled poets, one afternoon, sixteen poems. Human craft, human deception, human judgment fooled.
• Sokal (1996): One skilled physicist, one paper, one journal. Human craft, human deception, editorial judgment fooled.
• Grievance Studies (2018): Three skilled academics, twenty papers, multiple journals. Human craft, human deception, systematic testing.
• SCIgen (2005+): A program, unlimited papers, multiple publishers. No human craft required. The exploit has been automated. The troll is a script.

When a program passes peer review, the system being tested is not evaluating content. It is evaluating surface features — formatting, jargon density, citation patterns, structural conformity. It is checking whether the paper looks like a paper, not whether it is a paper.

This is the same failure mode at every level of the progression. The Ern Malley reviewers checked whether the poetry sounded like modernist poetry. The Social Text editors checked whether Sokal’s paper looked like cultural theory. The grievance studies reviewers checked whether the papers confirmed their expectations. And the IEEE and Springer reviewers checked whether the papers were formatted like computer science papers.

At no point in this sixty-year arc did the gatekeepers check whether the content was true.

The Penetration Test Analogy

In information security, the discovery that a vulnerability can be exploited by an automated tool — a script, a bot, a worm — is qualitatively different from the discovery that it can be exploited by a skilled human attacker. A vuln that requires expertise is serious. A vuln that can be exploited by a script is catastrophic, because the attack scales without limit.

SCIgen is the academic equivalent. When a human hoaxer fools a journal, the system has a security flaw. When a script fools a publisher, the system has no security.

Why This Matters

Hidden individual (Piltdown) → creative forger (Ossian) → ironic artists (Ern Malley) → political critic (Sokal) → systematic campaigners (Grievance Studies) → automated exploitation (SCIgen). This is the arc trolling follows everywhere — from lone tricksters to organized operations to bots. The academy is not exempt from the forces that shape every other arena of human discourse. It just dresses better and takes longer to notice.

Source URLs