alexloth.com

Dissertation Submitted: The AI-Driven Disinformation Ecosystem

—

by

Alexander Loth
in , ,

On September 30, I submitted my dissertation, “The AI-Driven Disinformation Ecosystem: Measurement Validity Across Generation, Perception, Detection, and Mitigation”. The thesis makes two contributions of its own: a reusable sequence of checks for building and evaluating detection benchmarks, and a map that relates claims about disinformation campaigns to the evidence needed to support them. Parts of the underlying studies were presented along the way at the ACM Web Conference (WWW ’26), the AAAI/ACM Conference on AI, Ethics, and Society (AIES 2026), the Harvard Kennedy School Misinformation Review, the ACM Web Science Conference (WebSci ’26), and INFORMATIK 2026. The full list is below.

The topic goes back to my time in Microsoft’s AI for Good Lab, including the lab’s contribution to Microsoft’s 2022 report on cyberattacks in Ukraine. This post covers how it came about, what the thesis argues, and where to find the papers, datasets, and tools.

Decisively Digital

Decisively Digital

Discover the power of artificial intelligence and digital transformation in the #1 best-selling business book Decisively Digital.

Buy on Amazon Learn More

The dissertation was co-supervised by Prof. Martin Kappes (Frankfurt University of Applied Sciences) and Prof. Marc-Oliver Pahl (IMT Atlantique, Chaire Cyber CNI) and submitted to the Doctoral Center for Applied Computer Science (PZAI).

How it started

The path to this topic began before the PhD:

The longer story, from grid computing at CERN to this research, is in From CERN Grid Computing to AI Disinformation Research.

What the dissertation is about

Large language models can write news texts that are hard to tell apart from human writing. The thesis examines what this means for deliberate false information across four stages: how such texts are produced, whether people recognize them, whether machines recognize them, and which mitigation strategies experts consider effective. The stages depend on each other, and the thesis calls this interplay the AI-driven disinformation ecosystem. Throughout, it asks what the collected data actually show and where they stop being sufficient.

The thesis makes two main contributions, both developed in the dissertation itself:

  • Detection benchmarks. A detector’s high hit rate on its own text collection does not show that it recognizes machine authorship in general. The classifiers may only be separating that collection from others. From this analysis, the thesis derives a reusable sequence of checks for building and evaluating such benchmarks.
  • A measurability map. It relates each claim about a disinformation campaign to the kind of observation needed to support it, distinguishing direct evidence, substitute measures (proxies), and what remains unobserved.

Both address the same problem: an observable substitute is taken for a broader phenomenon without testing the link between them.

What we found

The empirical work spans all four stages: a controlled corpus of machine-generated news, a study of human perception on JudgeGPT, an evaluation of machine detectors, and an expert survey on mitigation. Details are in the papers below.

The short version: claims about detection and mitigation are only as reliable as the link between what is measured and what is claimed.

Papers along the way

The dissertation is a monograph. These are the peer-reviewed papers along the way, by venue:

The data and tools are open: three datasets on Zenodo, JudgeGPT and RogueGPT, Origin Lens, and CRED-1. An updated overview of the papers, tools, and datasets is on the AI and Disinformation page.

CRED-1: from dataset to daily use

CRED-1 is the artifact from the thesis that is used most outside research. It is an open, domain-level credibility dataset. It merges two openly licensed source lists, OpenSources.co and the Iffy.news Index, and enriches each domain with four signals: domain age, web popularity from the Tranco list, fact-check frequency, and Google Safe Browsing flags. Each domain gets a score between 0 and 1, and every component stays inspectable.

Where it stands today:

  • Weekly releases. The latest release, v2026-10-13, covers 2,672 domains. Each release is archived on Zenodo under the concept DOI 10.5281/zenodo.18769459, which always resolves to the latest version. Across all versions, the record has more than 600 views and nearly 150 downloads.
  • For developers. The npm package @aloth/cred1 works as a command-line tool, a JavaScript library, and an MCP server for AI agents. It was downloaded 903 times in the 30 days to October 12.
  • The paper. The preprint, written with Martin Kappes and Marc-Oliver Pahl, is on SSRN: doi.org/10.2139/ssrn.6448466. I presented CRED-1 at ACM WebSci 2026 in Braunschweig.
  • In use. The Trackless Links Safari extension for iOS and macOS checks links against CRED-1 entirely on the device, with no data collection (post on X). RogueGPT uses it to identify low-credibility source domains.

The score is a domain-level signal, not a verdict on any single article. The thesis shows that such a signal can run on a device without sending browsing data anywhere. A natural next study is whether it changes how people judge sources.

More on CRED-1: Introducing CRED-1, the preprint, and how a research dataset ended up in a Safari extension (in German).

What comes next

The defense is the next step. After that, I plan to continue in three directions:

  • Applied research. Natural next studies are empirical: whether a credibility signal like CRED-1 changes how people judge sources, and how detectors perform on benchmarks built with the checks the thesis proposes.
  • Tools in use. CRED-1 ships weekly releases, runs in a Safari extension, and its MCP server lets AI agents check a source before they cite it. I plan to keep the datasets and tools maintained.
  • Industry and policy. Platforms, newsrooms, and regulators decide which countermeasures reach people. At AIES, the panel asked how academics can help policymakers. My answer:

The policy instrument should match the shape of the disagreement: where experts agree, bind; where they split, pilot with sunset clauses and treat the disagreement as a design input, not a reason to wait. I would like to keep working on research that feeds into these decisions.

The trust questions from this work also carry over to AI agents that act on our behalf, which I wrote about in From Misinformation to Agentic AI.

Thank you

To Martin Kappes and Marc-Oliver Pahl, for the first conversations in 2023 and for their guidance and feedback since. To my coauthors Mina Schütz, Dominique Conceicao Rosario, and Peter Ebinger. To everyone who judged news fragments on JudgeGPT and to the experts who took the time to answer our survey. And to my colleagues at Microsoft, where this journey started.