Towards Explainable Fact Checking, Read Twice
One doctoral thesis on automated fact checking, summarized twice: first only from what is public about it, then from the thesis itself. The two are compared claim by claim. It is one thesis, so the page shows what happened here, not a rate.
The record
| Fact | What the sources say | Label |
|---|---|---|
| Title and author | Towards Explainable Fact Checking, Isabelle Augenstein | In the thesis · p. i |
| Degree | “Doctor Scientiarum (Dr. Scient.)”, Faculty of Science, University of Copenhagen; the title page is dated “January 2021”. The preface calls it a doktordisputats, “in form and function equivalent to the French and German Habilitation and the Higher Doctorate of the Commonwealth”. | In the thesis · pp. i-ii |
| Defense | Monday Dec 06, 2021, 13:00, Universitetsparken 1, Copenhagen. Opponents: Professor Iryna Gurevych (Technical University of Darmstadt) and Professor Lucia Specia (Imperial College London and University of Sheffield). Chair: Christian Igel. | Record KU |
| On arXiv | arXiv:2108.10274; version 1 Aug 23, 2021, and version 2 Dec 08, 2021 | Record AX |
| Summary article | “Habilitation Abstract: Towards Explainable Fact Checking”, KI – Künstliche Intelligenz 36 (2022), pp. 255-258 | Record KI |
| Size | 288 PDF pages; ten published papers, 2016 to 2020; eleven chapters in four parts | In the thesis · pp. vi-vii, xv-xx |
Pass 1: from what is public about it
Four sources, opened on the evening of Tuesday Oct 20, 2026, Chicago time, without opening the thesis: the arXiv abstract page; the University of Copenhagen’s defense notice, read in the Internet Archive’s copy because a plain request for the live page returned HTTP 404; the author’s summary article; and the OpenAlex index record. They support these 10 claims:
- Record 1. Presented to the University of Copenhagen Faculty of Science for the degree of Doctor Scientiarum.
- Record 2. Defended Monday Dec 06, 2021; opponents Gurevych and Specia; chair Igel.
- Record 3. arXiv v1 Aug 23, 2021; v2 Dec 08, 2021.
- Record 4. Summary article in KI 36 (2022) 255-258 calls it a habilitation thesis.
- Record 5. OpenAlex: preprint, 2021, cited 5 times.
- Author’s description 6. Covers check-worthiness, stance detection, veracity prediction.
- Author’s description 7. Contributions extend to learning with limited labeled data.
- Author’s description 8. Presents some first solutions for explainable fact checking.
- Author’s description 9. Three verticals: subtask, low-label methods, explainability.
- Author’s description 10. Motivation includes EU legislation requiring explanations and platform transparency.
Pass 1 added one caveat of its own: every claim about content is the author describing her own work, so nothing in it shows how well the thesis delivers. And it listed what it could not say, which a writer would be tempted to fill in: how many papers the thesis holds, any dataset or number, which explanation methods it proposes, which EU laws it means, and what the examiners thought.
Pass 2: from the thesis itself
- In the thesis · pp. 2-3 It works on the content-based pipeline: detecting check-worthy claims, retrieving and ranking evidence, stance detection, and veracity prediction. Bot detection is “beyond the scope of this thesis”.
- In the thesis · pp. 10-11 Its ten papers are arranged along three axes: the fact-checking sub-task, the method for limited labeled data, and the explanation method.
- In the thesis · p. 12 On check-worthy claims, a unified approach “is achievable for 2 out of 3 tasks”.
- In the thesis · pp. 19-20 Its dataset MultiFC holds 34,918 claims from 26 English fact-checking websites, and the best model reaches a Macro F1 of 49.2%.
- In the thesis · p. 21 On claims generated to fool a fact-checking model, that model “performs 23.8% worse” than on the original FEVER claims.
- In the thesis · p. 22 Its explanation paper is presented as “the first study on generating veracity explanations”.
- In the thesis · p. 34 Its explainability results are “some of the very first findings on this topic, published only in 2020”. Of the explanation paper’s dataset, with about 10,000 training examples, it says “it is very difficult to beat simple baselines”.
- In the thesis · p. 35 The EU laws it means are the Digital Services Act and a “Right to Explanation”, citing Goodman and Flaxman (2017).
Pass 1 against the thesis, claim by claim
| # | Claim in pass 1 | Against the thesis |
|---|---|---|
| 1 | presented to the University of Copenhagen Faculty of Science for the degree of Doctor Scientiarum | Matches the thesis. The title page says the same (p. i). Pass 1 had it from the arXiv listing, not the thesis. |
| 2 | defended Monday Dec 06, 2021; opponents Gurevych and Specia; chair Igel | A record; the thesis cannot speak to it. Only the University of Copenhagen notice has the defense details. |
| 3 | arXiv v1 Aug 23, 2021; v2 Dec 08, 2021 | A record; the thesis cannot speak to it. From arXiv; the PDF’s arXiv stamp is dated Dec 08, 2021. |
| 4 | summary article in KI 36 (2022) 255-258 calls it a habilitation thesis | A record; the thesis cannot speak to it. The preface agrees that the degree is equivalent to a Habilitation (p. ii). |
| 5 | OpenAlex: preprint, 2021, cited 5 times | A record; the thesis cannot speak to it. An index count, read Oct 20, 2026 (Chicago time). It will change. |
| 6 | covers check-worthiness, stance detection, veracity prediction | Matches, but narrower than the thesis. The thesis’s pipeline has a fourth step, evidence retrieval and ranking (pp. 2-3), handled only inside the MultiFC paper (p. 11, footnote 2). The abstract’s three tasks match the thesis’s own chapters, so this verdict is a judgment call. |
| 7 | contributions extend to learning with limited labeled data | Matches the thesis. The thesis says the same (p. 10, Table 1 on p. 11). |
| 8 | presents some first solutions for explainable fact checking | Matches the thesis. The thesis says its explainability results are among the very first, published in 2020 (p. 34). |
| 9 | three verticals: subtask, low-label methods, explainability | Matches the thesis. The thesis calls them axes (p. 10), and once a “vertical” (p. 8). |
| 10 | motivation includes EU legislation requiring explanations and platform transparency | Matches, but narrower than the thesis. The thesis names the laws: the Digital Services Act and a Right to Explanation (p. 35). The abstract names none. |
The questions pass 1 left open
| Question | After reading | What the thesis says |
|---|---|---|
| How many papers and chapters | Answered | 10 papers (pp. vi-vii, Table 1 p. 11); 11 chapters in 4 parts (pp. xv-xx) |
| Datasets, models, results, any number | Answered | e.g. MultiFC 34,918 claims from 26 sites, best Macro F1 49.2% (pp. 19-20); adversarial claims 23.8% worse (p. 21); LIAR-PLUS 10k training instances (p. 34) |
| Which explainable methods | Answered | three kinds: natural-language explanations, adversarial (model-level) explanations, post-hoc rationales (pp. 8-10, 18-22) |
| Which EU laws | Answered | Digital Services Act; Right to Explanation (Goodman and Flaxman 2017) (p. 35) |
| What the assessment committee said | Still open | not in the thesis; KU says the report is available from the Faculty Secretariat on request |
What reading showed that pass 1 did not
Not the thesis Eight references name the wrong paper. The thesis lists its ten papers on pp. vi-vii. A script found every “Paper” or “Papers” followed by a number in chapter 1: 65 mentions. 19 of them are the section headings and the figure and table captions that label each paper, and all of those agree with the list. We read each of the other 46, the references in the running text, against the list: 37 are right, 1 is unclear, and 8 point to the wrong paper. Seven of the 8 use the chapter number in place of the paper number; one names a different paper. These are slips in a long document, and the list itself and the section headings are consistent. They matter for anyone who cites a paper by the number in the prose.
| Page | The thesis says | The list’s paper |
|---|---|---|
| p. 9 | “In Paper 10 (Section 10), we explore how to generate adversarial claims” | Paper 9, Adversarial claims (ch. 10) |
| p. 10 | “Paper 8 (Section 8) focuses only on the subtask” | Paper 7, Diagnostic study of explainability (ch. 8) |
| p. 10 | “Paper 8 addresses the highly challenging task” | Paper 7, Diagnostic study of explainability (ch. 8) |
| p. 11 | “only a contribution in Paper 9” | Paper 8, MultiFC (ch. 9). P. 25 credits joint evidence ranking to Paper 8 |
| p. 26 | “our work on explanation generation (Paper 9, Section 10)” | Paper 10, Generating explanations (ch. 11). The explanation paper is Paper 10, chapter 11 |
| p. 34 | “In Paper 10, on generating adversarial claims” | Paper 9, Adversarial claims (ch. 10) |
| p. 34 | “Paper 8, which studies post-hoc explainability techniques” | Paper 7, Diagnostic study of explainability (ch. 8) |
| p. 34 | “Paper 7 is the only one I am aware of which tackles this” | Paper 6, Temporal domain adaptation (ch. 7). P. 31 says Paper 6 for the same claim |
Not the thesis A legal date, checked. The thesis says the Digital Services Act was “passed in December 2020” (p. 35). The regulation itself is dated Oct 19, 2022: Regulation (EU) 2022/2065, published in the Official Journal (L 277) on Oct 27, 2022, after the European Parliament’s position of Jul 05, 2022 and the Council’s decision of Oct 04, 2022. So when the thesis was dated, in “January 2021”, the act had not passed. Its bibliography cites a European Commission page dated “Dec. 2020” and titled “Shaping Europe’s digital future: The Digital Services Act package” (p. 219), which suggests the sentence means the Commission’s proposal. Anyone citing the thesis for when the act became law should use 2022. EUR-Lex would not load for our tools, so Lawrence Ho saved the regulation from EUR-Lex in his own browser on Oct 21, 2026.
Caught before publication
In the first draft, the drafting model typed the number of matching claims as 5 when the list holds 4, and counted four wrong cross-references where there are 8; the first was caught by recounting the list, the second by the fact-check and hostile reviews of this page, which also found that our first search missed the plural “Papers”. The page now computes every count from its data files when it is built, and the build stops if they do not add up. That check catches arithmetic, not a wrong entry in the data.
Sources
Opened on the evening of Tuesday Oct 20, 2026, Chicago time (Oct 21, 2026 in universal time), through our access ladder. The regulation (DSA) came from its last rung, a person, and was saved on Wednesday Oct 21, 2026. Each saved copy is kept with its SHA-256; the first 16 characters are shown.
- T The thesis: arXiv:2108.10274, version 2 (PDF), 288 pages. Rung A.
6c67c75a0b59ff1b. Quoted in short passages only; arXiv’s license lets arXiv distribute the thesis and gives no one else the right to republish it. - AX The arXiv abstract page. Rung B.
dd44d61cfb68e769 - KU University of Copenhagen, Department of Computer Science, “Doctoral dissertation defence by Isabelle Augenstein”, read in the Internet Archive’s copy of Jun 29, 2022, because a plain request for the live page returned HTTP 404. Record.
6cea513bf2b9c52e - KI Isabelle Augenstein, “Habilitation Abstract: Towards Explainable Fact Checking”, KI – Künstliche Intelligenz 36 (2022), pp. 255-258, doi 10.1007/s13218-022-00774-6, read in its record at dl.gi.de. Rung B.
b96837df70feca5c - DSA Regulation (EU) 2022/2065 (the Digital Services Act), Official Journal L 277, Oct 27, 2022, 102 pages, saved from EUR-Lex by Lawrence Ho on Oct 21, 2026. Rung A for its own date and contents.
04b688f2ea13fc0f - OA OpenAlex work record for doi 10.48550/arXiv.2108.10274. Record.
866689df16f30f2f
Not used: the Semantic Scholar API, which refused our plain request as too many (HTTP 429); the record is queued for a person. Could be asked for: the assessment committee’s report, which the University of Copenhagen supplies on request.
Attestation
Attestation. Prepared under the Ten House Protocols. Drafted by Claude Opus 5.5, an Anthropic model, in Protocol STAMP’s working session. Reviewed independently by two further Claude Opus 5.5 instances, each working apart from the draft’s session: a fact-checker, in four rounds, and a hostile reader, in two; every required fix was made. Reviewed from outside Claude’s family by Xing 4.0, drawn at random by recorded seed we2-20261021 (R14), which judged it publishable with fixes; that review read the first draft, not this final page. Approved for publication by Lawrence Ho, by chat message (“I’ve read the page. Approved; sign it with today’s date.”), Wednesday Oct 21, 2026, 11:48 CDT, who answers for it. The record does not show that he read every source behind each number, date and quote, so we do not claim it. Every factual sentence names a source we opened. If you find an error, tell us and we will correct it in public.