Welcome!
The 1st Workshop on AI-Generated Content Detection and Trustworthy Governance (AIGCTrustDetect) will take place in Macau, China as part of COLING 2027 (May 9–14, 2027).
Going beyond binary AIGC detection toward full-stack, explainable, and evasion-resistant governance that generalizes across languages and deploys cost-effectively in the real world.
A binary label cannot say which parts a model touched, whether the events described are real, or how a detector fares in a language it was not built for. AIGCTrustDetect treats detection as one component of trustworthy governance rather than the whole of it: where machine authorship must be ruled out, detection has to be reliable; where AI assistance is welcome, the content itself has to be reliable.
The program combines invited keynotes, peer-reviewed research papers and a three-track shared task, bringing academic research and industrial practice into one forum.
A binary detector labels this text as AI-generated and stops there. It cannot say which parts a model touched, or that the events described are real.
{{verdict}} {{summary}}
{{n.text}}
News
| October 17, 2026 | Website is up! |
|---|
Important Dates
| First call for papers | Friday | October 30, 2026 |
|---|---|---|
| Second call for papers | Wednesday | November 25, 2026 |
| Direct paper submission deadline | Tuesday | January 5, 2027 |
| Pre-reviewed ARR commitment deadline | — | TBA |
| Notification of acceptance | Saturday | February 6, 2027 |
| Shared task call for participation | Tuesday | November 10, 2026 |
|---|---|---|
| Registration opens | Tuesday | November 10, 2026 |
| Training and development data release | Wednesday | November 25, 2026 |
| Test data release | Sunday | December 20, 2026 |
| System output submission deadline | Friday | December 25, 2026 |
| Evaluation results release | Wednesday | December 30, 2026 |
| System description paper deadline | Sunday | January 10, 2027 |
Common Dates
| Camera-ready deadline | Friday | March 5, 2027 |
|---|---|---|
| Proceedings due (hard deadline) | Monday | April 5, 2027 |
| Pre-recorded video due (hard deadline) | Thursday | April 15, 2027 |
| Workshop day | [TBD] | [TBD], May 2027 |
| COLING 2027 virtual conference | Thursday – Friday | May 6–7, 2027 |
| COLING 2027 main conference | Sunday – Friday | May 9–14, 2027 |
All deadlines are 11:59 pm UTC-12h (anywhere on earth).
Submission details are in the Call for Papers; shared task timelines are on the Shared Tasks page.
Topics of Interest
We invite submissions across five directions grounded in practical needs. We picture them as a stack: reliable detection sits at the base, and each layer above it addresses a question that deployment raises and a binary label cannot answer. The full list of topics is in the Call for Papers.
Select a layer to read its description.
{{l.desc}}
Interactive Examples
Each shared task track targets a question that a binary detector cannot answer. Try one example from each before reading the task descriptions.
Who wrote this?
One example per track, written for this page; not drawn from task data.
Track 1. Is this human-written, LLM-refined, or LLM-generated?
Our method is simple: we rewrite the text once and measure how much changes. Notably, this approach not only enhances robustness but also offers valuable insights into the nuanced dynamics of authorship.
{{fb1}} The first sentence reads as a plain human draft; the second carries typical model polish (“notably”, “not only… but also”, “valuable insights”, “nuanced dynamics”). Mixed authorship within one text is the reason Track 1 adds a third label.
Track 2. One of these is an 8th-century Tang poem; the other was generated by a language model. Which is the original?
Poem A
深山松影静， 白云自往还。 鸟归斜照里， 人在翠微间。
Poem B
空山不见人， 但闻人语响。 返景入深林， 复照青苔上。
{{fb2}} Poem B is Wang Wei’s 《鹿柴》. Poem A was generated for this page: it assembles conventional images (pine shade, white clouds, homing birds) without a perceptual turn, whereas Wang Wei’s poem hinges on voices heard in an empty mountain and light returning to the moss. Classical-style generation is fluent at the surface level, which motivates Track 2.
Track 3. One sentence claims more than its evidence supports. Which one?
{{fb3}} The evidence covers one language and one domain, while the claim extends to all of them. This evidence–statement mismatch is what Track 3 asks systems to identify.
{{scoreLine}}
Each answer came with a reason. Most detectors still return only a label. The shared tasks invite systems that can do more.