TRAIF2026—Nikolina Filiposki

Item: TRAIF2026—Nikolina Filiposki
Rating: 4
Author: PostSingularity Institute

by Nikolina Filiposki · View the channel on PSI

Applied Sciences & Engineering Arts & Architecture AThe Arts ATPerforming arts

🎙 Nikolina Filiposki 👥 395 📅 October 21, 2026 ⏱ 42 min 👁 0 📄 original study 🧭 2026-10-21

Available in: English (current) Français

Keywords

content moderationPolicyLinkLLMtransparencyAI governance

Summary

In this talk, Nikolina Filiposki presents PolicyLink, a framework for transparent and reviewable content moderation decisions using large language models. She begins by framing content moderation as a governance problem, noting that platforms set rules, interpret them, and enforce them at scale, but the link between written guidelines and decisions is often opaque. PolicyLink is designed to bridge this gap by retrieving relevant community guideline passages and returning verdicts (allow, remove, or context-dependent) grounded in explicit citations. The system separates multimodal perception, policy retrieval, and verdict logic, allowing inspection at each stage. It uses a hybrid retrieval approach combining BM25 and FAISS, and includes an explicit abstention category for cases with weak or ambiguous evidence. Preliminary results on a few hundred TikTok and YouTube Shorts videos show that the system most often returns ‘context-dependent’, indicating conservative behavior. The talk also discusses the broader research design, which will later compare PolicyLink judgments with platform decisions and human interpretations. The contribution emphasized is the transparent, uncertainty-aware decision architecture rather than a black-box classifier.

171 words

Critical Evaluation

Value of the Information & Strength of the Argument

The value of the information is high for researchers and practitioners interested in transparent AI systems for governance. The presentation clearly articulates the problem, the design choices, and the preliminary evidence, with a strong conceptual argument for why transparency and abstention are crucial in moderation contexts. The argumentation is solid: she justifies each design decision, explains the limitations of existing approaches, and provides empirical evidence that the system behaves as intended. The Q&A section further clarifies technical aspects, such as multimodal integration and validation methods, reinforcing the robustness of the methodology. However, the results are preliminary and the system has not yet been fully evaluated against human judgments or platform decisions, so the current evidence is suggestive rather than conclusive.

Scientific Rigor, Source Quality, Title Accuracy

The scientific rigor is commendable: the presentation is well-structured, methodology is detailed, and the author acknowledges limitations (e.g., OCR not working). Sources are primarily institutional (TUM, alignAI, Responsible AI Forum) and the provided description includes links to these projects. However, no specific academic papers are cited during the talk itself, and the work is presented as ongoing research with preliminary results. The title is generic and does not reflect the content’s substance; it merely labels the conference session. Overall, the rigor is good, but the lack of peer-reviewed sources and the preliminary nature of the findings temper the overall assessment.

235 words

Title / Content Match

The title is just the conference label and presenter name, not descriptive of the content. It does not convey the topic of the presentation, but this is typical for academic conference recordings.

Quality & Reliability

7/10

Presentation of ongoing research with a clear methodology and preliminary results. The system design is rigorous, but results are not yet validated with the full human study, and the independence of the evaluation is limited as it is the author's own work.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

alignAI — Mentioned as the funding project for the Responsible AI Forum 2026.
Institute for Ethics in Artificial Intelligence (IEAI) — The presenter's research institute at TUM.
The Responsible AI Forum — The event where this talk was given.
Amerikahaus — Venue providing video footage for the event.
IEAI Events — Information about other events by IEAI.
IEAI Newsletter — Signup for newsletter to receive updates.

Contribution & Novelties

The original contribution of PolicyLink is a transparent, retrieval-augmented framework for content moderation that explicitly separates perception, policy retrieval, verdict commitment, and explanation. Unlike black-box classifiers, it provides a reviewable link between content and specific platform guidelines, and it includes an abstention category for uncertain cases, making the system uncertainty-aware. This design choice has governance relevance, as it helps clarify when an automated system should defer to human judgment.

Pour aller plus loin :

Retrieval-Augmented Generation (RAG) — Concept underlying PolicyLink’s hybrid retrieval approach.
Content moderation — Overview of the policy and practice that PolicyLink addresses.
Zero-shot learning — Background on how the system makes decisions without training on platform labels.
AI ethics — Broader context for transparency and accountability in AI systems.

123 words

Radar Profile

The radar chart shows balanced scores across quantity, quality, and reliability, with a slightly lower technical level. This indicates a presentation that is informative and well-structured, but not overly technical, making it accessible to a broader audience. The reliability score is moderate due to the preliminary nature of results, which is consistent with an ongoing research talk.

Reliability 7/10