6.2 KiB
IGF 2026 WS #416 Survival of the Convenient: Model Selection and Tracing Harm
Description
AI models spread because they’re well adapted to their environment: procurement rules, incentive to publish, and token budgets. Meanwhile, better models sometimes vanish because companies deem them computationally expensive, or hard to train. Therefore, the fitness landscape is often determined by those who hold the power. Developers, scientific journals, investors, governments, and the affected populations do not share that power equally.
Not until something goes wrong does it become apparent that an AI model has failed. Chatbots endlessly offering evasive answers. Minorities getting credit refused. Medical attention diverted to another patient. Yet, these AI-assisted decisions are challenging to counteract. People sometimes do not know an AI model was involved, and if they suspect so, they lack the data, the expertise and the counsel to prove a pattern.
This roundtable explores three questions: what is the threshold to deploy a model; who is told when issues emerge; and what record must accompany a model so that failures can be detected and corrected.
The roundtable brings two juxtaposing positions into debate. Technical experts argue that there is no perfect risk-benefit balance. Peer-review certifies model use based only on how it was applied to the original dataset. It does not certify the model itself. Policy experts argue that acknowledging model limitations is insufficient unless there is accountability. There must be mechanisms for scrutinizing decisions, and detecting when models go astray.
Short interventions open the session. Most of the time is working time. Participants will be presented with three cases: a public-service chatbot that does not escalate to a human, an unexplained credit refusal, and a clinical risk score that redirects care.
Working from these cases, participants draft the minimum record that should travel with a deployed model, corrected live and published as the session's output.
Policy Question(s)
(1) What evidence should be required before an AI model is deployed in a setting where it affects access to public services, credit or care, and who should set that threshold?
(2) When a deployed model is found to be failing, who should be informed, by whom, and on what timescale — the affected individual, the deploying institution, the regulator, or the public?
(3) What record should accompany a consequential AI model into deployment so that failures can be detected and corrected, and who should be able to obtain it?
Expected Outcomes
The session will produce a short, plain-language statement of the minimum record that should accompany a consequential AI model into deployment: data provenance, the population the model was validated on, justification of the performance measure, version history, decision logs, escalation pathways to a human, and drift monitoring. Participants draft it during the working segment; it is corrected aloud in the room, circulated to registrants within two weeks, published as the session report, and offered to ongoing IGF and WSIS+20 follow-up work on AI accountability.
Participants will take away an account of why weaker models often displace better ones, from people who build and deploy them; a clear understanding of what peer review does and does not certify; and a concrete sense of what someone affected by an automated decision would need in order to establish a pattern and contest it.
Format
Roundtable Duration (minutes): 60 Format description: Roundtable. Most of the session is participant working time rather than presentation. A roundtable lets mixed groups form and re-form without rearranging the room mid-session, and seats speakers among participants rather than facing them, which matters when the output is drafted collectively rather than delivered. Classroom and theatre layouts impose a presenter-and-audience geometry that would work against the format.
Sixty minutes is chosen deliberately. Interventions are capped at four minutes, which obliges speakers to state a position rather than deliver a paper, and leaves roughly half the session for participants to work the three cases and report back. The tight clock is what keeps the format honest: with less time to fill, the discussion has to reach the evidentiary question instead of circling the framing. The roundtable layout supports this by removing the cost of moving people into and out of groups.
Hybrid Format
Hybrid Format: Interaction is structural rather than additive, and a short session makes that more important: when time runs out, remote participants are usually what gets cut. The design removes that discretion.
The online speaker opens the session, so remote participation frames it rather than trailing it. She then facilitates the online group, which works its assigned case in parallel with the onsite tables, giving remote participants an expert working alongside them rather than a side channel. Report-backs run online first, before any onsite table speaks. A dedicated online moderator, who is not a speaker, holds a standing right to interrupt.
Both rooms are given the same task and write into the same document, so no contribution is separable by location.
Tools: one Mentimeter poll at the open, displayed in-room and on-screen; a shared collaborative document for capture; the IGF-provided video channel and captioning.
Organizers
Organizer 1: Deliz-Aguirre Rafael, Weizmann institute of Science – Fulbright Program Organizer 2: Raashi Saxena, 🔒 Accessibility Lab
Speakers
Speaker 1: Deliz-Aguirre Rafael, Technical Community, Western European and Others Group (WEOG) Speaker 2: Melanie Nuesch Germano, Technical Community, Latin American and Caribbean Group (GRULAC) Speaker 3: Raashi Saxena, Civil Society, Asia-Pacific Group Speaker 4: Effoduh Jake, Civil Society, African Group
Moderator
Raashi Saxena, Civil Society, Asia-Pacific Group
Online Moderator
Melanie Nuesch Germano, Technical Community, Latin American and Caribbean Group (GRULAC)
Rapporteur
Deliz-Aguirre Rafael, Technical Community, Western European and Others Group (WEOG)
Participant Benefits
What will participants gain from attending this session?
SDGs
3.8 8.10 10.3 16.10 16.6