alexloth.com

Five Things I Took Home from AIES 2026 in Malmö

—

by

Alexander Loth
in

AIES 2026, the AAAI/ACM Conference on AI, Ethics, and Society, ended
yesterday in Malmö. On Monday I presented our paper The Mitigation
Paradox in the discussion session on Participation, Sovereignty, and
Power in AI Governance; I wrote about that session here.
This post is about the rest of the week, and about how much of it kept
circling back to one finding from our survey of 54 experts in 21
countries: the field agrees on the threats of AI disinformation and
splits on what to do about them.

1.
Agreement on the problem is cheap, agreement on the fix is not

The Best Poster award went to Kevin Baum, Markus Langer, and
colleagues for work that treats human oversight of AI as a security
problem, with the overseers themselves as an attack target. Human
oversight is a good test case for our finding. Almost everyone agrees
that someone should be able to watch and stop an AI system. Whether to
mandate it, and how, is where camps form.

That matches what we see in our data. 78% of the experts we surveyed
name election deepfakes as the most urgent risk, yet government
regulation is both the most-chosen and, for 15% of the same panel, the
least effective countermeasure.

2. Research
does not move into policy on its own

A recurring line in Malmö was that policymakers should act on the
ideas presented. Suresh Venkatasubramanian pushed
back on that on Bluesky: individual papers rarely arrive “policy
ready”. I agree, and it is the reason our paper does not end with a
single recommendation. It maps patterns of expert disagreement to four
kinds of action: bind where experts converge, pilot where they split
into camps, differentiate where regions disagree, and monitor where
signals are weak.

3. Audits only help
when they reach real use

Several posters asked what happens after an audit. Hannah Cha argued
that harm in personalized generative AI can only be identified at the
interaction level, with users, not in simulated audits. Mowafak Allaham
and Nicholas Diakopoulos audited four generative search engines on 712
real queries and found that about one in six cited sources is
AI-generated, most often on health and environment topics. Iman Peljto,
Xenia Heilmann, and Mattia Cerrato asked whether German algorithm
registers are actually transparent.

The synthetic-sources result is the one I will carry into my own
work. If generative search engines cite AI-generated pages as evidence,
provenance stops being a property of a single piece of content and
becomes a property of the whole retrieval chain.

4. Who decides what counts as
harm

Shiyun Tang and Zezheng Lin’s “Simulation as Alibi” showed how design
choices in LLM agent simulations can hide behind the appearance of
emergent behavior. Minyeong Bae, Hyunjae Kim, and Kyung Myun Lee studied
gender bias and gender performance in AI-generated music. Poster 898 on
the future of generative AI for Black communities compared public
discourse with scholarly research and found the two misaligned.

5. The keynotes set the frame

Sandra Wachter opened the conference with a keynote on how
geopolitics, deregulation, and the AI hype cycle threaten human rights
protection. Stefan Larsson told the story of moving on from copyright
law to AI, only to find that his own book on copyright had been used to
train LLMs. Both made the same point from different sides: the rules we
rely on are under pressure exactly where the technology moves
fastest.

Read the paper

If you work on AI disinformation, platform governance, or expert
elicitation, I would be glad to hear where our findings hold up and
where they do not.

@inproceedings{loth2026aiesexperts,
  author    = {Loth, Alexander and Sch{\"u}tz, Mina and Kappes, Martin and Pahl, Marc-Oliver},
  title     = {The Mitigation Paradox: How {AI}-Disinformation Experts Converge on Threats but Polarize on Solutions},
  booktitle = {Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society},
  year      = {2026},
  volume    = {9},
  number    = {3},
  pages     = {1722--1733},
  doi       = {10.1609/aies.v9i3.43261}
}

Thanks to my co-authors Mina Schütz, Martin Kappes, and Marc-Oliver
Pahl, to Haiyi Zhu for chairing our session, and to everyone who stopped
by the poster.