misaligned bits #11: Preparing For Impact

Welcome to a new edition of misaligned bits, the newsletter by misaligned where we sum up recent news, with a lighter touch.

In this edition: Insurers are unsure about AI. Grok has another lapse. And LLMs are seriously biased against people who speak dialect.

As usual, we will mark all non-medium links with “➚” (external link) and all possibly paywalled links with “🔒”.

Misaligned Recap

For your convenience, we have compiled a list of studies from 2025 that are related to ethical questions around AI. The studies give a glimpse at the multitude of ethical questions and may serve as a warning at a time when the AI industry overpromises the capabilities of AI and attempts to push its application into more and more critical areas.

​In Hello Computer!, published in “bootcamp”, we wonder if talking to AI is actually a good idea for human-machine interfacing, and if it means that AI companies have access to your most private data, is even something that we should desire.

Assessing risk and preparing for impact

Liability, or the question of who is responsible when LLMs go haywire, is still an issue under intense discussion. We recently tackled this question in the article “Anyone responsible? — After The Demise Of The EU AI Liability Directive”.

Insurers appear to become increasingly nervous about covering those cases, as covered by the Financial Times, and some have started to exclude “any actual or alleged use” from their coverage.

It has long been predicted that liability (and by extension insurance) will be one of the major stumbling blocks for the AI uptake in companies. This problem is becoming even more pronounced with the emergence of Shadow AI.

Insurers are supposed to evaluate the risk of AI services, but as an article by law firm Hogan Lovells points out in an article from October, “as AI tools begin to interact directly with customers or influence decision-making, the risk profile increases. […] Liability insurers, in particular, will want to think carefully about the interaction of AI risk with traditional coverages.” (➚Hogan Lovells)

As Jocelyn Auger writes for Fasken: “Recent incidents involving AI-generated misinformation, automated customer-service tools making inaccurate statements, or sophisticated impersonation frauds have only heightened these concerns. While individually manageable, they illustrate how unpredictable AI behavior can be — and how difficult it is for insurers to assess where future claims might arise.” (➚Fasken)

For further reading, an article from BrownJacobsen also compares the situation of AI liability in the EU and UK (➚BrownJacobsen).

Risky business

Without regulation requiring transparency and a central (and public) incident database, it will be extremely difficult for insurers to judge the risks.

The MIT model risk catalogue might therefore worth to look at. The repository is part of the MIT AI Risk Initiative, which aims to increase awareness and adoption of best practice AI risk management across the AI ecosystem.

Safeguards?

While we are on the topic of unpredictability: Lapses in safeguards (or more likely the absence of any) led to a wave of sexualized images generated by Elon Musk’s grok this week. xAI says it is working to improve its systems while blaming the users. (➚The Guardian)

It may be good to remember that in other industries, you have to implement safeguards (and test them) before you release a product to the public.

Datacentres everywhere?

Meanwhile, local resistance against building AI data centres everywhere appears to be growing: Between April and June 2025 alone, the latest reporting period, 20 proposals valued at $98 billion in 11 US states that were blocked or delayed amid local opposition and state-level pushback. (➚Associated Press)

Cheating everywhere!

The Association of Chartered Certified Accountants (ACCA), which has almost 260,000 members, has said that from March it will stop allowing students to take online exams.

The ACCA said it had concluded that online tests have become too difficult to police, given the rise in artificial intelligence (AI) tools available to students. (➚The Guardian)

Disinformation

Poland has asked the European Commission to investigate TikTok after the social media platform has been flooded with AI generated videos calling for Poland to withdraw from the EU. (➚Reuters)

The EU AI Act will mandate labelling for AI-generated content like deepfakes and synthetic media to ensure transparency, with rules coming into force on August 2, 2026, requiring both visible labels and machine-readable watermarks for AI generated output.

Bits of research

It appears that LLMs have some serious bias issues when talking to people with local dialects, a study has found, with serious impacts on AI in recruiting.​

The models were asked to describe the speakers of these texts with personal attributes, and to then assign individuals in different scenarios. The models were asked who should be hired for low-education work or where they think the speakers lived. (➚Paper)

In nearly all tests, the models attached stereotypes to dialect speakers. The LLMs described them as uneducated, farm workers and needing anger management. What is worse, the bias grew when the LLMs were told the text was a dialect. (➚DW)

Minh Duc Bui, Carolin Holtermann, Valentin Hofmann, Anne Lauscher, and Katharina von der Wense. 2025. Large Language Models Discriminate Against Speakers of German Dialects. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, pages 8223–8251, Suzhou, China. Association for Computational Linguistics.

A happy new year to all readers!

Wolfgang Hauptfleisch, misaligned

Follow and subscribe

PS: If you stumbled across this on the web, subscribe to the “misaligned bits” newsletter here and follow misaligned on Medium, LinkedIn or Mastodon.