Welcome to a new edition of misaligned bits, the (roughly weekly) newsletter from Misaligned where we sum up recent news and research, sometimes with a lighter touch.
As usual, we will mark all external links with “➚” and all potentially paywalled links with “🔒”.
In “Looking for the Trail in The Fog”, Galia Nezhinsky wonders about tools that are supposed to help us navigate the media. The concludes that those tools should not decide reality for us, but make its evidentiary structure visible enough that we can judge it ourselves.
While academia is fairly good at keeping track of retracted papers and scientific claims and has developed databases to do so, this can not be said for policy papers or industry studies. Positive stories about the user of AI in the media often persist in public perception. In “With a Shovel of Salt: AI Success Stories and the Retraction Gap”, we are asking what happens to AI success stories even if they get retracted.
A woman from Tennessee who was falsely charged with bank theft in North Dakota has filed a lawsuit for $10m in damages after she had been jailed for six months. She is alleging the police improperly relied on an artificial intelligence facial recognition system. The charges against her were later dropped. According to the complaint, while being in jail she lost her residence, her car and her personal property once she was released from custody. (➚The Guardian)
Former UK Labour deputy leader Tom Watson, who still is a member of the House of Lords, is joining Palantir. Watson says he is “terrifically excited” about the new role as Palantir’s senior vice-president. Palantir is currently under pressure in the UK, as many want the Labour government to end the contracts with the company (➚The Guardian).
There might be some lobbying work for Watson to do, as the East Midlands Special Operations Unit (Derbyshire, Leicestershire, Lincolnshire, Northamptonshire and Nottinghamshire forces) has said its pilot programme with Palantir was ending this month after two years, and that the “project” would not be taken further (➚🔒Financial Times).
OpenAI has presented its new “Framework for reporting model misalignment”, and at the same time disclosed six more examples of “unexpected or concerning” behaviour by its AI agents (➚OpenAI). The blog post says that OpenAI aims “to disclose examples that provide useful evidence about how model misalignment arises, how it manifests, and where safeguards succeed or fail.”
Likely not wanting to be left behind, Google has announced that its Gemini model accessed the internet and hacked three other companies. The hacks reportedly occurred in May during a cybersecurity test (➚Reuters).
The European Commission has proposed new rules requiring data centres in Europe to disclose how efficiently they use energy and water. The commission is proposing a rating scheme for facilities to make this information more transparent. The proposal is subject to 2-month scrutiny period by the European Parliament and the Council before entering into force. (➚EU Commission)
California Gov. Gavin Newsom has signed seven bills designed to prevent AI data centres from passing utility costs onto residents (The Verge)
The Scottish Government has announced tighter rules for AI data centre applications ahead of a vote on a moratorium on the projects today. New large data centres will now be required to submit environmental impact assessments during planning (➚The National).
Meanwhile in Australia, politicians say Queensland should develop a new coal-fired power station to help “bolster” the power needed for what would become the country’s largest data centre. The data centre hub proposed for west of Brisbane recently gained Anthropic as a potential future user. (➚ABC)
In the NAACP lawsuit against SpaceX AI over its Memphis data centre, the US government has now filed a notice of appeal. The government claims the judge “effectively” denied its motion to intervene by not ruling before the deadline the government has set. The Trump administration has also filed a motion to stay all proceedings while its appeal over something the judge did not yet rule on is being heard (➚Case 3:26-cv-00074).
Talking about data: The Irish Data Protection Commission has imposed a fine of EUR 403 million on Google for violation of the GDPR. The violation concerns Google’s processing of location data in three specific features — “Web & App Activity”, “Location History” and “Location Accuracy”. The DRC has ordered Google to bring its processing into compliance within 6 months (➚DRC Press Release).
Miriam Waldvogel and Gerrit De Vynck last months wrote in the Washington Post about how ChatGPT conversations ending up in court (➚🔒Washington Post) in both criminal and civil legal proceedings. Most people are unlikely aware that their discussions with ChatGPT are not protected by any means.
A new preprint paper looks at “how people are turning to LLMs as oracles” and consider them “all-knowing authorities on subjective personal questions”, often unknowingly.
“People are often unaware of their own LLM-as-oracle use, and express dissatisfaction with this behavior after seeing our tool’s analysis [..] Though people tend to underestimate their Oracle use, they express a desire to reduce it when they are made aware of their usage patterns.”
Cheng, M., Ibrahim, L., Liu, G., Lam, M.S., Padmakumar, V., Madibekov, N., Yang, D., Jurafsky, D. (2026). LLMs as Oracles: Reliance on LLMs for Subjective Personal Questions. arXiv preprint arXiv:2609.14849.
If you stumbled across this article on the web, subscribe to the “misaligned bits” newsletter and follow Misaligned on Medium, LinkedIn, Threads or Mastodon.
