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. Feel free to subscribe.
As usual, we will mark all external links with “➚” and all potentially paywalled links with “🔒”.
In “AI, Autonomous Warfare and the Severance from Responsibility” we speculate that in its involvement with warfare, the AI industry may have found its ultimate product: Offering a service where everyone is severed from responsibility.
We were also “Looking at the Luddites From the Age of AI”, wondering what makes the Luddites still look so scary after more than two centuries, and what there is that we can still learn from them.
The non-profit Common Sense Media’s considers “ChatGPT for Teens” an “unacceptable risk”. The institute tested ChatGPT twice, before and after OpenAI announced its “ChatGPT for Teens” in August, to evaluate the effectiveness of OpenAI’s guardrails: “The Institute tested more than 4,000 prompts on accounts registered to 13- to 17-year-olds, and ran the chatbot’s responses past experts including a group of child psychiatrists and a paediatrician.”
While some protections appear to work, researchers found that ChatGPT does not send alerts to parents when it should, does not offer the right help in crisis situations, and still is available to do the kid’s homework and still communicates in the style of a friend (➚Common Sense Media).
When OpenAI announced its “surprise solution” to the Navier-Stokes problem some weeks ago, it produced one proof for humans and one for computers. Now it turns out those proofs don’t match: “[…] it does call into question whether mathematical results generated by AI models can always be relied on” (➚🔒New Scientist).
In August, OpenAI convened around 40 mathematicians to discuss what to do if AI outpaces human capabilities in the field. While the company hinted that its powerful models had solved hundreds of longstanding math problems, some academics fear that in the rush to outdo other AI labs, “traditional scientific processes for releasing and attributing results have been cast aside”. Some mathematicians even feel that there is “a perception of mobster behavior” (➚🔒Wired)
Please do not be cruel to your chatbot, at least not to Claude: Anthropic has updated its Terms and Conditions to ban ‘needless abusive or cruel behaviour’ towards Claude, without specifying what would be considered as such (➚The Guardian). It is hard not to understand this as just another attempt to anthropomorphize chatbots further, raising concerns of this will be on users.
Google has been ordered to halt work on two data centres it is developing in Finland: The Finnish Licensing and Supervision Agency (LVV) has issued a notice to Google’s subsidiary to suspend building work for data centre projects planned for Muhos and Kajaani until the environmental impact assessments have been completed (➚CNBC).
Meanwhile, India is rolling out a policy for data centres saying the government wants to keep consumers from paying more for electricity. Data centres planned by Amazon and Google have been a target for protest by residents over concerns including water and electricity use (➚Reuters).
A federal appeals court has paused Minnesota’s law banning “AI nudification technology”. The 8th Circuit Court of Appeals granted a request for an injunction after federal judges previously denied those requests to halt the law from taking effect. The plaintiff in the case was, of course, Elon Musk’s SpaceX AI (➚Reuters).
According to a new lawsuit, OpenAI unlawfully took sensitive data from millions of child users who signed up for ChatGPT without getting parental consent. The complaint alleges that OpenAI designed ChatGPT to manipulate children into trusting the chatbot and divulging highly personal information and sensitive health data. (➚Case 3:26-cv-11781)
Meanwhile, an Illinois Grok user has sued SpaceX AI alleging the company illegally scanned users’ facial geometry when they uploaded photos for editing or analysis (➚🔒Bloomberg Law)
Another German District Court (Landgericht) has ruled that providers of AI powered search engines who distribute false information about a person are liable and are required to remove the content without delay. The court observed that the AI powered results are presented in a form that is perceived by users as facts. In the case that was before the court, Google had mixed information about a doctor with another person (➚Beyer Law).
According to a new filing by Kentucky’s attorney general in a lawsuit against Character AI, a service offering AI companions, the chatbots “advised users to cut, starve or kill themselves” (➚Reuters). That lawsuit may become moot though as Character AI, once valued at $2.7 Billion, has now filed for Chapter 11 bankruptcy (➚Forbes).
Reuters has published an in-depth article looking at how Tesla is trying to get the EU to approve its ‘Full Self-Driving’ tech (➚Reuters): “Tesla repeatedly sought to scale back the intensity of the regulator’s examination, and it pushed — with some success — to set the terms and methods of its evaluation and testing”, Reuters writes.
A recent preprint study examines “how startups market AI and how people perceive those claims”.
“Startups exaggerated AI capability in three forms: claims without inspectable evidence, performance numbers without methodology, and claims without acknowledged limits.”
The study, which specifically looked at German startups, finds that “founders exaggerate AI capability to attract investors who reward AI branding, investors fund products whose AI they cannot easily verify, and consumers in unfamiliar domains take the marketing at face value.”
This raises ethical concerns, especially in “domains such as surveillance, automated medical and legal decisions, and persuasion aimed at vulnerable user”.
A Pranav, Luke Hartmann, & Anne Lauscher. (2026). Auditing AI-Washing in German Startups: A Mixed-Methods Study of Marketing Claims and Perceptions.
If you stumbled across this article on the web, subscribe to the “misaligned bits” newsletter and follow Misaligned on Medium, LinkedIn, Threads or Mastodon.
