ChatGPT’s learning successes retracted, questionable productivity stats, Palantir gets more of UK’s health data, hallucinations are accelerating, and AI assistants make us a bit dumber.

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 non-medium links with “➚” (external link) and all possibly paywalled links with “🔒”.

New in Misaligned this week: “Confident, Polished, and Incorrect: A Case Study in AI Reliability” by David Dill. “When users ask questions that affect real-world decisions, the system must do more than sound right. It must respect the limits of what it knows, stay inside the user’s constraints, and clearly identify when verification is required.”

Also, in our new academic reading list, we look at seven studies that investigated AI bias, stereotyping and looked at the effect on vulnerable users.

Data centres

The developers of Google’s data centres in the UK have significantly misstated how much CO2 the two proposed AI datacentres will contribute to the UK’s total emissions, according to a review by The Guardian. The review found that the developers understated the significance of their emissions by a factor of five. (➚The Guardian)

Legal bits

French prosecutors have opened a criminal investigation into Elon Musk and X. This follows Musk’s failure to respond to a summons to appear for a voluntary interview in April. In a statement, the Paris prosecutor’s office said it had requested that the investigating judges charge xAI, as well as Musk and Yaccarino “by summoning them […] or, in the event of non-compliance, by issuing a warrant equivalent to an indictment.” (➚Le Monde)

In the lawsuit filed by the NAACP against SpaceX/xAI over its data centre in Memphis, district judge Michael P Mills has recused himself on his own initiative. It is not known why. The case has been reassigned to judge Debra M. Brown (➚Recusal Order). Also, the US government has indicated it might intervene in the case (➚Notice of possible intervention). It would be the second lawsuit where the US government would intervene on behalf of SpaceX/xAI within weeks, the other one being the lawsuit of xAI against AI regulation in Colorado.

Regarding AI slop

Music fans are becoming increasingly uncomfortable with AI songs, according to a recent study published by the entertainment insights company Luminate (➚NPR). It found that overall interest dropped from -13% to -20% during that time period from May to November 2025.

Regulatory Bits

On 8 May the European Commission published its draft guidelines to help providers and deployers of generative AI systems comply with their transparency obligations. Feedback to the draft is expected by 3 June (➚European Commission).

On April 28 Maryland became the first state to prohibit certain differential pricing practices: Effective October 1 a bill will prohibit food retailers and third-party delivery services from using protected class data to offer or price goods in a way that denies consumers equal access to benefits or services (➚HB 0895).

Meanwhile on federal level in the US, the federal Take it Down Act (TiDA) will go into effect on May 19. TiDA will make it illegal to “knowingly publish” or threaten to publish intimate images without a person’s consent, including AI-generated “deepfakes.” (➚Take it Down Act)

Palantir, again

NHS England has granted external staff from Palantir “unlimited access” to identifiable patient data, according to the Financial Times. The change relates to the “National Data Integration Tenant”, described as a “safe haven for data” before it is “pseudonymised”. The change marks a significant departure from the current practice. (➚Financial Times)

The family of one person killed in an April 2025 shooting at Florida State University has filed a federal lawsuit against OpenAI. Plaintiffs allege that the suspected gunman carried out the attack “with input and information provided to him during conversations with ChatGPT” (➚Complaint). It is one of several wrongful death lawsuits filed against OpenAI over the last year.

Growth claims evaluated

Under the title “Measuring up” the Ada Lovelace Institute has released its study evaluating claims about AI and productivity in the UK public sector: “Government communications, industry reports and third-party policy analyses frequently invoke the technology’s potential to cut costs, save time and drive economic growth.”

The study finds that “the bridge between measurement and meaning is unclear”, that “Productivity figures rarely consider costs and trade-offs” and that “Workers and citizens are absent from research design”. (➚Ada Lovelace Institute)

Retracted bits

An influential study from 2025 that claimed OpenAI’s ChatGPT can positively impact student learning has been retracted, nearly one year after publication. The publisher, Springer Nature, “decided to retract this paper owing to concerns regarding discrepancies in the meta-analysis.” (➚Nature) The journal publisher also stated that “the authors had not responded to correspondence regarding the retraction.” The paper had collected hundreds of citations and was often quoted as evidence for ChatGPT’s positive impact.

Science Bits

A study analysing 2.5 million biomedical papers spanning 3 years shows that fabricated references “are embedded in the peer-reviewed literature at scale, and that the rate of fabrication is accelerating.”

Topaz M, Roguin N, Gupta P et al. Fabricated citations: an audit across 2·5 million biomedical papers. The Lancet, 407, 1779–1781

A study investigating the impact of AI assistants and finds that “current AI systems are fundamentally short-sighted collaborators — optimized for providing instant and complete responses, without ever saying no […]”. The study looks at the impact the use of those systems has on performance without them: “Across a variety of tasks […] we find that although AI assistance improves performance in the short-term, people perform significantly worse without AI and are more likely to give up.”

Liu, G., Christian, B., Dumbalska, T., Bakker, M.A., Dubey, R. (2026). AI Assistance Reduces Persistence and Hurts Independent Performance.

Follow and subscribe

If you stumbled across this article on the web, subscribe to the “misaligned bits” newsletter and follow Misaligned on Medium, LinkedIn, Threads or Mastodon.

You can now also subscribe to this newsletter directly with your email.