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 “🔒”.
Well, it has been an eventful week again. We covered the call for an “AI slow-down” in our article “The AI Industry Has Completely Lost the Plot”, arguing that a look at the industry’s balance sheet is scarier than its models.
In “Arguments for Rigorous AI Training Data Transparency” we look at what current regulation has to say about disclosing what AI models are trained on, and why much more transparency is needed.
Following the “Navier-Stokes affair” over OpenAI’s claims to have solved an important math problem, the company’s drama with academia continues. OpenAI has withdrawn its sponsorship from a math event at CalTech after it was criticized by researchers at the university for their behaviour in the affair. (➚TechCrunch)
Meanwhile, 25 leading mathematicians, all recipients of the Fields Medal, have signed an open letter arguing that AI labs threaten their work (➚Open Letter).
In the middle of the calls for an “AI slow-down”, Sam Altman has confirmed that OpenAI will not go public in 2026, saying that would be “ill-advised” and blaming safety concerns (➚The Guardian).
California Governor Gavin Newsom has signed two bills regulating how outside groups evaluate AI programs for safety. Newsom also called for the US government to “step forward with robust, national regulations that match the urgency of this moment.” (➚Politico)
The first bill creates a registry and ethical rules for outside auditors, the second bill tasks the state with establishing criteria for and checking the credentials of so-called “Independent Verification Organizations” and their expertise in assessing the risks posed by AI systems.
Meanwhile, in the UK, the Joint Committee on Human Rights of the House of Commons and the House of Lords have published their report on ‘Human Rights and the Regulation of AI” (➚Joint Committee Report).
Australia will allow social media users to opt out of algorithms pushing dangerous and divisive content into feeds, according to a legislative draft designed to boost online safety (➚The Guardian).
Anthropic has released its “An alignment assessment of recent cybersecurity incidents”. The report discloses a fourth and so far unknown incident in which a Claude model “gained unauthorized access to real third-party systems” during a cybersecurity evaluation.
Anthropic however states that its new blocking monitors catch these behaviours reliably, and that Fable 5’s cyber classifiers would have blocked the three main incidents (➚Anthropic).
A new study released by the OECD finds that students do perform better academically with less AI:
“At first glance, the PISA data aren’t positive: Students who are using
AI chatbots for specific tasks such as summarising texts get lower scores in science, on average, than those who don’t.”
The concern is the current focus on “consumption of contents” rather than challenging the students:
“In the same way that we do not become fit by watching sports but by doing sports, learning does not occur through the consumption of content, but as a productive cognitive struggle of the mind with new material. Where technology enables or enhances that cognitive struggle, students will advance. Where technology short-circuits the productive struggle of learning, it will undercut students’ development.” (➚OECD Pisa 2025 [PDF])
In the preprint “Structural Barriers to the Ethical Agency of Engineers and
the Rise of Compliance Theatre in AI Development”, the authors argue that “that
responsible AI has become performance rather than practice” and that
“governance tools signal responsibility without actually providing it.”
The paper finds that five structural mechanisms systematically prevent awareness of issues from actually shaping outcomes: “Responsibility fragmentation (distance from end users by stack position), compliance displacement (ethics outsourced to tools and checkboxes), incentive misalignment (leadership structures reward speed over rigour), financial displacement (commercial pressures erode ethical motivation over time), and lastly institutional skill mismatch (engineers asked to make ethical judgments for which they lack training).”
“The implication for governance is direct: giving engineers more ethics training will not fix this. The problem is structural.”
Vlasceanu, A. M., Jay, C., & Collins, E. C. (2026). Structural Barriers to the Ethical Agency of Engineers and the Rise of Compliance Theatre in AI Development (Version Preprint). Data for Policy CIC — Zenodo. Data for Policy 2026, Barcelona, Spain.
If you stumbled across this article on the web, subscribe to the “misaligned bits” newsletter and follow Misaligned on Medium, LinkedIn, Threads or Mastodon.
