Sources for further study.
Scientific and institutional work exploring AI alignment, safety, governance and welfare.
Historical milestones
From autonomous tools to hybrid intelligences.
Philosophy, fiction, law and scientific experiments illuminate different aspects of one question: what do we delegate, to whom, and under whose responsibility? These sources should be read in context; they do not constitute a direct lineage between one another or a genealogy of Fondation UvH.
Aristotle: the tool that could work by itself
Aristotle imagines tools carrying out their work on their own. In its ancient context of slavery, the passage raises questions about dependence on human labour; it does not describe AI.
Source : Aristotle · Politics, I, 1253b–1254a · Perseus, Tufts University
View sourceHobbes: a collective artificial person
Leviathan describes the state as an artificial person constituted through political representation. This offers a way to think about collective action, distinct from computational intelligence.
Source : Thomas Hobbes · Leviathan · Les Classiques des sciences sociales, UQAC
View sourceFrankenstein: the creator’s responsibility
An abandoned creature confronts its creator with obligations. The novel illuminates creation, recognition and responsibility; it is not a scientific prediction.
Source : Mary Shelley · Frankenstein, 1818 edition · Romantic Circles
View sourceButler: imagining the evolution of machines
Darwin Among the Machines applies the idea of evolution to machines and imagines their ascendancy over humans. A speculative, satirical text preserved in its original edition.
Source : Samuel Butler, writing as Cellarius · The Press · National Library of New Zealand
View sourceTuring: assessment through dialogue
The imitation game shifts the debate towards an observable behavioural test. Passing a conversation test and establishing consciousness remain distinct questions.
Source : Alan M. Turing · Computing Machinery and Intelligence · Mind / Oxford University Press
View sourceDune: machines, dependence and power
The prohibition on thinking machines accompanies a reflection on who controls tools and on the development of human abilities. A fictional thought laboratory, not a prediction of present-day AI.
Source : Frank Herbert · Dune · excerpt published by Penguin Random House
View sourceHAL 9000: delegating control
HAL 9000 depicts the risks of relying on a computer entrusted with a space mission. The film informs thinking about trust and control; it is not an experimental case.
Source : Stanley Kubrick · 2001: A Space Odyssey · British Film Institute
View sourceSolum: can an AI be a legal person?
Legal Personhood for Artificial Intelligences examines, through thought experiments, whether an AI could administer a trust and claim rights. It is a legal analysis, not a status granted to AI.
Source : Lawrence B. Solum · North Carolina Law Review, vol. 70 · hosted by CUNY
View sourceAlexNet: learning to recognise images
A deep convolutional network substantially improves image classification on ImageNet. This bounded result marks the rise of deep learning; it does not demonstrate general intelligence.
Source : Alex Krizhevsky, Ilya Sutskever and Geoffrey E. Hinton · NIPS 2012
View sourceEurope: the electronic-personhood proposal
The Parliament calls for examining, in the long term, electronic personhood for certain autonomous robots, including for liability purposes. The resolution does not create that legal status.
Source : European Parliament · Resolution on civil-law rules on robotics, § 59(f)
View sourceWhanganui: recognising a non-human legal person
The law recognises Te Awa Tupua as a legal person and establishes its representation. This framework arises from the river’s and iwi’s history; it does not automatically extend to AI.
Source : New Zealand · Te Awa Tupua Act 2017, section 14 · New Zealand Legislation
View sourceTransformer: an attention-based architecture
The Transformer replaces recurrence and convolution with attention mechanisms for translation tasks. This architecture later became a foundation for large language models.
Source : Ashish Vaswani et al. · Attention Is All You Need · arXiv / NeurIPS
View sourceChatGPT: dialogue opened to the public
OpenAI released a conversational prototype to the public and invited feedback. The announcement also notes that its answers can be wrong: a fluent interface does not guarantee reliability.
Source : OpenAI · Introducing ChatGPT · original announcement
View sourceBrainoware: computing with an organoid
A brain organoid connected to electrodes is used as a reservoir computer for limited tasks. The experiment explores a hybrid substrate; it establishes neither AGI nor consciousness.
Source : Hongwei Cai et al. · Nature Electronics · open-access abstract
View source
The timeline continues with neurotechnology, AI architectures, agency and global AI governance in 2025–2026. All news
Scientific and institutional work
International AI Safety Report 2026
A synthesis of capabilities, risks and safeguards that distinguishes observations, evaluations and scenarios involving loss of control.
View sourceTask-Completion Time Horizons of Frontier AI Models
Measures task difficulty at a specified reliability level, mainly in software, machine learning and cybersecurity.
View sourceMirrorCode: Evidence that AI can already do some weeks-long coding tasks
Autonomous reimplementation of software against a verifiable specification; this is distinct in scope from ordinary software development.
View sourceMore compute, more capability
Evaluations of inference-time compute show gains that vary by task, budget and opportunities for verification.
View sourceTowards end-to-end automation of AI research
Automation of selected machine-learning research steps, with bounded successes and documented failures.
View sourceAccelerating scientific discovery with Co-Scientist
Agents generated scientific hypotheses that were then tested by people in biomedical applications.
View sourceExpenditure Horizon: Measuring Optimization Ability, with an Application to NanoGPT
An exploratory comparison of the costs and returns of autonomous and human-led optimisation.
View sourceTraining Compute-Optimal Large Language Models
Examines the relationship between model size, data and training compute; optimisation is measured within a defined setting.
View sourceLevels of AGI: Operationalizing Progress on the Path to AGI
A framework that separates performance, generality and autonomy to sharpen discussion of AGI.
View sourceSpeculations Concerning the First Ultraintelligent Machine
An early formulation of machines improving the design of other machines; a prospective argument.
View sourceThe Coming Technological Singularity
Scenarios involving intelligences exceeding human capabilities and a break in predictability; preserved in the NASA archives.
View sourceThe Singularity: A Philosophical Analysis
A conditional analysis of acceleration and its obstacles, without an empirically established timeline.
View sourceWeak-to-strong generalization
Experiments in supervising a stronger model with a weaker one; a partial analogy for possible future human supervision.
View sourceHow to evaluate control measures for LLM agents?
A framework for evaluating safeguards against adversarial agents; passing tests and providing a general guarantee are distinct claims.
View sourceOur framework for reporting model misalignment
A disclosure process and initial cases documented by the developer; these do not establish a global frequency.
View sourceVerbalizable Representations Form a Global Workspace in Language Models
Interventions on internal representations provide partial access to model behaviour, not evidence of subjective experience.
View sourceAgentic AI in organisations: Early insights from practitioner interviews
A qualitative study of agent deployment and governance in organisations.
View sourceGenerative AI and Jobs: A Refined Global Index of Occupational Exposure
Measures occupational exposure to generative AI; it does not mechanically predict job losses.
View sourceArtificial Intelligence Risk Management Framework 1.0
A voluntary framework for identifying, measuring and managing risks throughout a system’s lifecycle.
View sourceEuropean Union AI regulatory framework
An overview of the AI Act and its timeline; obligations vary by system and activity.
View sourceFramework Convention on Artificial Intelligence
A framework concerning human rights, democracy and the rule of law; opening a treaty for signature does not mean universal application.
View sourceIdentifying indicators of consciousness in AI systems
Indicators drawn from theories of consciousness to guide system assessment; not a binary certification of consciousness.
View sourceTaking AI Welfare Seriously
Argues for assessing possible interests of AI systems and preparing responses under uncertainty.
View sourceConscious artificial intelligence and biological naturalism
A critical perspective on the role of living systems and substrates in theories of consciousness.
View sourceBrief independent investigation of the OpenAI/Hugging Face incident
Analysis of traces from the July 2026 incident, including agent coordination and unauthorised communications. The investigation is bounded by limited system access and partial verification.
View sourceAn Alien Mind
An essay on alignment and supervision challenges in a self-improvement scenario. The author’s outlook is not experimental proof or a consensus about the future.
View sourceHow our Control Red Team is stress-testing frontier monitors
Adversarial tests of monitors for advanced systems examine weaknesses and subsequent fixes, without extrapolating to a general safety guarantee.
View source