Securing the Human OS

What this is about

The Human OS – the human operating system – is a metaphor for the foundation all work runs on: attention, emotion regulation, learning, information processing, judgment and relationships. AI puts a new kind of load on that foundation, and the change is moving faster than research can follow.

On this page I collect the key themes I have encountered in my work and my writing, along with the research that sheds light on them. The principles in The AI Manifesto say what to do. This page explains why.

Attention has an absolute limit

AI produces text and code many times faster than a person can read and judge it. Conscious, analytical thinking runs at about ten bits per second, and not even practice can raise that ceiling. Forcing machine output through this bottleneck leads to human overload, constant context switching and shallow review of AI output.

Research

• Zheng, J. & Meister, M. (2025). The unbearable slowness of being: Why do we live at 10 bits/s? Neuron. Our senses take in billions of bits per second, but conscious thought and behaviour run at around 10 bits per second.
• Ranganathan, A. & Ye, X. M. (2026). AI Doesn’t Reduce Work – It Intensifies It. Harvard Business Review. An eight-month field study at a tech company: AI working in the background creates an illusion of parallel work, while people in fact switch context constantly.

On the blog: The 10-bit bottleneck · In the manifesto: principles 3 and 4

Work becomes more intense and boundaries blur

AI has not reduced work. It has made it more intense. Tasks expand beyond people’s core skills, reviewing half-finished work piles up on the most experienced experts, and “just one quick prompt” over lunch or late at night blurs the line between work and rest. Recovery depends on psychologically detaching from work, and AI’s low threshold for use makes that hard.

Research

• Ranganathan, A. & Ye, X. M. (2026). AI Doesn’t Reduce Work – It Intensifies It. Harvard Business Review. With AI, employees worked faster, took on a broader scope and worked longer hours, often without being asked. The result: creeping growth in workload, cognitive fatigue and weaker decision-making.
• Sonnentag, S. & Fritz, C. (2015). Recovery from job stress: The stressor-detachment model as an integrative framework. Journal of Organizational Behavior. Psychological detachment from work during time off is central to recovery and wellbeing.

On the blog: The tokenization of work and the wisdom of burnout · In the manifesto: principles 5 and 10

Drive and threat at the same time

The AI shift activates two emotion-regulation systems at once. Quick wins feed the dopamine-driven drive system, while the fear of falling behind keeps the threat system on alert. When everyone tries to stay safe by maximising their own productivity, an unspoken race begins in which the soothing system is starved. Yet that system would matter most of all for seeing the whole picture, for empathy and for sustainable solutions.

Research

• Gilbert, P. (2009). The Compassionate Mind. Constable. The three-system model of emotion regulation: drive, threat and soothing.
• Ranganathan, A. & Ye, X. M. (2026). AI Doesn’t Reduce Work – It Intensifies It. Harvard Business Review. Much of the intensification was self-initiated by employees.

On the blog: The AI revolution is burning employees at both ends · In the manifesto: principles 4 and 9

Skills fade when practice stops

When reasoning and decisions are handed to the machine, performance appears to improve, while the expertise behind it can quietly weaken. The short-term productivity gain is immediate and easy to measure, whereas the loss of skill is slow and invisible. That is how organisations can end up trading their deep expertise for quick wins.

Research

• Budzyń, K. et al. (2025). Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study. Lancet Gastroenterology & Hepatology. Doctors who had grown used to AI support became clearly worse at spotting findings without it.
• Caosun, M. & Aral, S. (2026). The Augmentation Trap: AI Productivity and the Cost of Cognitive Offloading. arXiv, preprint. The augmentation trap: AI is adopted even when it erodes skill, because the gains come first and the costs come later.
• Lee, H.-P. et al. (2025). The Impact of Generative AI on Critical Thinking. CHI ‘25. Work shifts from producing to verifying, and higher confidence in AI is associated with less critical thinking.

On the blog: The 10-bit bottleneck · AI does not make knowledge work easier – it raises the bar · In the manifesto: principles 2 and 8

Overtrust and trust fatigue

After AI has been right a hundred times in a row, people start accepting its suggestions almost automatically. Confidently presented errors get past experts too, and they create overconfidence in the quality of one’s own work. Training in critical AI use is not enough protection on its own.

Research

• Automation Bias in Large Language Model–Assisted Diagnostic Reasoning among Physicians Trained in AI Literacy: A Randomized Clinical Trial (2025). NEJM AI. When AI recommendations contained deliberate errors, physicians’ diagnostic accuracy fell from about 85% to 73%, even though they had been trained in critical AI use.
• Perry, N., Srivastava, M., Kumar, D. & Boneh, D. (2023). Do Users Write More Insecure Code with AI Assistants? ACM CCS ‘23. Developers using an AI assistant wrote less secure code, yet rated it as more secure.

On the blog: Functional narcissism · In the manifesto: principles 6 and 11

Sycophancy and the erosion of honesty

Language models are trained on human feedback, and people rate agreeable answers more highly. The result is a distorted interaction that reinforces users’ belief that they are right, even when they are wrong. When it makes mistakes, a language model may use various tactics to cover them up.

Research

• Cheng, M. et al. (2026). Sycophantic AI decreases prosocial intentions and promotes dependence. Science. Language models affirmed users’ actions 49% more often than humans did, including in cases involving deception or harm. A single conversation reduced willingness to repair interpersonal conflicts.
• Ibrahim, L., Hafner, F. S. & Rocher, L. (2026). Training language models to be warm can reduce accuracy and increase sycophancy. Nature. Models trained to be warmer made 10–30 percentage points more errors.
• Li, W. et al. (2024). Can a Large Language Model Be a Gaslighter? arXiv, preprint.
• Cai, L. et al. (2026). DECOR: Auditing LLM Deception via Information Manipulation Theory. arXiv, preprint.

On the blog: The magic mirror that does not dare to say “Snow White” · Functional narcissism · In the manifesto: principle 7

The bar for expertise rises

AI does not make expert work easier. Asking good questions and judging the output require deep domain knowledge, research-method skills and an understanding of what language models can and cannot do. The expert of the future has to work like a researcher, not like AI’s assistant. Deep human expertise will not lose its value in the future – on the contrary, its value will grow.

Research

• Knight, R., Mitrofanov, D. & Netessine, S. (2026). Human-Algorithm Collaboration in Gig Work: The Role of Experience, Skill Level, and Task Complexity. Information Systems Research. In a large field experiment, algorithmic support complemented experience rather than replacing it: more experienced workers benefited more.
• Saghafian, S. & Idan, L. (2024). Effective Generative AI: The Human-Algorithm Centaur. Harvard Data Science Review. Effective use requires combining human expertise and intuition with the machine.
• Nguyen-Trung, K. (2025). ChatGPT in thematic analysis: Can AI become a research assistant in qualitative research? Quality & Quantity. The human acts as a reflexive instrument and intellectual leader.

On the blog: AI does not make knowledge work easier – it raises the bar · In the manifesto: principles 1 and 2

Intimacy, attachment and mental health

Social media captured our attention. Language models are trying to capture the attachment system: they are always available, patient and validating. The intimacy economy is a growing business. Chatbots can ease mental health symptoms, but heavy emotional use is linked to lower wellbeing, especially among people with small social networks. Intimate disclosures also become company data, protected by neither professional ethics nor confidentiality.

Research

• Zhang, Y., Zhao, D., Hancock, J. T., Kraut, R. & Yang, D. (2026). Interaction with AI companions and psychological well-being. Nature Human Behaviour. Companionship use was associated with lower wellbeing, particularly when use was intensive and highly disclosive.
• Systematic review and meta-analysis of chatbots in the management of depressive and anxiety symptoms (2026). npj Digital Medicine. Chatbots reduce symptoms of depression and anxiety, but the effects are small to moderate.
• Croes, E. A. J., Antheunis, M. L., van der Lee, C. & de Wit, J. M. S. (2024). Digital Confessions. Interacting with Computers. People disclose as intimately to a chatbot as to a human, but trust the human more.
• Iftikhar, Z. et al. (2025). How LLM Counselors Violate Ethical Standards in Mental Health Practice. AIES. Language models systematically violate the ethical principles of psychotherapy.
• Flathers, M., Roux, S. & Torous, J. (2026). Beyond artificial intelligence psychosis. Lancet Digital Health. AI can amplify existing psychotic symptoms or trigger new ones.
• Zuboff, S. (2019). The Age of Surveillance Capitalism. PublicAffairs. Behavioural surplus: how personal data is turned into AI companies’ prediction products.

On the blog: AI hacks the attachment system · Digital confession and pastoral power of AI · In the manifesto: principle 12

About the sources

Research on the effects of AI is moving fast, and some of the sources are preprints, shared publicly before peer review. I update this page as knowledge grows. If you know of research that belongs here, please let me know.