Human-like output encourages people to infer intelligence and project human qualities onto chatbots.

A world model can represent features, locations, movement, and changes within an environment.

Consciousness, planning, deliberation, and thinking about thinking require separate terms and tests.

Superhuman performance on a narrow task such as chess does not imply broad or human-level general intelligence.

Intelligence evolved to support movement, survival and purposeful interaction with the environment.

Intelligence involves perceiving, deciding and actively manipulating an environment to improve survival and prosperity.

He questions definitions of intelligence centred on task completion and control, proposing that ecological balance and relations with other organisms should also be considered.

He argues that intelligence is embodied and relational and cannot be fully captured as digitised task performance.

Large language models can reorganize existing knowledge but do not think, reason or originate ideas in the human sense.

Generative AI can support research when users verify its output against reliable sources.

Public writing alone does not capture an expert’s unpublished knowledge, voice or position on contested questions.

People and businesses should avoid becoming reliant on current generative AI tools because their prices may rise as subsidies disappear.

Society tolerates human mistakes more readily than machine mistakes, making rare AI failures a major adoption barrier.

Investment in transformers is driven by impressive public utility, repeated scaling gains and uncertainty about where improvement will stop.

Repeated dependence on generative AI may reduce a person’s ability to perform delegated tasks unaided.

Healthy cognition requires continued intake of novelty alongside stored patterns and error correction.

Generative AI can replace productive struggle in research and writing, creating dependence and weakening independent thought.

ChatGPT may expose existing problems in higher education, including workload, cost and weak alignment between assignments and learning goals.

Students who believe ChatGPT is smarter than they are may weaken their own creative development.

Young people can learn fast-moving technical subjects effectively by building small projects and using online resources as needed.

Enterprises are actively reconsidering hiring, associate programs, training and how to use AI without weakening employee development.

Conflict will grow where AI causes job loss, reduces opportunity or devalues skills, especially for people entering the workforce.

California should adopt universal basic income in response to AI automation and use technology policy to renew social and economic progress.

Economic privilege makes reflective transformation easier, while poverty and success can each obstruct it in different ways.

Creativity consists of novelty plus value.

This Climate Does Not Exist uses generated local imagery with the aim of reducing psychological distance from climate impacts.

Digital media can transmit cultural outputs while omitting the embodied experience and context needed to understand how they were produced.

Creative-AI harms involve both unauthorized training data and systems intentionally designed to replace human artists.

People may disclose quickly to chatbots because they expect less judgment than in a human relationship.

As machines become more intelligent, humans must become wiser by relying on embodiment, relationships, nature and emotion.

The internet has shifted from a place people visited to an ambient physical-digital environment embedded throughout culture.

Current evidence cannot establish whether generative AI causes psychosis or makes existing vulnerability easier to express.

Self-development depends on self-awareness, while emotional education often stops before people learn how to regulate and examine emotions.

His gaming concern centres on children being exploited or groomed within platforms such as Roblox.

Some people can become drawn into AI-mediated beliefs and alternate realities in a way that resembles cult dynamics.

Strong group attachment can turn into righteous anger or violence when identity-defining values feel betrayed.

AI dependence can form when it makes disliked tasks easier and repeated successful outcomes reinforce avoidance of effort.

AI should help people implement their values and make decisions, with explanation and auditability supporting human oversight.

He argues that accepting some friction and consciously choosing technology can restore individual agency and change one’s relationship with it.

Humility, doubt and willingness to seek exceptions allow people to revise deeply held beliefs.

High-dimensional optimization could extend human decision-making beyond the number of variables people can consciously manage.

Disruption can lead through reflection and transformation to the evolution of institutions including family, education, politics, economics and religion.

Transhumanism is a social movement that uses science and technology to radically improve humans and human experience.

Privacy can be collective as well as individual, especially when systems classify or expose a community.

Privacy includes inferences made about a person, even when the original data is not directly disclosed.

He sees access to and control of large amounts of personal data as central to current technology power and associated harms.

Expert-specific AI should be built with permission and tied to an identifiable human source.

People affected by an AI system should have a say in the trade-offs built into it.

Researchers may create the ideas, but companies and governments usually control how those ideas are used.

AI rules can be set by governments or companies; the unresolved question is who gets to decide and who is accountable.

Regulation compels behaviour, policy identifies objectives and governance determines who makes and enforces decisions.

Restrictions on research ideas would be difficult to enforce because ideas can be modified and reused in ways that are hard to detect.

He argues that AI resembles heavy industry because additional use requires ongoing compute and resource inputs, and asserts that scaling creates diseconomies.

Open-source research can accelerate development, invite scrutiny and distribute the benefits of a new AI architecture.

Evolutionary systems exploit loopholes in objectives and environments, so promising solutions should be tested in simulations before real deployment.

Reward-maximizing agents may seek control of their reward mechanism, so safety research must constrain unintended paths or use alternative goal systems.