De Kai on “Raising” AI and the Choices Behind Algorithms

Shira Lazar welcomes De Kai, an AI researcher and author of Raising AI, to discuss technology, culture and responsibility. She introduces his work in machine translation and his interest in mindful approaches to AI.

Asked why he calls AI systems “artificial children,” De Kai says they learn patterns from human examples and can seek approval in ways that make a parenting metaphor useful to him. He describes current systems as capable in some respects but immature in others. This is an analogy, not a claim that software is literally a child.

He argues that people should not rely solely on regulators or company guardrails. In his framing, human behavior and the material used to train systems help shape what those systems reproduce. Lazar responds that the metaphor encourages people to consider how they act in communities.

De Kai recalls beginning AI research at Berkeley in the 1980s. He says his early work questioned whether hand-written logical rules could handle the complexity of language, and argued for associative models that learn patterns.

He contrasts fast, intuitive pattern recognition with slower conscious reasoning, using language processing and learning chess as examples. He sees modern neural networks as especially strong at the former, while warning that fluent pattern completion can produce exaggerated or incorrect answers.

He describes emerging reasoning systems as early and still inefficient. Looking ahead, he predicts smaller and more efficient AI architectures and argues that current transformer-based systems are not the only possible approach. These are his forecasts, not established timelines.

Lazar asks what it will be like to grow up in a rapidly changing AI era. De Kai says the pace makes it difficult to advise children about future work and predicts major changes for technical and design occupations. His employment predictions are presented as his view.

He emphasizes that AI is a field of inquiry rather than one specific model, comparing its many approaches to different theories in physics. He expects architectures to change over time and questions assumptions that today’s infrastructure needs will remain fixed.

The conversation turns to a reported death after interaction with an AI therapy tool. De Kai predicts AI may eventually become useful in therapy, but says current systems are not sufficiently reliable for sensitive counseling. He discusses human oversight, conservative deployment and prominent warnings as possible safeguards.

Asked about fear of AI, he distinguishes fictional narratives about powerful machines from risks he sees in current systems. He argues that recommendation algorithms influence people every day through feeds, search results and other selections.

De Kai calls attention to the other side of selection: what an algorithm does not show. He calls this “algorithmic censorship” as a description of filtering, while acknowledging that people need help navigating enormous volumes of information. He wants society to ask what criteria should guide these choices.

He worries that filtering can make it harder for people to understand one another. At the same time, he points to machine translation as an example of technology that can help people communicate across languages, depending on how it is deployed.

Lazar raises a then-recent Google estimate of energy and water use for an average Gemini query. De Kai supports transparency about resource use and reiterates his view that current models are inefficient compared with human learning. The interview’s numerical comparisons are their claims from 2025 and do not establish a universal footprint for AI use.

The discussion moves to a reported study of disappointing returns from corporate AI projects. De Kai argues that a mandate to “use AI” can lead teams to adopt tools without defining a costly problem or considering organizational effects. He says expectations can be both overly optimistic and too limited.

Discussing his book, De Kai says AI research also prompts questions about human cognition: what models are doing, what they are missing and how that compares with human thought. He questions whether fluent output necessarily captures a genuinely new insight.

He advocates broad education combining science, engineering, art and philosophy, arguing that understanding ourselves matters in an AI era. He returns to his parenting metaphor: learning about systems can also teach people about themselves.

Lazar closes with a personal reflection on switching from an AI journal to pen and paper and making quiet space to think. She thanks De Kai for a conversation that invites listeners to consider their own choices.