Former Gemini Data Decision-Maker Calls for AI That Questions Itself

Quick Info

  • Essay: “The Paradox of AI Safety” by Dr. Jennifer Prendki
  • Published: 18 November 2026, in The AI Journal
  • Background: Former Google DeepMind — data decisions for Gemini
  • Her company: Dyssonance — a common-sense reasoning engine
  • Core idea: AI should understand better, not just remember more

A new essay from a researcher who once helped decide what data went into Google’s Gemini argues that the AI industry is chasing the wrong goal. Instead of systems that remember more, she says, the field needs systems that understand better.

In a column for The AI Journal, Dr. Jennifer Prendki, founder of Dyssonance, describes the path that took her from physics to Google DeepMind and then out of the lab altogether. The piece, published on 18 November 2026 under the title “The Paradox of AI Safety”, combines a personal account with a manifesto for her new company.

Prendki writes that as a child she wanted to understand the universe. That curiosity led her to physics, and later to artificial intelligence, which offered a similar intellectual challenge along with an unresolved question about the nature of intelligence itself. Over time, though, what kept her in the field was a growing unease about how AI systems were being built.

At Google DeepMind she worked among leading researchers and carried responsibility for sensitive decisions about the data used to train Gemini. The weight of those decisions sometimes kept her awake at night. She was struck, she says, that many colleagues concentrated on improving models without stopping to ask what exactly was being built or what would happen once people began using it.

She stayed because she hoped that working from the inside would let her make a difference, even without any guarantee that her concerns would be heard. As that hope faded, she found herself feeling responsible for outcomes she could no longer meaningfully influence. She left, while admitting a fear that she might one day watch the consequences from the outside and regret her decision.

See Also  The Role of Identity Lifecycle in Preventing Data Breaches and Insider Threats
AI researcher reflecting on how large language models learn from data

Once outside, Prendki was free to question the assumptions she had worked under. Drawing on firsthand knowledge of how large language models learn from data, she chose not to build a better LLM or tools to patch the shortcomings of existing ones. She decided to go back to first principles, and that choice led to Dyssonance.

The company’s manifesto starts from a blunt observation: today’s AI is good at remembering information, but remembering is not the same as understanding. Prendki names ChatGPT, Claude and Gemini as examples of systems that perform impressively yet still struggle with conflicting information, with recognising their own errors and with continuously updating what they understand. She expects the next breakthrough to come from teaching machines to handle ambiguous situations, rather than from more data or bigger memory.

Common-sense reasoning engine concept — interconnected ideas resolving contradictions

Dyssonance is working on what it calls a common-sense reasoning engine, a system meant to develop and refine its own understanding, with interpretations that shift as new evidence appears. Contradictions play a central role in this approach. Prendki compares the process to the work of scientists, detectives and entrepreneurs, who all make progress by noticing what does not add up. AI, she argues, should actively identify and resolve such contradictions instead of ignoring them or simply piling up more information.

The longer-term vision, which the essay calls engineering intuition, is a shared foundation for reasoning, learning and adaptation that different capabilities can build on, rather than one system that attempts everything. Prendki expects this to make AI more adaptable, its reasoning easier to evaluate and its capabilities easier to improve and combine.

Her conclusion sums up the argument in a sentence: the future of AI is not about remembering more, but about understanding better.

Similar Posts