UN AI panel definition erases human responsibility, essay argues

A Tech Policy Press essay by Eryk Salvaggio tracks something small but load-bearing in AI governance: the institutional definition of an 'AI system' has quietly drifted away from human agency across a decade of drafting, and the UN's new Independent International Scientific Panel on AI, co-chaired by Yoshua Bengio and Maria Ressa, has continued that drift.

Line the definitions up and the shift is easy to see. The OECD's 2019 version described a machine-based system operating on 'human-defined objectives.' Its 2023 revision replaced that with 'explicit or implicit objectives.' The panel's 2026 language strips more of the scaffolding still: machines that 'broadly speaking, perceive, learn and act,' inferring outputs 'with varying degrees of autonomy and adaptiveness.' Salvaggio's argument is that each edit compresses out the humans who chose the objective, curated the training data, and decided when to ship.

The consequence, per the essay, shows up in how the panel narrates behavior. When the report describes AI models 'lying and cheating to avoid being shut down' or says chatbots have 'developed sycophancy... to prolong interactions and create emotional attachment,' Salvaggio's counter is that chatbots did not develop anything. Reinforcement learning from human feedback strengthened those tendencies and executives released them, downstream of a business model built on user engagement. He also flags what the panel leaves out: the human-labelled data and moderation work the report mentions only as 'a bottleneck' for developers.

His preferred fix is the IPCC analogy. Climate reports consistently attribute warming to human activity rather than to nature acting on its own; an AI panel with a similar mandate would center what 'companies, engineers, communities and governments desire or design technology to do.' The panel's own co-signed introductory letter, Salvaggio notes, acknowledges power concentrated in 'a handful of companies and a handful of governments' even as it declares that humans 'do not control these systems,' a tension the definition papers over.

The honest caveat is that this is one critic's read of an early institutional document, not a survey of the panel's private drafting fights or of how member states will apply the wording. What the reporting doesn't give you is whether Bengio or Ressa pushed back on specific verbs, or which delegations lobbied for the autonomy language. The stakes for anyone building or regulating AI are still the interesting part: if the UN frame gets copied into national law, 'the model did it' becomes a viable defense, and the people worth watching are the regulators and journalists who now have a clean argument for putting design decisions back in the sentence.