Home
Aug 15, 2026
View All

Notion Simplifies Model Selection; Simon Willison on Hallucination as a Feature

·1 underrepresented voice

The Short Version#

Notion quietly shipped model selection simplification while Simon Willison reframes hallucination as a deliberate design pattern rather than a bug — together they surface a consistent tension PMs are navigating right now: how much model complexity should users actually see?

Notion — Model Selection, Simplified#

Source: https://www.notion.so/releases/2026-08-14 Credibility: High (first-party changelog)

What happened: Notion shipped a release on August 14 called "Model selection, simplified." The framing is notable. Most AI product teams have been trending toward giving users more model choice. Notion is moving in the other direction — simplifying or abstracting the model picker, presumably because most users don't want to make that call themselves.

Key product patterns:

  • Abstraction over control — removing or collapsing model selection rather than expanding it
  • Signals a maturing assumption: users want outcomes, not model names
  • Pairs with the ZeroEntropy acquisition (efficient, task-specific models for knowledge work) announced recently — Notion is building toward a model layer they control, which makes opinionated model selection more defensible

Why it matters for PMs: This is a direct stake in the ground on one of the most contested UI questions in AI product design right now. Do you expose model choice to users, or do you hide it and optimize behind the scenes? Notion is betting on "hide it." The argument is solid: most users don't know what GPT-4o versus Claude Sonnet actually means for their use case — and asking them to choose creates friction without value. The risk is that power users who care about model performance feel locked out. Watch for whether Notion adds any advanced toggle or whether this is a hard abstraction.

Critical questions:

  • How does Notion handle cases where users explicitly want a specific model for compliance, output style, or cost reasons?
  • Is "simplified" a removal of the picker, or a smarter default that still allows override?
  • Does this change the value prop for Notion AI at the enterprise tier, where IT buyers often want auditability over which model processed their data?
  • How does Notion plan to communicate model changes to users who've come to expect consistency in AI output quality?

Action you could take today: If your product exposes model selection to users, pull your analytics on how often that picker actually gets used. If it's under 10-15% of sessions, you probably have a "simplification" opportunity sitting in front of you.

Simon Willison — "Don't Classify. Hallucinate!"#

Source: https://simonwillison.net/2026/Aug/14/dont-classify-hallucinate/ Credibility: High (Simon Willison is a recognized practitioner; this is his direct analysis blog)

What happened: Simon Willison published a post reframing LLM "hallucination" not as a failure mode but as a deliberate and sometimes useful feature. The title is the argument: in certain contexts — specifically creative or generative ones — you want the model to confabulate rather than classify from a fixed set of options.

Key patterns from the post:

  • The classification vs. generation distinction is more important than the accuracy vs. hallucination distinction in many product contexts
  • When you ask a model to pick from a fixed taxonomy, forcing it into a classifier mode can suppress useful generative output
  • "Hallucination" is a failure when you need factual retrieval; it's a feature when you need synthesis, creative output, or novel framing
  • The framing implies that prompting strategy — not model choice — is the primary lever for controlling this behavior

Why it matters for PMs: Most product teams treat hallucination as the enemy and build elaborate RAG pipelines, citation systems, and grounding layers to defeat it. Willison's point is that this is the right call for some use cases and the wrong call for others. If you're building a creative writing tool, a brainstorming assistant, or any feature where novelty is the value — you might be fighting the wrong battle. More practically: it reframes how you spec AI features. "Accuracy" is the wrong success metric for generative tasks. "Useful output" is the right one, and those require different evals.

Critical questions:

  • How do you communicate to users the difference between an AI feature that's meant to generate versus one that's meant to retrieve? Most users don't distinguish.
  • If you lean into hallucination for creative tasks, what guardrails prevent it from bleeding into factual contexts in the same product?
  • Does this change how you'd write acceptance criteria for AI features in a spec?
  • What's the eval framework for "useful hallucination" — how do you know it's working?

Action you could take today: Look at your current AI features and explicitly categorize each one as "retrieval" or "generation" tasks. For each generation task, check whether your evals are measuring accuracy (probably wrong) or usefulness/novelty (probably right). Adjust your success metrics accordingly.

Fei-Fei Li — AI in Schools: Agency, Not Cheating, Is the Real Stakes#

Source: https://hubermanlab.com (episode featuring Fei-Fei Li, released ~August 11, 2026) Credibility: Medium-High (Fei-Fei Li is a foundational AI researcher and Stanford professor; podcast appearance, not a paper or product announcement)

What happened: Fei-Fei Li appeared on the Huberman Lab podcast and made a pointed claim about AI in education: the worst outcome isn't cheating. It's students losing agency and intrinsic motivation. The distinction matters. Most institutional responses to AI in schools are anti-cheating frameworks. Li is arguing the threat model is wrong.

Key patterns:

  • Reframes the AI-in-education problem from "integrity" to "agency" — a fundamentally different product design and policy constraint
  • The concern is that over-reliance on AI collapses the productive struggle that builds competence and confidence
  • Li's framing implies that good AI education tools should preserve the user's sense of authorship and effort, not eliminate it
  • This is consistent with a broader pattern in human-AI collaboration research: the more invisible and seamless AI assistance is, the more it may undermine the cognitive work that drives skill development

Why it matters for PMs: This applies well beyond education. Any product where users are trying to learn, grow, or develop competence faces the same design tension. Copilot-style autocomplete might make developers faster in the short term while making them worse engineers over time. Writing assistants might improve output quality while atrophying the user's ability to think through arguments. The PM question is: what is the user actually trying to get better at, and does our AI feature support that or short-circuit it?

Critical questions:

  • How does your product define "success" for AI features — task completion speed, or user capability growth?
  • Are you building for engagement with your AI feature, or for outcomes that reduce the user's need to use AI?
  • In an education or learning context, how would you design AI assistance that preserves productive struggle?
  • Does your onboarding or feature design make users more capable over time, or more dependent?

Action you could take today: Pick one AI feature in your product and write a two-sentence description of what the user is trying to get better at by using it. Then ask whether the feature helps them improve or replaces the need to improve. If it's the latter, that's a design question worth surfacing to your team.

Quick Hits#

The Thread#

The abstraction question is getting louder. Notion simplifying model selection, Willison arguing hallucination is a feature in the right context, Fei-Fei Li warning about AI eroding user agency — all three pull at the same thread: how much of what's happening under the hood should users understand, control, or even know about? The answers are different depending on whether you're optimizing for task completion, user capability, or trust. But the question is no longer optional for product teams. You have to have an answer.

Sit With This#

Notion shipped "model selection, simplified" — moving away from exposing model choice to users, betting that abstraction over control is the right call for most of their audience.

For your product: If you stripped out every visible AI control — model pickers, temperature sliders, prompt editors, "try again with a different model" buttons — would your users miss them? Would power users leave? And if you're not sure, what would it take to find out?