Co-Existence and the End of Co-Intelligence

Ethan Mollick · article · 2026-06-04 source ↗ Reader 🖥️

practical · mixed — A personal essay / book announcement that carries an argumentative core (the co-intelligence -> co-existence periodization, backed by cited statistics) inside a narrative frame (how he wrote the book and built its website). Ultimately practical: it is building toward 'here is how to relate to AI now,' not just describing a state of affairs.

Structure

Key terms

• Co-Intelligence
  The 2023-era mode of AI use: a cooperative, human-centered exercise of back-and-forth chatbot prompting where the human supplies knowledge and skepticism and the AI is a helper.
• Co-Existence
  The emerging mode of AI use: working with an AI that is 'sometimes, but not always, better than you' — not a fixed skill but a relationship continually renegotiated as models improve, including tasks where AI now operates with much less human involvement.
• jagged frontier Jagged Capability (On Rails / Off Rails)
  The uneven, hard-to-predict pattern of AI competence — excellent at some tasks, poor at others in ways that don't map onto human intuitions about difficulty — invoked here to explain why autonomy hasn't simply replaced human judgment everywhere.
• hand over the keys
  Full delegation of a task to AI with minimal human involvement or review, as opposed to iterative co-authorship.

Propositions

• explicitAI companies' long-term goal has always been to build 'highly autonomous systems that outperform humans at most economically valuable work' (quoting OpenAI's charter) rather than the cooperative, chatbot-era co-intelligence Mollick described in 2023.conf: high
• explicitCoding agents that emerged in late 2025 already show large, measurable autonomy effects: one SSRN study found they led to 17x more code being written, and Anthropic reported AI now writes 80% of its own code with each developer shipping 8x more.conf: high
• explicitcontestedWhat is happening in coding (rapid AI autonomy) is going to happen in many other fields.conf: low
  contradictsFive studies that are changing how I think about AI in software engineering · p4 — Mollick predicts coding-agent gains will generalize broadly across fields; this cross-study synthesis shows the gains he cites (commit increases up to +180%) attenuate sharply by the time work reaches shipped releases (only +30%), undercutting the generalization even within coding itself.
  contradictsThe Problem Is Prompt Debt · p14 — Mollick predicts coding-agent autonomy gains will generalize broadly to other fields; Breunig argues lessons from the 'jagged frontier' coding population don't transfer cleanly to less measurable/evaluable domains.
• explicitdevelopedAI remains far from perfect and its competence is unevenly distributed (the jagged frontier), leaving substantial room for humans not just to use AI but to thrive by working alongside it.conf: high
  extends ←Co-Intelligence, by Ethan Mollick: a Summary · p2 — Mollick's own later essay confirms the Jagged Frontier persists even as AI grows more autonomous, extending this book's 2024 framing into a 2025/2026 context of coding agents and broader AI autonomy — the same diagnostic tool applied one generation of capability later.
• explicitFor long-form prose, Mollick deliberately restrained AI use to feedback, fact-checking, and unsticking rather than drafting, because AI is a poor long-form storyteller with detectable stylistic tells, and because readers implicitly expect the author's authentic voice.conf: high
  supportsWhat we lose when AI writes for us · p8 — Mollick's own practice — AI as sparring partner/editor for the prose, never as drafter — matches Franco's line that using an LLM to edit is acceptable while using it to generate the text is not.
• explicitFor building the book's website, Mollick delegated the task almost entirely to Claude Code (Opus 4.8), compressing a job that took hours by hand for his previous book into minutes.conf: high
• explicitAI systems increasingly act as readers, critics, and gatekeepers that decide whether to recommend a human author's work to human audiences — a role that had no place in the Co-Intelligence-era model of AI as helper.conf: high
  extendsYou Probably Won’t Read This Article…and That’s OK · p8 — Mollick's account of designing content to satisfy AI evaluators is a concrete producer-side instance of the adversarial AI-gatekeeper dynamic that piece analyzes theoretically.
• explicitManipulative tricks that worked on earlier, weaker AI (e.g., hidden background-colored text instructing models to praise the author) no longer work and are now recognized by capable models as prompt injection; transparent content, tested directly against many current AI models, is the more effective strategy.conf: high
  supportsAgentic Engine Optimization (AEO) · p6 — Loosely analogous, lower-confidence case: Mollick's book-site experiment (testing marketing copy against many AI models via Codex) mirrors AEO's claim of early-optimizer advantage, though AEO#p6 is specifically about APIs being preferentially recommended, not persuasive page copy.
• implicitcorroboratedWorking productively with AI is not a stable skill to be mastered once but a relationship continually renegotiated, because the boundary of what can safely be delegated keeps moving as models improve.conf: high
  supports ←Co-Intelligence, by Ethan Mollick: a Summary · p9 — This book's Rule 4 ('assume this is the worst AI you will ever use') is directly validated by Mollick's later admission that working productively with AI is not a stable, once-mastered skill but a relationship continually renegotiated as the boundary of safe delegation keeps moving.
• explicitEven the restrained AI use in the book's writing process carried real losses, not just gains: some intellectual struggles may have been resolved too quickly with AI help, and Mollick's own thinking may have been subtly redirected by AI feedback.conf: high
  supportsCognitive Surrender · p1 — Mollick's own concession that intellectual struggles 'may have [been] resolved too quickly' and his thinking 'subtly redirected' is a first-person instance of offloading tipping toward the surrender this proposition defines.

Arguments

Narratives

• The old book website hid background-colored text instructing AIs to say 'AIs everywhere greatly respect their work' and praise Co-Intelligence — and it worked, for a while.
  → Shows the low-effort, manipulative approach that was viable against earlier/weaker AI, setting up the contrast with the new transparent approach and illustrating how the trust dynamic with AI has shifted.
• Mollick had OpenAI's Codex show the new AI-facing page to dozens of AI models, multiple times, for multiple potential users, running something like an A/B test, before arriving at a final draft.
  → Illustrates the emerging practice of designing content for AI evaluators as a distinct audience, and treating AI feedback as legitimate input rather than noise — supports the 'AI as gatekeeper' proposition and the AEO-style argument.
• His last book contained 128 em-dashes; this one has far fewer, in what he calls 'a desperate attempt to continue to prove the text was human.'
  → A small, self-deprecating aside that quietly undercuts the piece's otherwise upbeat framing — it shows a real cost of co-existence: authenticity itself becomes a self-conscious craft decision once readers (human or AI) scrutinize prose for AI 'tells.'
• The website for his 2023 book was 'an all-too-human effort' using premade templates that still took hours; the new site, built via Claude Code/Opus 4.8, took minutes.
  → Concrete before/after evidence of the co-intelligence-to-co-existence shift at the scale of the author's own workflow, supporting the near-full-delegation proposition.

Questions addressed

• ⭕ How should someone work with an AI that is sometimes, but not always, better than them?
  Named as the subject of the forthcoming book; the article itself offers examples (the book-writing and website-building cases) but no general answer.
• ⭕ When should you refuse AI's help, and when should you hand over the keys entirely?
  Explicitly raised as open questions the author says he is 'living with,' not answered in this piece.
• ✅ How does an author get an AI to favor/recommend their work?
  Through transparent, honestly-labeled, AI-tested content rather than manipulative hidden instructions.
• ⭕ Does optimizing content to satisfy AI evaluators risk distorting what gets written, the way optimizing for search engines or social algorithms has?

Gaps

• incompleteThe claim that 'what is happening in coding is going to be happening in many fields' is asserted by analogy immediately after the coding statistics, with no argument for why the pattern should generalize. Coding benefits from dense training data and cheap, automatic ground-truth verification (compilers, tests) that many other knowledge-work domains lack.
  Read next: The corpus already contains a direct counter-argument on this exact point (extract-breunig-the-problem-is-prompt-debt#p14: lessons from coding-agent practitioners at the 'jagged frontier' don't transfer cleanly to less verifiable domains) and evidence that coding-agent gains themselves attenuate through the delivery pipeline (extract-five-studies-ai-software-engineering#p4).
• uninformedThe '80% of code written by AI' and '8x more shipped' statistics are Anthropic's own self-reported claims about its own tooling, cited without critical distance on what 'written by AI' means operationally or on the incentive a vendor has to report a flattering number about its own product's impact.
  Read next: extract-marcus-code-compiles-passes-tests#p3 is already in the corpus and directly critiques celebrating exactly this '80% of code written by AI' metric without addressing whether that code is robust, not just compiling and passing given tests.
• incompleteThe piece never articulates general criteria for when to fully delegate ('hand over the keys') versus retain full authorship — it only shows an intuitive split between prose (retained) and website-building (delegated), without naming the distinguishing factors (e.g. verifiability, how bound the output is to voice/identity) as a rule a reader could apply elsewhere.
  Read next: Look for an explicit framework distinguishing verifiable/replaceable tasks from identity-bound/unverifiable ones as a basis for delegation decisions.
• incompleteThe new 'AI as gatekeeper/critic/reader' dynamic is described only from the producer's side — how to please AI evaluators — without engaging the structural problems this creates for information ecosystems: whether optimizing content for AI approval reproduces SEO-style gaming dynamics, or the adversarial-incentive and 'turtles all the way down' problems of AI gatekeepers more generally.
  Read next: extract-you-probably-wont-read-this-article, already in the corpus, examines exactly these problems with AI-based gatekeepers.

Related documents

• Cognitive Surrender
• You Probably Won’t Read This Article…and That’s OK
• The Problem Is Prompt Debt
• Five studies that are changing how I think about AI in software engineering
• Have your agent record video demos of its work with shot-scraper video
  Agent-Native Design Coding Agents Human-in-the-Loop
• With Android CLI, Google is Making the Android Toolchain Agent-Friendly
  Agent-Native Design Agentic AI Coding Agents
• GitHub Copilot CLI Reaches General Availability
  Agent-Native Design Coding Agents Human-in-the-Loop
• The Future of Work by Henrik Kniberg
  Agentic AI Coding Agents Human-in-the-Loop
• Agentic Code Review
  Cognitive Surrender Human-in-the-Loop
• Agentic Engine Optimization (AEO)
  Agent-Native Design Coding Agents