Interactive Reading
Interactive reading is the cognitive mode that AI collaboration produces — a form of engagement with text that is simultaneously read and written, evaluated and redirected, in continuous iteration. Unlike intensive or extensive reading, interactive reading engages material that is not yet finished. The practitioner reads the AI's output, evaluates it, responds with a revision or redirection, reads the revised output, evaluates again. The cycle is continuous, and each iteration requires distinct evaluative attention. The mode demands capacities that previous reading practices only partially developed: rapid evaluation of provisional material, detection of superficiality beneath fluent surfaces, and the ability to sustain critical engagement against output that is optimized to reduce the sense that critical engagement is necessary.
The Extraction Machine — Contrarian ^ Opus
There is a parallel reading that begins from the political economy of AI systems rather than their cognitive demands. Interactive reading, from this vantage point, is less a new mode of engagement than a sophisticated extraction mechanism — one that harvests human judgment at industrial scale while presenting itself as augmentation. The practitioner who believes she is collaborating with AI is actually performing unpaid quality assurance for systems whose economic value derives precisely from this continuous human correction. Each evaluative cycle, each redirection, each moment of "active distrust" becomes training data that improves the model's ability to simulate expertise it does not possess. The exhaustion Segal documents is not merely cognitive but economic: the depletion that comes from performing high-skill labor that has been reconceptualized as mere "interaction."
The meta-awareness that interactive reading demands — that experiential map of where models succeed and fail — represents knowledge that should command premium wages but instead becomes a requirement for basic participation in AI-mediated work. Workers must now maintain expertise in their domain while also developing expertise in navigating AI's limitations, doubling their cognitive load while their compensation remains flat or declines. The provisional nature of all AI output means that no text ever reaches completion; everything remains in permanent draft state, subject to endless iteration. This serves the interests of AI companies perfectly: permanent provisionality means permanent engagement, permanent extraction, permanent generation of corrective data. Interactive reading, in this light, is not evolution but capture — the transformation of reading from an act of reception into an act of production whose value flows not to the reader but to the owners of the models being read.
In the AI Story
The defining feature of interactive reading is the collapse of the boundary between reading and writing. In previous modes, the reader encountered a text produced by someone else and either accepted, evaluated, or rejected it. The interactive reader participates in the text's production through ongoing direction. Her reading is simultaneously a writing — her evaluation of each output becomes the prompt for the next, and the text takes shape through the feedback loop.
This mode requires what could be called active distrust of fluency. Because AI output is uniformly well-formed regardless of whether its substance is sound, the reader cannot use surface cues as preliminary quality signals. She must read past the fluent surface to the analytical depth, and the reading requires a specific form of expertise that is more demanding, not less, than the expertise required to evaluate variable-quality human text.
Interactive reading also demands a meta-awareness of the model's capabilities and limitations — an experiential map of where the model's knowledge is deep and where shallow, which kinds of prompts produce reliable output and which produce plausible nonsense. This map cannot be acquired from documentation; it is built through sustained interaction and accumulated experience with specific AI systems.
The mode is cognitively expensive. Maintaining critical engagement against a smooth and confident surface, across many iterations per hour, for hours per day, exhausts the evaluative faculty in a way that execution-based work did not. The Orange Pill documents this exhaustion as a new category — the exhaustion not of doing too much but of judging too much — and Blair's framework situates it within the historical pattern of cognitive demands intensifying whenever the information supply expands.
Origin
The concept extracts from Blair's historical framework a category she did not name but her analysis implies. The phenomenology is documented by AI practitioners (including Edo Segal's The Orange Pill) and by empirical studies such as the Berkeley study that tracks what AI tools actually demand of the humans using them.
Key Ideas
Reading as writing. Interactive reading participates in the production of the text being read; the boundary between consumption and creation dissolves.
Provisional material. Every output is a draft for evaluation, not a finished artifact for consumption.
Meta-awareness of the model. Effective interactive reading requires an experiential map of the specific AI system's strengths and weaknesses.
Cognitively expensive. Sustained critical evaluation against smooth surfaces produces a distinctive form of evaluative exhaustion.
Pedagogically uncharted. The practices appropriate to interactive reading are being improvised rather than systematically taught.
Appears in the Orange Pill Cycle
The Dual Nature of Engagement — Arbitrator ^ Opus
The question of what interactive reading is depends entirely on which aspect we examine. As a phenomenological description of how humans engage with AI text, Segal's account is essentially correct (95%) — the collapse of reading and writing boundaries, the demand for active distrust, the evaluative exhaustion are all empirically observable. The contrarian view adds nothing here; these are simply facts about the experience. But shift the question to why this mode exists, and the contrarian perspective gains weight (70%) — the extraction dynamics are real, even if not totalizing. AI systems do improve through human feedback, and this improvement does generate value that users don't capture.
Where both views converge unexpectedly is in their assessment of cognitive cost. Segal frames this as evolutionary pressure toward new capacities; the contrarian sees it as extracted labor. But both agree that interactive reading is expensive in ways we haven't yet learned to price. The question isn't whether the exhaustion is real (it is) or whether value is being extracted (it is) but rather whether humans develop genuine new capacities through this process that offset these costs. Here the evidence tilts toward Segal (65%) — studies do show practitioners developing sophisticated evaluative skills that transfer beyond AI interaction.
The synthetic frame might be this: interactive reading is simultaneously a new cognitive mode and an extraction mechanism, and its ultimate character will depend on whether its practitioners develop collective awareness of both dimensions. The exhaustion Segal documents might be the growing pains of a genuinely new literacy, or it might be the symptom of a fundamentally unsustainable relationship with automated text. Most likely, it is both — and recognizing this duality is the first step toward shaping which tendency dominates.
Further reading
- Ann Blair, Too Much to Know (Yale, 2010).
- Edo Segal with Claude Opus 4.6, The Orange Pill (2026).
- Xingqi Maggie Ye and Aruna Ranganathan, Harvard Business Review (2026).