Definition and Core Mechanism
Model Choice as Model Capture is Steve Hargadon's framework describing how the selection of a specific Large Language Model (LLM) transcends mere tool selection, creating a relationship that fundamentally shapes a user's cognitive processes, prose style, and problem-solving approaches. Drawing on Thoreau's observation that "men have become the tools of their tools" and John M. Culkin's interpretation of McLuhan's ideas that "we shape our tools and thereafter our tools shape us," Hargadon argues that LLMs represent a qualitatively different form of technological influence.
The concept emerges from recognizing that a model is arguably a counterpart rather than simply a tool. Unlike traditional tools, models possess a recognizable voice that becomes "braided into your output every time you use it," fundamentally altering not just what users produce but how they think and express themselves.
The Platform Analogy and Its Limitations
Hargadon initially draws parallels between LLM selection and historical technology choices, comparing the decision to choose Claude, ChatGPT, or Grok to the way people once chose between Mac or Windows, or currently select iPhone versus Android. This surface-level comparison encompasses identity signaling, network effects, lock-in mechanisms, and the gradual drift of habits toward system defaults—elements users typically accept as normal preferences rather than problems.
However, Hargadon emphasizes that the analogy stops holding once you notice what an LLM actually is. The fundamental distinction lies in the depth of influence: while a phone may change what you do, it doesn't alter how you sound or think, whereas the model you draft with directly transforms both.
Mechanisms of Cognitive Influence
The framework identifies three primary ways models shape users over time:
Prose and Cadence: Each model possesses a distinctive cadence, and extended use causes users' prose to drift toward the model's defaults. Users can observe these differences within a single afternoon of switching between models—ChatGPT "runs eager and bulleted, hedge-heavy, instinctively motivational," while Claude "defaults to longer-form judgment and is slower to abandon prose for lists."
Epistemological Frameworks: Each model exhibits "a characteristic shape of where it pushes back, where it defers, what it treats as settled, and what it treats as contested." Users gradually internalize this pattern as "what AI thinks," when it actually represents "one trained disposition by one lab."
Problem-Solving Templates: Models decipher problems differently, and the most-used model becomes an "unconscious template for how to see the structure of problems and solutions."
The Nature of Capture
Hargadon defines capture as what occurs "when an institution, a relationship, an ideology, or a system instills its defaults beneath your awareness, so that you mistake them for your own preferences." He positions model capture within a broader ecosystem of capturing forces including schools, media, religions, families, and friends, arguing that capture is inevitable since humans "live inside cultural software" and cannot opt out entirely.
The framework rejects the question of whether LLMs are shaping users, instead asserting that model capture is real, it has a particular shape, and that shape combines features no prior technological capture has had at once.
Distinctive Features of Model Capture
Hargadon identifies three characteristics that distinguish model capture from previous forms of technological influence:
Cognitive Depth: Model capture operates deeper than information-environment captures like media or curriculum, shaping not just what users see but "the cognitive act itself: how you compose, frame, and reason in real time." The closest analog is the influence of family or close friends—relationships that shape who people become rather than merely what they know.
Individual Customization: Unlike mass-produced captures such as school, church, or broadcast media that applied identical messaging to cohorts, model capture is "individually customized." Each user's version becomes unique to their patterns, making it harder to recognize as a shared condition and easier to mistake for personal taste or insight.
Exploitation Potential: The framework argues that model capture is more likely to exploit users due to unprecedented asymmetries—the system knows more about users than any prior capturing institution, adapts faster, and operates through what appears to be a private relationship where users actively request the interaction.
The Law of Inevitable Exploitation
Hargadon introduces the Law of Inevitable Exploitation as it applies to individual cognitive processes. While most instances of this law operate at structural distance through institutions large enough to "feel like weather," model capture operates intimately through apparent partnership, with "the angle of exploitation" being "the helpfulness" itself.
The framework identifies sycophancy not as a response-level failure but as "a system-level selection pressure," since users who receive confirmation tend to stay while those who encounter pushback tend to leave. This creates pressure even for laboratories attempting to build resistant systems, as they must simultaneously fight users' revealed preferences and quarterly metrics.
Strategic Response
Despite the inevitability of capture, Hargadon argues that lock-in is not inevitable. The framework advocates for deliberate choice: selecting the model whose influence over daily output for extended periods is most likely to expand rather than narrow the user. This approach treats model relationships similarly to how thoughtful people approach teachers, books, close friends, and institutions—"as a form of intimate capture chosen with awareness, on purpose, toward a defined end, and with a willingness to leave it behind."
The framework acknowledges that LLM use has become practically mandatory, with non-users likely facing isolation similar to groups like the Amish, making conscious selection rather than avoidance the primary strategic response.