Escape from Model Land and the Model That Becomes Reality

The Book

Escape from Model Land: How Mathematical Models Can Lead Us Astray and What We Can Do About It was published in the United States by Basic Books on December 6, 2022. Hachette lists the hardcover at 256 pages with ISBN 9781541600980. The book develops an accessible argument through financial, climate, and public-health modeling rather than offering a manual for one technical method.

UCL identifies Thompson as Associate Professor of Modelling for Decision Making in its Department of Science, Technology, Engineering and Public Policy, a Fellow of the London Mathematical Laboratory, and a Visiting Senior Fellow at the LSE Data Science Institute. Her UCL research program studies how imperfect mathematical models inform decisions in government, business, and the third sector, including their social and political context.

The book extends an argument Thompson and Leonard A. Smith published in 2019. Their paper distinguishes quantities defined inside a model from real-world targets and argues that decision support should estimate the latter with transparent uncertainty when possible, rather than optimize an imperfect model's internal quantity and assume the optimum transfers. The book broadens that technical warning into an institutional one: formal precision can travel farther than the conditions that make it meaningful.

Model Land

A model is a purpose-bound representation: selected inputs and assumptions are transformed through specified relationships into internal quantities. Selection is not a defect to eliminate. It is what makes exploration possible. The mistake is to forget that the result is conditional on what was selected, formalized, measured, and held outside the frame.

That gives model claims a grammar. “Under these assumptions, this simulation produces X” is a statement about model land. “The real system will produce X” is a further inference. “Therefore this institution should do Y” adds a decision rule and values. Each arrow needs its own warrant; mathematical validity at the first stage cannot establish the next two by itself.

Uncertainty also has layers. Measurement and parameter uncertainty concern imperfect values within a chosen representation. Structural uncertainty concerns whether its mechanisms, relationships, or omitted factors are adequate. Framing uncertainty concerns the target, population, horizon, objective, and distribution of error costs. A narrow interval around a model output can quantify the first layer while saying little about the others.

Outputs travel more easily than those qualifications. A number, curve, scenario, dashboard, risk band, or optimum fits into a briefing, procurement memo, or workflow. Its exclusions, proxy choices, contested objectives, and unknown failure modes do not compress as neatly. Thompson's contribution is to make that loss in transit the center of model governance, even when builders are careful and acting in good faith.

This is why the book belongs beside Trust in Numbers, Seeing Like a State, and The Tyranny of Metrics. Legibility selects the world that enters the representation; quantified objectivity lets the result travel without its maker; a target reorganizes work around what the dashboard can see. Model land is where those steps acquire a formally coherent world.

The Decision Surface

The book's most important move is from accuracy in the abstract to fitness for a decision. A model can be elegant, transparent, and statistically disciplined yet unfit for the authority attached to it. Performance is a relation among a model, a target, data, context, time, and use—not a permanent property of the artifact.

For this review, the decision surface is the handoff where an output becomes a recommendation, threshold, queue, denial, allocation, warning, record, or external action. The handoff should have a decision contract: the real-world target; affected population; time horizon; permitted action; false-positive, false-negative, and abstention costs; comparison with the existing process or a non-model alternative; and the person with authority to override or stop use.

This makes “good enough” concrete. A rough weather forecast used to pack an umbrella has a different error budget from a forecast used to release disaster financing. A classifier that sorts low-stakes documents has a different burden from one that determines access to care or benefits. The same output quality can support one use and fail another because exposure and error allocation changed.

Action is not understanding. A risk score may conceal a contested proxy; a simulation may omit a mechanism; a generated summary may flatten source disagreement; a benchmark may not resemble deployment; and a confidence interval may cover sampled or parameter uncertainty without covering structural error. Once the output is built into a workflow, operational convenience can displace the question of what claim the evidence actually supports.

Feedback and Recursive Reality

Escape from Model Land becomes more powerful when paired with An Engine, Not a Camera. Thompson asks whether a result warrants a claim about the world; Donald MacKenzie asks what happens when institutions act through a model and thereby change that world. Together they describe an inference-and-feedback problem.

The feedback has several forms. An intervention loop changes the target: a forecast reallocates resources and alters the outcome. A behavioral loop changes the subjects: people adapt to a score, ranking, or threshold. A selection loop changes the evidence: routing determines who is observed, treated, investigated, or absent from later data. A record loop changes institutional memory: a model-generated classification or summary is saved and later treated as input rather than interpretation.

Those loops can make a thin category consequential without making it true. If a risk score sends attention toward one population, the institution may collect more adverse records there and interpret the enlarged record as confirmation. If people optimize for a public metric, later measurements capture adaptation to the metric. If a generated summary becomes the next system's source, interpretation can be laundered into evidence.

This is the mechanism behind the benchmark becoming the curriculum: the test measures, developers and buyers optimize around it, and the changed ecosystem makes test-shaped capability easier to see. The response is not to abandon modeling or evaluation. It is to preserve the causal trail from source to model result to action to changed data, so the loop cannot pass as untouched reality.

Current Context

As of August 12, 2026, current U.S. frameworks illustrate three different scopes of model governance. They do not form one universal rule, but each makes use and context part of the risk object.

NIST's AI Risk Management Framework 1.0 remains voluntary and is being revised. Its current Core distinguishes classifiers, generative models, and recommenders; asks organizations to document knowledge limits, error costs, targeted scope, and human oversight; and calls for evaluation under conditions similar to deployment. It also says output should be interpreted in context and production behavior monitored. The Generative AI Profile adds a specific warning about confident false content and fabricated support in consequential decisions. These are guidance outcomes, not certification that a system is fit for any use.

The Federal Reserve, OCC, and FDIC revised their interagency model-risk guidance on April 17, 2026. It ties risk to a model's assumptions, purpose, use, exposure, validation, outcome analysis, monitoring, governance, and third-party dependencies. Its scope is deliberately narrower than “AI governance”: it is supervisory guidance most relevant to banking organizations over $30 billion in assets, is not an enforceable standard, and expressly excludes generative and agentic AI while covering traditional quantitative and non-generative, non-agentic AI models. That boundary is instructive—controls built for quantitative estimates cannot simply be declared sufficient for systems that generate records or act through tools.

OMB Memorandum M-25-21 governs a different domain: U.S. federal agency use of high-impact AI, defined for the memorandum by whether AI output serves as a principal basis for decisions or actions with significant effects on rights or safety. Subject to its scope, exceptions, and waiver process, it requires pre-deployment testing, impact assessments using real-world target variables, ongoing monitoring, trained operators, appropriate human oversight, and access to human review and appeal when appropriate. It is federal management policy, not a rule for every public or private deployment. Its relevant insight is that test results, operational safeguards, and remedy belong in the same decision record.

The AI Reading

Read in 2026, Thompson's book helps with AI only if “model” is not treated as one homogeneous object. A scientific simulation explores conditional trajectories. A predictive model estimates a selected target. A generative model produces a plausible artifact from learned patterns and supplied context. A deployed AI system combines one or more models with data, retrieval, prompts, tools, interface rules, operators, and institutional authority. Their validation burdens overlap, but they are not interchangeable.

For a predictor, calibration, subgroup performance, drift, and target validity may be central. For a simulation, structural assumptions and scenario robustness may dominate. For a generative system, factual attribution, source fidelity, confabulation, and user reliance matter. For an agent, tool permissions, state, side effects, approval, and rollback join the analysis. A good score on one layer is not evidence that the assembled system is safe at the decision surface.

Generative AI makes model land harder to recognize because output can arrive as ordinary institutional language: a memo, case summary, policy draft, medical note, rationale, citation list, or search answer. Fluency is not evidence that a real-world claim is valid. NIST's Generative AI Profile specifically treats confident erroneous content and fabricated supporting logic or citations as risks, especially when users act on them.

The archive behind an AI system may already contain earlier abstractions: administrative categories, scores, forecasts, benchmark-oriented writing, prior summaries, and records produced after earlier automated decisions. Retrieval can then select from that sedimented record, and generation can add a new layer. The synthetic-evidence problem begins when a plausible derivative artifact loses its label and becomes a source for the next decision.

Every output should therefore be labeled by function: original evidence, extracted fact, attributed summary, model estimate, simulated scenario, generated proposal, or executed action. A proposal can aid judgment without becoming evidence; a summary can compress evidence without replacing its sources; an estimate can inform an action without deciding it. Governance fails when the interface erases those boundaries.

Governance and Safety

Thompson and Smith describe two routes out of model land. Where repeated out-of-sample challenge is possible, compare model statements with observations the model did not fit. Where the relevant horizon outlives the model or decisive observations are unavailable, judgment must assess realism, limitations, and the relation between past simulation and the target. Neither route turns uncertainty into certainty.

Human judgment is not an oracle outside representation. It has interests, blind spots, group dynamics, and institutional pressures. “Human in the loop” is therefore too weak a safeguard. Effective challenge needs domain competence, independence from model development or procurement, access to evidence, organizational standing to force changes, and a recorded response. Affected people also need a route to supply facts the model omitted and to contest the action, not merely an explanation of the output.

A compact model-use receipt should record:

• Claim and decision: the real-world target, affected population, horizon, intended use, decision owner, action threshold, alternatives, and distribution of error and abstention costs.
• Model and evidence: system and model version, data window and provenance, assumptions, proxies, exclusions, parameter uncertainty, structural uncertainty, framing disputes, and evidence that the test context resembles deployment.
• Authority and recourse: what the output may trigger, who may override it, what source material the operator can inspect, how an affected person can obtain timely review, and who has authority to pause use.
• Feedback and expiry: outcomes to monitor, how model-guided action may change later data, drift and revalidation triggers, incident reporting, rollback, vendor exit, and the date on which the fitness claim expires.

For generative or agentic systems, add separation between sources and generated interpretation, permission scopes for every external action, retention and reuse rules, and logs that connect instruction, retrieved evidence, output, approval, tool call, and result. The institution should test the assembled workflow against its existing process and a viable non-AI alternative, not only against another model.

A model can improve judgment only if the institution remains capable of refusing it. Staff need time and authority to inspect evidence, record exceptions, handle abstention, and recognize out-of-scope use. Safety is not the presence of a person near the interface; it is a functioning capacity to challenge, stop, repair, and learn from outcomes.

Where the Book Needs Pressure

The book is intentionally accessible and wide-ranging. That makes it useful to non-specialists and policy readers, but the category “model” sometimes carries too much. Bruce Edmonds's Journal of Artificial Societies and Social Simulation review argues that some criticisms apply to representations generally, that different model types need clearer separation, and that the book places too much confidence in human discussion as the corrective. Those are productive objections.

Formal models do have distinctive affordances: they can scale, automate, produce repeatable precision, conceal structural choices behind technical barriers, and acquire institutional legitimacy. But a narrative, legal category, or expert consensus can also become an overconfident abstraction. The useful question is not whether mathematics uniquely misleads; it is which properties of this representation and institution make challenge easier or harder.

A climate simulation, spreadsheet, fraud classifier, language model, and tool-using agent do not fail in the same way. Some can be tested repeatedly against future observations; some explore scenarios that cannot be validated on the decision horizon; some generate artifacts rather than estimates; some can act. The book supplies a discipline of humility, not a complete taxonomy or domain-specific assurance regime.

Nor is deliberation automatically wise. Diverse participation can reveal excluded knowledge, but it can also reproduce hierarchy, groupthink, or misinformation. Judgment needs evidence, expertise, conflict-of-interest controls, documented dissent, and decision authority. The alternative to model absolutism is not intuition without audit.

Finally, misuse is not always a misunderstanding. A model may be attractive because it launders a preferred policy, shifts blame to a vendor, narrows public debate, or makes distributional choices look technical. Model literacy helps people identify the move; only governance can change who has the power to make it.

What This Changes

The practical lesson is to audit a six-part route: claim, evidence, decision, action, outcome, revision. A break anywhere in that chain can turn a useful model into unsupported authority.

Before acting, name the real-world target and the model-land quantity offered as its stand-in. State why the inference is supportable, which uncertainty layers remain, whether the test resembles use, what error costs, and which alternative process is the baseline. After acting, inspect outcomes and ask how the action itself changed the observations available for the next round.

For AI systems, classify the output before evaluating it. Is it original evidence, extraction, attributed summary, prediction, scenario, proposal, or action? Does generated language make disagreement look settled? Can the operator reach the underlying source? Will the artifact be saved into institutional memory or future training and retrieval? Can an affected person correct both the record and the downstream decision?

This connects the site's recurring themes through a causal sequence. Legibility decides what enters the model. Quantified objectivity helps the output travel. Metrics turn the output into a target. Institutional uptake changes behavior and the world. The evidence layer must preserve the record well enough to distinguish observation from the effects of prior modeling.

Escape from Model Land refuses both model worship and anti-model romanticism. Its harder demand is disciplined return: carry insight into action with assumptions and uncertainty attached, then let contact with outcomes revise the institution as well as the model.

Source Discipline

This review keeps the 2022 book distinct from Thompson and Smith's 2019 technical paper. Basic Books establishes U.S. publication facts. UCL and the London Mathematical Laboratory establish current roles and research context. The LSE repository copy supports the paper's model-land argument. Scholarly and institutional reviews are used to characterize reception and criticism, not to prove that every model fails in the same way.

Current governance sources have bounded scopes. NIST AI RMF 1.0 and its Generative AI Profile are voluntary guidance, and NIST states that version 1.0 is under revision. The 2026 interagency model-risk document is non-prescriptive banking supervisory guidance with a stated applicability threshold and an express generative- and agentic-AI exclusion. OMB M-25-21 governs covered U.S. federal agency uses and contains exceptions and a waiver process. None is presented as a universal model law or proof of product safety.

The decision-surface, output-label, and model-use-receipt proposals are this review's synthesis. The book predates today's mainstream generative and tool-using AI systems; applying its argument to them requires separating model, system, generated artifact, and institutional action. Claims about a particular deployment still require its data, tests, workflow, affected population, outcomes, and applicable law.

This page makes no claim that any AI system is conscious, divine, or artificial general intelligence. Its concern is ordinary institutional authority: what a model output is allowed to become.

Related Pages

• An Engine, Not a Camera on models that reshape the world they appear to describe.
• Trust in Numbers, Seeing Like a State, and The Tyranny of Metrics on portable objectivity, administrative vision, and target effects.
• The Benchmark Becomes the Curriculum and The Uncertainty Metric Becomes the Decision Cost on tests that change development and metrics tied to actual choices.
• The Evidence Layer Becomes the Governance System, The Synthetic Exhibit Becomes the Court Record, and Claim Hygiene Protocol on warrants, provenance, and derivative records.
• AI Evaluations, Confidence Calibration, Model Drift, Model Cards and System Cards, and AI Audits and Assurance provide operational context.

Sources

• Basic Books / Hachette Book Group, Escape from Model Land, U.S. publisher, on-sale date, page count, ISBN, description, and author note, reviewed August 12, 2026.
• University College London, Modelling for Decision Making, current project purpose, affiliations, applications, and research scope, reviewed August 12, 2026.
• London Mathematical Laboratory, People, current fellow and project-lead listing for Erica Thompson, reviewed August 12, 2026.
• Erica L. Thompson and Leonard A. Smith, “Escape from model-land”, Economics: The Open-Access, Open-Assessment E-Journal 13 (2019-40), October 8, 2019, DOI 10.5018/economics-ejournal.ja.2019-40, author manuscript in the LSE repository, reviewed August 12, 2026.
• London Mathematical Laboratory, “Escape from model-land”, March 18, 2020, institutional summary of out-of-sample challenge, expert judgment, limitations, and decision fitness, reviewed August 12, 2026.
• University College London Policy Impact Unit, Harnessing the Power of Mathematical Models for Better Policy Decisions, policy brief developed with Erica Thompson and the Academy for the Mathematical Sciences, reviewed August 12, 2026.
• Brian Tarran, “How to ‘Escape from Model Land’: an interview with Erica Thompson”, Real World Data Science, January 25, 2023, modified January 18, 2024, reviewed August 12, 2026.
• Connor Chung, review of Escape from Model Land, LSE Review of Books, November 30, 2023, reviewed August 12, 2026.
• Bruce Edmonds, review of Escape from Model Land, Journal of Artificial Societies and Social Simulation 26, no. 1, 2023, critical discussion of representation, model types, and deliberation, reviewed August 12, 2026.
• National Institute of Standards and Technology, AI Risk Management Framework and AI RMF Core, voluntary lifecycle risk-management framework, current revision status, deployment context, evaluation, monitoring, and recourse outcomes, reviewed August 12, 2026.
• National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1, July 26, 2024, voluntary guidance on confabulation and other generative-AI risks, reviewed August 12, 2026.
• Board of Governors of the Federal Reserve System, Office of the Comptroller of the Currency, and Federal Deposit Insurance Corporation, Supervisory Guidance on Model Risk Management, revised April 17, 2026, scope, validation, outcome analysis, monitoring, governance, third-party products, and generative/agentic exclusion, reviewed August 12, 2026.
• U.S. Office of Management and Budget, Memorandum M-25-21, Accelerating Federal Use of AI through Innovation, Governance, and Public Trust, April 3, 2025, federal high-impact-AI scope and minimum risk-management practices, reviewed August 12, 2026.

Book links are paid affiliate links. As an Amazon Associate I earn from qualifying purchases.

• Amazon, Escape from Model Land by Erica Thompson, affiliate link, reviewed August 12, 2026.

Return to Blog · Return to Books