The Pattern Matching Conditions: A Framework for Predicting Which Jobs AI Will Actually Displace

The coverage of Claude Mythos Preview has generated the familiar cycle: a genuine capability advance in a specific domain gets extrapolated into sweeping claims about the end of professional work as we know it. In the first piece of this trilogy I explained why the Mythos cybersecurity findings, while real and significant, do not constitute AGI and were predicted by existing analytical frameworks. In this piece I want to do something more useful than either alarm or reassurance. I want to give you a precise analytical tool for identifying which jobs face genuine structural displacement risk and which do not.

Call it the Pattern Matching Conditions framework. It consists of three conditions that must all be satisfied simultaneously for a job to face genuine displacement risk from current and near-term LLM capability. Where all three hold, displacement is plausible on a 10 to 20 year timeline regardless of the other economic absorption mechanisms. Where any one fails, the job is substantially protected. The framework is falsifiable, specific, and grounded in the same analytical logic that predicted the Mythos cybersecurity findings before they arrived.

The Three Conditions

The first condition is binary or near binary outcomes. The output of the work must be evaluable as correct or incorrect without requiring contextual human judgment. Code either runs or it does not. A vulnerability either exists or it does not. A contract clause either complies with a standard or it does not. A tax return either correctly applies the code or it does not. When correctness is binary and verifiable the LLM has an oracle to work against. Without that oracle the model cannot reliably distinguish its good outputs from its bad ones and human judgment must fill the gap.

The second condition is large high quality training data. The domain must be extensively documented in digital form at a quality level sufficient for the model to have learned the relevant patterns. Legal documents, financial statements, code repositories, medical imaging labels, actuarial tables, and regulatory filings have all been produced in enormous quantities and are heavily represented in training corpora. Domains where knowledge is primarily tacit, embodied, relational, or undocumented do not satisfy this condition regardless of how binary their outcomes might be.

The third condition is complete digital manipulability. The entire job must be executable through a computer interface without requiring physical world interaction, contextual relationship management, or embodied judgment. A radiologist interacts with images on a screen. A document reviewer reads and classifies text. A junior coder writes in an IDE. A financial auditor works with spreadsheets and databases. These jobs are fully digital in the sense that every input and output passes through a computer interface. A plumber diagnosing a specific pipe configuration in a specific building with specific history does not satisfy this condition regardless of how binary the diagnosis might be.

The Jobs That Satisfy All Three

Working through the major professional categories systematically produces a specific and bounded list of genuinely at-risk occupations. These are not the jobs the doomer discourse typically focuses on. They are not truck drivers or factory workers. They are educated professional jobs that have historically been considered safe from automation precisely because they require cognitive sophistication. The Pattern Matching Conditions reveal that cognitive sophistication is not the relevant protection. The relevant protection is whether the work requires contextual human judgment that cannot be reduced to pattern matching against a binary oracle.

Legal document review and discovery analysis satisfy all three conditions almost perfectly. Contracts and legal documents either contain specific clauses, admissions, or relevant evidence or they do not. The training data from decades of legal filings, case law, and regulatory documents is vast and high quality. The entire task is executed through a computer interface. Junior associates at large law firms spending their days on discovery review are doing work that fits the Pattern Matching Conditions precisely. The timeline for significant displacement in this specific task category is probably closer to 5 to 10 years than 10 to 20.

Entry level software development satisfies all three conditions. Code either works or it does not. The training data from decades of open source repositories, Stack Overflow, documentation, and academic papers is the most extensive of any professional domain. The work is entirely digital. Junior developers writing boilerplate, implementing documented APIs, debugging known error patterns, and writing unit tests are doing Pattern Matching Condition work. The Mythos findings and the Claude Code architecture together confirm this. The gradient steepens sharply at the senior level where architecture decisions, product judgment, and novel problem framing dominate.

Cybersecurity vulnerability analysis as the Mythos findings directly demonstrate. Security flaws either exist or they do not. The training data from CVE databases, security research papers, and open source code is extensive. The work is entirely digital. Junior security analysts doing routine vulnerability scanning and penetration testing face the most direct displacement pressure from systems like Mythos. Senior security architects making strategic decisions about system design and incident response are substantially protected by the steepness of the gradient at that level.

Routine financial auditing and accounting at the junior end. Financial statements either comply with accounting standards or they do not. The training data from decades of financial filings, audit workpapers, and regulatory guidance is vast. The work is entirely digital. Junior auditors checking reconciliations, flagging anomalies, and verifying standard disclosures are doing Pattern Matching Condition work. The Big Four accounting firms have already begun reducing junior hiring in anticipation of this transition.

Actuarial modeling at the routine end. Risk either meets standard criteria or it does not. The training data from decades of actuarial tables, claims history, and regulatory filings is extensive. The work is entirely digital. Junior actuaries applying standard mortality tables and loss models to standard insurance products are doing Pattern Matching Condition work. Senior actuaries designing novel products, advising boards, and interpreting regulatory changes are substantially protected.

Insurance underwriting for standardized products. A risk either meets underwriting criteria or it does not. The training data from decades of claims data and underwriting manuals is substantial. The work is entirely digital for standard personal and small commercial lines. Junior underwriters applying standard criteria to standard applications are highly exposed. Complex commercial underwriting requiring site visits, relationship management, and novel risk assessment is much safer.

Radiology and medical imaging support at the diagnostic end. An image either shows a pathology or it does not. The training data from decades of labeled medical images is now substantial enough that AI diagnostic tools already outperform human radiologists on specific conditions in controlled settings. The work is entirely digital in the sense that the radiologist interacts with images on a screen. This is the domain where the Liability Requirement is doing the most work to keep humans in the loop, but even with that protection the displacement pressure on junior radiologists is real and accelerating.

Tax preparation for standard returns. A tax return either correctly applies the code or it does not. The training data from decades of tax filings, IRS guidance, and tax court decisions is vast. The work is entirely digital. H&R Block style tax preparation for standard W-2 returns is already substantially automatable. Complex tax strategy for high net worth individuals and corporations involving novel structures, regulatory interpretation, and relationship management is substantially protected.

Pharmaceutical drug interaction checking at the routine end. Two drugs either have a known interaction in the literature or they do not. The training data from pharmacological research, prescribing information, and adverse event databases is extensive. The work is entirely digital for standard interaction checking. Clinical pharmacists advising on complex polypharmacy cases for critically ill patients are substantially protected by the contextual judgment required.

Content moderation at the classification end. Content either violates a defined policy or it does not. The training data from billions of labeled examples is the most extensive of any category on this list. The work is entirely digital. This category is already heavily automated and represents the leading edge of what Pattern Matching Condition displacement looks like at scale.

The Jobs That Fail At Least One Condition

The framework is equally useful for identifying jobs that look exposed but are actually substantially protected. Electricians have binary outcomes, the circuit either works or it does not, but fail the digital manipulability condition because diagnosing a specific wiring configuration in a specific building with specific history requires physical embodied judgment. Therapists have extensive training data but fail the binary outcomes condition because therapeutic progress is not verifiable against an automated oracle. Surgeons fail the digital manipulability condition for the same reason as electricians despite having extensive training data and often binary outcomes. Teachers fail all three conditions in meaningful ways: learning outcomes are not binary, the relational and contextual dimensions of teaching are not digitally manipulable, and the most important knowledge transmitted in classrooms is tacit and embodied rather than documented.

The framework also clarifies why the most alarming displacement predictions focus on the wrong occupations. Truck drivers are frequently cited as the canonical automation casualty. But truck driving fails the digital manipulability condition significantly. Real world driving involves constant embodied judgment about physical environments, weather, mechanical conditions, and novel situations that do not reduce to pattern matching against a binary oracle. The timeline for full autonomous trucking has been consistently overpredicted for exactly this reason.

What the Framework Tells Us About Timelines

The Pattern Matching Conditions framework produces a specific and bounded displacement prediction. The jobs that satisfy all three conditions employ roughly 15 to 20 million Americans at the junior and routine end of their respective pyramids. That is significant but it is not the economy. It is a specific demographic, educated urban professionals in the 25 to 45 age range in pyramid-shaped knowledge work professions, facing genuine structural displacement pressure over a 10 to 20 year timeline.

The third piece in this trilogy will explain why even that specific and bounded displacement, real as it is, does not produce mass unemployment. The absorption mechanisms that protect the broader labor market operate differently for Pattern Matching Condition jobs than for broader cognitive work, but they still operate. The framework tells you where to look for displacement. The economic dynamics tell you why displacement is not the same as unemployment.

About the Author

Sean Richey, Ph.D., is a Professor of Political Science at Georgia State University specializing in AI information environments and digital political communication.

Expert Witness & Consulting Services

Dr. Richey provides expert witness testimony, case review and analysis for counsel, survey methodology evaluation, and policy consulting on AI-associated information environments. Visit my website or email consulting@seanrichey.com.