Chapter 20 - Teaching Machines to Listen for Echoes

In 1981, testing whether a specific learned behavior was becoming more common across geographically separated rat populations meant physically shipping data - trial counts scribbled in lab notebooks - between institutions by mail, tabulating it by hand or with primitive mainframe statistical packages, and hoping your sample size, typically dozens to a few hundred animals, an entire research group's multi-year output, was large enough to detect an effect nobody had any way of estimating in advance. Today, a single graduate student, using publicly available datasets and open-source machine learning libraries running on a laptop or a modest cloud computing budget, can query behavioral records spanning millions of animals, run statistical tests that automatically correct for multiple comparisons and publication bias, and train models capable of detecting correlational patterns far too subtle and far too high-dimensional for any human analyst, however careful, to notice by eye.

This chapter opens Part IV with the case for why that shift in available tooling changes the terms of the entire morphic resonance debate - not by making the hypothesis more likely to be true, but by making it, for the first time in its history, genuinely and rigorously testable at the scale the question actually demands. It then turns to the specific technology most readers will have some intuitive familiarity with already: large language models, and whether fine-tuning them on historical behavioral, morphological, or cultural datasets could reveal genuine "resonance signatures" that classical statistical methods would have missed entirely.

The Four-Decade Instrumentation Gap

It's worth being precise about exactly what has changed, because "technology has improved" is a vague truism that undersells how specifically relevant the recent changes are to this particular hypothesis.

Data scale has changed by many orders of magnitude. Sheldrake's original rat-learning and staring-detection studies worked with sample sizes in the hundreds to low thousands of trials, accumulated over years of painstaking manual data collection. Today, behavioral datasets from citizen-science platforms, mobile applications, wearable sensors, and digitized historical archives routinely contain millions to billions of individual data points, and the statistical power to detect small effect sizes - exactly what morphic resonance, if real, would most plausibly produce, given four decades of evidence pointing to modest rather than dramatic effects - scales with the square root of sample size. A dataset a million times larger doesn't just make a small effect a little easier to detect; it makes it a thousand times easier to distinguish from noise, a qualitative rather than merely incremental improvement in what's detectable at all.

Computational pattern recognition has changed even more dramatically in kind, not just in scale. The statistical tools available to Sheldrake and his critics in the 1980s and 1990s were almost exclusively designed to test a single, pre-specified hypothesis against a single, pre-specified pattern - a t-test comparing two group means, a chi-squared test comparing observed to expected frequencies. Modern machine learning, particularly the unsupervised and self-supervised methods examined in Chapter 21, is specifically designed to search enormous, high-dimensional datasets for statistical structure no human analyst specified in advance. This matters enormously for a hypothesis whose core claim - resonance strength depends on multidimensional "similarity" between morphic units - is inherently a claim about high-dimensional pattern structure, not a simple, single-variable comparison.

Automation of experimental delivery has changed the entire methodological landscape of exactly the human-subject research - Chapter 9's staring experiments, Chapter 9's telephone telepathy studies - that has historically been most vulnerable to the experimenter-cuing confounds examined in detail in Part II. Fully automated, app-based, sensor-verified protocols, deliverable to a globally distributed subject pool with no live human experimenter in the loop, were logistically inconceivable in the 1990s and are now a matter of building, or more often adapting, an existing smartphone application.

Simulation and synthetic-world modeling have advanced to the point where researchers can build entire agent-based digital ecosystems and observe, computationally, exactly what patterns would emerge under a morphic-resonance-like inheritance rule versus a purely conventional one, over thousands of simulated generations, in an afternoon rather than a multi-year breeding program - letting researchers generate precise, falsifiable predictions about what a real effect should look like before touching real data, dramatically reducing the researcher-degrees-of-freedom problem from Chapter 12.

Recall the crossword-puzzle thought experiment from Chapter 1. In 1981, running it properly would have meant recruiting ten thousand participants by mail, distributing physical puzzles, and collecting solve-time data by hand - a logistical undertaking easily costing more than an entire research grant, with no realistic way to guarantee true isolation between participants at that scale. Today, that exact experiment could run through a smartphone puzzle-game app with an existing user base, at the marginal cost of app development: solve times captured to the millisecond, participant isolation verifiable through usage logs and social-network metadata, and sample sizes in the hundreds of thousands rather than thousands. This is not hypothetical infrastructure. It exists today, built for entirely unrelated commercial purposes, sitting unused for this specific scientific question purely because nobody with the right combination of technical skill and institutional support has yet pointed it here with the rigor this book has argued the question deserves.

None of this changes the prior plausibility of morphic resonance as a physical hypothesis - the concerns raised in Chapter 16 remain exactly as serious as before, and nothing in this Part should be read as walking them back. What has changed is our capacity to actually run the kind of decisive, well-powered, methodologically airtight experiments that both Sheldrake and his critics have, in good faith, called for across four decades of talking past each other.

One caution before the technical material begins. New methodological capabilities have a long history of generating a wave of enthusiastic, under-scrutinized findings before the community learns how to apply appropriate skepticism to the new method itself - genomics in the early 2000s produced a wave of "gene for X" findings that mostly failed to replicate once genome-wide association study methodology matured; neuroimaging produced a wave of "brain region for X" findings that suffered from exactly the small-sample, multiple-comparisons problems of Chapter 12, before the field developed appropriate corrections. The tools described across the next seven chapters are genuinely powerful. Hold onto everything Chapter 12 taught you about the file-drawer problem as you read them, because powerful new tools amplify a careless researcher's capacity to fool themselves exactly as readily as they amplify a careful researcher's capacity to find something real.

What a Language Model Actually Learns

With that case made, we turn to the first tool: large language models. A transformer-based language model, at its technical core, is a system trained to do one deceptively simple thing extraordinarily well: predict the next piece of text - a "token," typically a word fragment - given everything that came before it. From this simple training objective, applied to enormous quantities of text, emerges a system capable of an astonishing range of downstream capabilities - translation, summarization, multi-step reasoning - none of which were directly, explicitly specified in training. This section asks a specific question: can this same technology, redirected at historical and behavioral datasets rather than internet text, detect the kind of accelerating-pattern-acquisition signature that would constitute genuine evidence for or against morphic resonance?

A transformer is trained by adjusting millions to billions of internal numerical parameters so that, given a sequence of preceding tokens, the model assigns high probability to whatever token actually comes next in its training data, and low probability to tokens that don't. Repeated over enormous datasets - modern frontier models train on trillions of tokens, roughly corresponding to many millions of books' worth of text - this process forces the model to internally represent an enormous amount of statistical structure: not just word frequencies, but complex, long-range, hierarchical patterns of grammar, factual association, and something resembling structured world knowledge, because accurately predicting text about the world turns out to require implicitly encoding a great deal about how the world actually works.

The key technical concept for our purposes is perplexity, a formal measure - directly related to the Shannon entropy introduced in Chapter 17 - of how "surprised" a trained model is by a given piece of text. Mathematically, perplexity is the exponentiated average negative log-probability the model assigns to the actual sequence of tokens it observes. A model trained extensively on a specific type of pattern shows low perplexity when it encounters a new instance of that pattern; a model with no prior exposure shows high perplexity. This gives us, for the first time, a rigorous, quantitative, model-based operationalization of the intuitive concept at the heart of morphic resonance - how easy or surprising is this specific pattern, given everything that has come before it - in a form that can actually be measured, computed, and compared across enormous datasets, rather than argued about impressionistically.

The Core Proposal: Historical Perplexity Trajectories

Here is the concrete research proposal, stated as precisely as I can manage. Take a large, carefully dated historical corpus - newspaper archives, patent filings, scientific publications, folklore collections, or digitized field-naturalist observation logs spanning many decades - and train a sequence of language models, each restricted to data available up through a specific historical date, in a rolling-window fashion: a model trained only on data through 1950, another through 1960, another through 1970, and so on. Then, for a specific target pattern - say, the first documented instance of a particular technique, idiom, or behavioral report - measure how that pattern's perplexity, under models trained on data available up to increasingly later dates, changes as the number of independently documented prior instances of similar patterns increases.

The mundane, expected outcome - the null hypothesis a fair experiment must be designed to distinguish from anything more exotic - is that perplexity decreases for entirely conventional reasons: growing vocabulary and cultural familiarity with a general category of pattern. The interesting, non-mundane question is whether perplexity decreases for a specific, narrowly matched pattern faster than a rigorously constructed statistical control would predict, where the control accounts for the model's growing general sophistication, the pattern's overall category frequency, and any documented instance of direct cultural transmission between the compared instances.

Designing the Control Condition Properly

This is the single most important methodological detail in the entire chapter: the mundane explanation - models get generally better at predicting a whole category of pattern as they see more examples - is so powerful and so certain to operate to some degree that any claimed "resonance signature" must be shown to exceed it by a rigorous margin, using a properly matched control, or the result is worthless. A defensible design needs, at minimum, three things: a large set of matched control patterns, structurally similar to the target but with no plausible resonance-relevant connection to it; explicit modeling and statistical removal of direct citation or transmission chains, so that apparent "acceleration" isn't just ordinary documented cultural copying; and a pre-registered, quantitative prediction of exactly how much faster the target pattern's perplexity should decrease under morphic resonance than the matched controls, ideally derived from a formal agent-based simulation of the kind built in Chapter 22. That third element matters because almost any real dataset will show some noisy version of "faster than control." The comparison needs to be sharper than that - faster than control by an amount consistent with a specific, quantitatively predicted effect size.

A Second Application: Cross-Linguistic Convergence

A second, related application involves cross-linguistic and cross-cultural pattern convergence, where language models' multilingual training gives them a genuinely novel comparative advantage over any pre-existing statistical method. Multilingual models, trained jointly across dozens to hundreds of languages, learn shared internal representations that let them recognize structurally similar patterns across languages with no shared vocabulary or common linguistic ancestry - providing, for the first time, a scalable tool for asking whether specific folkloric motifs or symbolic patterns (recall Jung's archetypes from Chapter 5) show a higher degree of cross-cultural convergence than a rigorously specified null model, accounting for known trade routes, migration, and documented contact, would predict. Comparative folklorists have grappled with exactly this question - independent invention versus diffusion - for over a century using laborious manual methods; embedding-based, LLM-powered comparative analysis, applied to existing digitized world folklore corpora, offers a rigorous way to revisit it at a scale no individual human comparative scholar could match.

The Serious Pitfalls to Watch For

Training data contamination is the single largest risk: modern language models are frequently trained on enormous, poorly documented scrapes of internet text that may already contain indirect discussion or direct copies of the very historical patterns being studied, meaning an apparent "resonance signature" could simply reflect the model having been pre-exposed, during its original training, to exactly the cross-referencing material the experiment is trying to control for. This makes it essential to use models trained from scratch on a carefully curated, fully documented, date-restricted corpus, rather than convenient pre-trained models whose training data provenance is often opaque.

Anthropomorphizing perplexity is a subtler risk: it is tempting, but not fully justified, to treat a language model's perplexity as directly equivalent to some biological or physical system's "resonant ease," when perplexity measures a specific artificial neural network's statistical surprise given its own architecture, which may or may not correspond to whatever process is actually happening in the real-world system the corpus describes. Perplexity-based methods are best understood as a rigorous instrument for detecting anomalous correlational structure in the documentary record, not as direct evidence about underlying causal mechanism.

Questions for the Curious

• Design a rolling-window language model training pipeline using a well-documented, fully-dated public domain corpus - Google Books Ngram data, or a digitized newspaper archive with clear publication dates - and a matched-control methodology as described above. Could you detect any category of pattern whose perplexity decreases faster than matched controls, after correcting for citation and direct transmission? What effect size would need to survive a false-discovery-rate correction across all the pattern categories you test?

• Using an existing digitized world folklore or mythology database, apply multilingual embedding models to quantify cross-cultural motif similarity, and build a formal statistical null model incorporating known historical trade and migration data. Does any specific motif category show convergence exceeding what documented contact could explain?

• Historical patent and scientific citation databases are exceptionally well-dated. Could a perplexity-trajectory analysis, applied to independently filed but conceptually similar inventions with no citation relationship, detect anything beyond what's explained by the shared technological "adjacent possible" - a concept we examine in Chapter 25 - available to inventors at a given historical moment?

• Draft a full pre-registered adversarial collaboration protocol for this method: a skeptic and a proponent jointly specifying, in advance, exactly what perplexity-trajectory result would count as meaningful evidence either way, before any model is trained. Write it as a mock registered report, following the format introduced in Chapter 13.

Looking Forward

Language models give us a genuinely new instrument for measuring "surprise" and pattern-familiarity at a scale and precision no previous statistical tool could match, provided the serious methodological pitfalls around training data contamination and control-condition design are taken as seriously as this chapter has insisted they must be. Chapter 21 broadens the toolkit considerably, moving from text-specific language models to the wider world of unsupervised and self-supervised machine learning methods designed to search vast, heterogeneous scientific databases for exactly the kind of non-local correlational structure a human analyst, however careful, would have no realistic way of noticing unassisted.