# Against Reverse Inference **Why honest neuroscience is the defensible position, and the design principle behind a live brain that refuses to overclaim** Neuro ASI / Phroneme · Draft for review · 2026 --- ## Abstract The most common move in popular neuroscience is also its most common error: a region "lights up" during a mental state, and the activation is read backward as proof that the state occurred. This is reverse inference, and it is not a rhetorical quibble but a documented logical limit. This paper states the limit precisely, shows through meta-analysis that emotions do not occupy dedicated regions, marks the ceiling of what a functional localizer licenses even where specialization is real, states the honest limit of the correspondence between deep networks and visual cortex, situates the error inside the reproducibility problems of neuroimaging, and reads the whole pattern historically as the latest instance of describing the brain in the image of the era's newest machine. We then draw the practical conclusion. An interface that types a feeling and asserts what the brain is doing is committing the field's signature mistake at product scale and inheriting the exact legal exposure that has punished overclaim before. The defensible design does the opposite: it surfaces associations, attaches confidence and citation, and names the caveat on screen. Honesty here is not caution for its own sake. It is the position least likely to age into embarrassment and most likely to earn institutional trust. ## 1. The seduction of the lighting-up brain A brain image with a bright spot is one of the most persuasive objects in modern communication. It converts a claim about the mind into what looks like a photograph of a fact. The persuasion is real and the fact usually is not, because the inference running underneath the image is invalid in a specific and well-characterized way. ## 2. What reverse inference is Forward inference asks: given this mental state, what happens in the brain. Reverse inference runs the other way: given this brain activity, what mental state is present. The reverse direction is not deductively valid. Its evidential value depends entirely on how selective the region is, that is, on how much more likely activation is under the process of interest than under everything else the region also does. Where a region is engaged by many processes, and most are, observing its activation licenses almost nothing about which process occurred (Poldrack, 2006). "Region X is responsible for feeling Y," stated flatly, is nearly always this error. ## 3. Emotions do not live in single regions The point is not abstract. A meta-analysis of the neuroimaging of emotion compared the intuitive "one emotion, one region" model against the view that emotions are constructed from general-purpose networks, and the general-network account won. Discrete emotions did not map onto dedicated regions. The amygdala, popularly the brain's "fear center," is not fear-specific; it responds to salience across many states, including sadness (Lindquist et al., 2012). Any interface that assigns a feeling to a region as though the mapping were fixed is contradicted by the field's own meta-analytic record. ## 4. What a functional localizer does and does not license The strongest case for regional specialization is not the amygdala, it is the face area. A functional localizer, showing faces against objects and taking the contrast, reliably identifies a patch of fusiform cortex that responds far more to faces than to almost anything else, in nearly every healthy participant (Kanwisher et al., 1997). The same logic identified a parahippocampal region that responds to scenes and places (Epstein and Kanwisher, 1998). These findings are real, they replicate, and they are about as close to modular as the human cortex gets. They are also the ceiling of what localization delivers, and the ceiling sits lower than it looks. The first limit is that a region being present does not mean its function is intact. A localizer answers one question: does this patch of cortex respond more to faces than to objects, in this person, right now. It does not answer whether that person can recognize a face. The two come apart, which is why face recognition is assessed with behavioral tests of face recognition rather than with a picture of the fusiform gyrus. Anatomy is not capacity, and an image of a present region is not evidence of a working one. The second limit is that resolution does not get you causality. A sharper picture of where activity occurs is still a picture of correlation. Better hardware narrows the region and shortens the interval, but it does not establish that the region performs the process, because that was never what the measurement was answering (Poldrack, 2006). This matters commercially because it is the upgrade path a hopeful product tends to promise. There is no resolution at which correlation becomes cause, so the limit bounds the product rather than its roadmap. ## 5. The honest ceiling of brain-AI correspondence The strongest empirical link between artificial networks and the brain deserves stating precisely, because the precise version and the popular version point in opposite directions. Human observers categorize a novel natural scene fast enough that the response leaves little room for extended feedback, which is the evidence that one narrow visual task, core object recognition, runs largely feedforward (Thorpe et al., 1996). That is the task deep networks were built for, and on that task the correspondence is real: networks optimized for object recognition predict responses in higher visual cortex better than models designed by hand, accounting for roughly half of the explainable variance in inferior temporal cortex, for core object recognition only (Yamins et al., 2014). The link has since been tested causally rather than correlationally. Images synthesized from such a model drove targeted V4 neurons beyond the range natural images produce, which is a stronger result than a good fit (Bashivan et al., 2019), though the control was partial and confined to one area of the visual system. The same laboratories report where it fails. The images that feedforward networks classify incorrectly are the ones the primate ventral stream solves with an additional tens of milliseconds of processing, implicating recurrent circuits that the feedforward models do not contain (Kar et al., 2019). The field's best models are incomplete by the field's own test, on the single function they were built for. Two consequences follow for a product. The first is that "AI works like the brain" is not what this literature says. It says that one class of model matches one visual computation about half as well as the data permit, and misses a component the brain demonstrably uses. The second is a level confusion that predates the machine-learning era. Marr distinguished the computation being performed, the algorithm that performs it, and the physical implementation, and observed that a description at one level does not hand you the others (Marr, 1982). A measurement of implementation, whether it is a voltage at the scalp or an activation inside a network, is not a description of the computation. That is the same limit as reverse inference, arriving from the modeling side rather than the imaging side. It is why this project does not claim to read cognition from a model, from a screen, from a typed sentence, or from a questionnaire, and why the instrument measures behavior and treats any neural signal as convergent evidence only. ## 6. The reproducibility layer Even the forward direction is shakier than the public believes. Modestly sized task-based imaging studies replicate poorly (Turner et al., 2018). When seventy teams analyzed one dataset, analytic choices alone produced materially different conclusions (Botvinik-Nezer et al., 2020). Stable brain-behavior correlations appear to demand samples in the thousands (Marek et al., 2022). A method this sensitive to pipeline and sample cannot support confident point claims about an individual's inner state, and the caution compounds the reverse-inference limit rather than replacing it. ## 7. The pattern is historical Each era has described the brain as its most advanced machine, declared the mind nearly solved, built practice and industry on the description, and watched it date. Hydraulics gave "pressure" and "drives." The telephone switchboard gave "wiring" and "circuits." The computer now gives "processing," "encoding," and "decoding." Phrenology mapped character to skull bumps with commercial confidence before it collapsed. The most reliable prediction about today's computational framing is that it, too, is a metaphor with a shelf life. The defensible intellectual stance is to hold the framing as a hypothesis under examination, not as settled ground. ## 8. Honesty as position, not caution There is a direct commercial reading. A prior brain-training company paid a two-million-dollar federal settlement for marketing claims that outran its evidence (FTC, 2016). Overclaim is not only intellectually weak, it is a documented liability. The counter-move is not to say less; it is to make candor the product. Publishing limits, intervals, and failure modes is a position a marketing-led competitor structurally cannot copy, because copying it would indict the competitor's own claims. In a category defined by overreach, the honest instrument is the differentiated one. ## 9. Design implication: an instrument that refuses to overclaim The live brain in this project is built as the applied form of this argument. A user enters a feeling or a metaphor. The system returns the regions the literature *associates* with that class of input, never an assertion about the user's actual brain. Every region carries an explicit confidence band, capped below certainty on purpose. Every response carries the reverse-inference caveat, attached by the system rather than left to the model's discretion. The neural interpretation is framed as probabilistic association with a citation attached, which is the exact inverse of the lighting-up-brain fallacy. The visualization keeps the persuasive power of a brain that responds to language while removing the false claim that usually rides along with it. ## 10. Conclusion Reverse inference is the field's signature error, emotions do not sit in dedicated regions, the imaging that fuels the error is itself fragile, and the whole configuration is a repeat of a two-century pattern. A product that ignores all of this can look impressive briefly and then age badly, or invite a regulator. A product that builds the limits into its interface earns the one asset that compounds in a low-trust field. For a measurement company whose entire value rests on being believed, refusing to overclaim is not the safe choice. It is the winning one. --- ## References 1. Bashivan, P., Kar, K., and DiCarlo, J. J. (2019). Neural population control via deep image synthesis. *Science, 364*(6439), eaav9436. 2. Botvinik-Nezer, R., Holzmeister, F., Camerer, C. F., et al. (2020). Variability in the analysis of a single neuroimaging dataset by many teams. *Nature, 582*(7810), 84-88. 3. Epstein, R., and Kanwisher, N. (1998). A cortical representation of the local visual environment. *Nature, 392*(6676), 598-601. 4. Federal Trade Commission (2016). Lumosity to Pay $2 Million to Settle FTC Deceptive Advertising Charges. FTC Press Release. 5. Jonas, E., and Kording, K. P. (2017). Could a neuroscientist understand a microprocessor? *PLOS Computational Biology, 13*(1), e1005268. 6. Kanwisher, N., McDermott, J., and Chun, M. M. (1997). The fusiform face area: A module in human extrastriate cortex specialized for face perception. *The Journal of Neuroscience, 17*(11), 4302-4311. 7. Kar, K., Kubilius, J., Schmidt, K., Issa, E. B., and DiCarlo, J. J. (2019). Evidence that recurrent circuits are critical to the ventral stream's execution of core object recognition behavior. *Nature Neuroscience, 22*(6), 974-983. 8. Lindquist, K. A., Wager, T. D., Kober, H., Bliss-Moreau, E., and Barrett, L. F. (2012). The brain basis of emotion: A meta-analytic review. *Behavioral and Brain Sciences, 35*(3), 121-143. 9. Marek, S., Tervo-Clemmens, B., Calabro, F. J., et al. (2022). Reproducible brain-wide association studies require thousands of individuals. *Nature, 603*(7902), 654-660. 10. Marr, D. (1982). *Vision: A Computational Investigation into the Human Representation and Processing of Visual Information.* W. H. Freeman (MIT Press edition, 2010). 11. Poldrack, R. A. (2006). Can cognitive processes be inferred from neuroimaging data? *Trends in Cognitive Sciences, 10*(2), 59-63. 12. Thorpe, S., Fize, D., and Marlot, C. (1996). Speed of processing in the human visual system. *Nature, 381*(6582), 520-522. 13. Turner, B. O., Paul, E. J., Miller, M. B., and Barbey, A. K. (2018). Small sample sizes reduce the replicability of task-based fMRI studies. *Communications Biology, 1*, 62. 14. Yamins, D. L. K., Hong, H., Cadieu, C. F., Solomon, E. A., Seibert, D., and DiCarlo, J. J. (2014). Performance-optimized hierarchical models predict neural responses in higher visual cortex. *Proceedings of the National Academy of Sciences, 111*(23), 8619-8624.