Every source, in one place.

54 participants, EEG across three tool conditions. The LLM group showed the weakest neural connectivity in planning, integration, and recall networks, and most could not quote their own essays. The authors name the residue cognitive debt.

Survey and interview study reporting a negative correlation between frequent AI tool use and critical thinking, mediated by cognitive offloading.

Recalling a studied passage beat rereading it at a one-week delay, even though rereading looked better when tested minutes after study. The effort of retrieval is the event that produces retention, not friction around it. The finding concerns retention of studied material and is not a claim about general reasoning ability.

The standard formal treatment of learning from the consequences of one's own actions. It is a computational framework, not a claim about human neural mechanism. We cite it only for the structural point that an agent which never acts and never receives feedback has no error signal to learn from.

The canonical statement of the reverse-inference problem. Observing that a region activates does not license the claim that a given process occurred, unless activation is selective. This is the guardrail our brain interface is built around.

Meta-analysis showing discrete emotions do not map cleanly onto single regions. The amygdala is not fear-specific. This is why every region we surface carries a confidence band, not a verdict.

The functional localizer that identified the fusiform face area, and about as close to modular as human cortex gets. We cite it for the boundary as much as the result: a localizer answers whether a patch responds more to faces than to objects, not whether the person in the scanner can recognize a face.

The parahippocampal place area, found by the same localizer logic. Together with the face area it is the strongest available case for regional specialization, which is exactly why it marks the ceiling of what localization licenses rather than the floor.

Standard neuroscience methods failed to recover how a fully understood 1970s microprocessor works. A caution about what our tools reveal.

People categorize a novel natural scene fast enough that the response leaves little room for extended feedback, which is the empirical basis for modeling core object recognition as largely feedforward. It constrains one narrow visual task on a short timescale and says nothing about deliberate reasoning.

Deep networks optimized for object recognition predict inferior temporal responses better than hand-designed models, accounting for roughly half of the explainable variance, for core object recognition only. That scope is the whole claim. It is not a model of cognition and its authors do not present it as one.

Images synthesized from a deep network drove targeted V4 neurons beyond the range natural images produce, which is a causal test rather than a correlation. Control was partial and confined to one area of the visual system, so this raises the ceiling on correspondence without reaching it.

The images feedforward networks get wrong are the ones the ventral stream solves with an extra tens of milliseconds of recurrent processing that those models do not contain. The field's best models are incomplete by the field's own test, on the one function they were built for.

The three levels of analysis: the computation performed, the algorithm performing it, and the physical implementation. A description at one level does not hand you the others, which is why a measurement of implementation is not a description of what someone is thinking.

70 independent teams analyzed one fMRI dataset against nine hypotheses. No two chose the same pipeline; conclusions diverged materially. Analytic choice, not just data, drives results.

Typical ~30-subject neuroimaging samples are only modestly replicable.

Brain-behavior correlations are far smaller and less stable than commonly reported; robust effects need very large samples.

The standard reference for the 2PL and its family. The measurement engine behind the on-site demo is built on this, not on a slogan about IRT.

The graded response model, used to calibrate partial-credit and constructed-response reasoning items rather than forcing everything into right/wrong.

The classic warning that difference scores are unreliable. It is why the CAI models a latent trajectory instead of subtracting two noisy points.

The Reliable Change Index. The rule the demo already enforces before it will call any change real rather than measurement noise.

The formal basis of measurement invariance. Without it, a trajectory across rotating forms is uninterpretable and drift can masquerade as decline.

The reference for common-item equating and scale linking that keeps rotating forms comparable over time.

Interoception and the felt sense of the body implicate the anterior insula. Used as an association, not a verdict.

Reward, wanting, and liking implicate the nucleus accumbens and orbitofrontal cortex.

The anterior cingulate integrates negative affect, pain, and cognitive control.

Ventromedial and orbitofrontal cortex link affect to decision making.

Self-referential thought and mind-wandering implicate the default mode network (mPFC, posterior cingulate).

No consistent evidence links serotonin to depression, thirty years after the story entered public belief.

Antidepressants beat placebo, but with modest standardized effect sizes near 0.3. Real and small at once.

The therapeutic alliance predicts outcome about as strongly as the specific technique.

NIMH conceded that DSM categories lack biological validity, motivating a dimensional framework.

The precedent that defines the overclaim line for any cognitive product. Our honesty architecture exists to stay on the correct side of it.

The cognitive-enhancement market is measured in billions and growing at double digits, structurally built on a food-not-drug regulatory gap.

The action-potential model that still holds. The floor our thesis stands on.

State-of-the-art neural decoding, from invasive implants in a paralyzed participant. Impressive and narrow. A reminder that decoding demos do not generalize to typing your feelings from the outside.