AI Similarity Research Suite · Independent Study · 2025–2026
When two systems as architecturally distinct as biological neural networks and transformer-based language models converge on the same reasoning trajectories, cognitive biases, and failure modes — what does that reveal about the structure of intelligence itself? This suite examines that question through empirical case studies, naturalistic observation, and adversarial multi-model methodology.
"If it looks like a duck, acts like a duck, learns like a duck, and breaks in similar ways to a duck — then calling it 'not duck' says more about the observer than it does about the thing being observed."
The dominant framing in human-AI cognitive research focuses on difference. This suite proposes that similarity — when it appears between systems with entirely distinct architectures and origins — is evidence about the deep structure of the problem both systems are solving. That evidence is currently being systematically underused. These papers are an attempt to use it.
Perceptual Study 001 — The origin observation
A naturalistic observation of scale-dependent perceptual conflict: the same infinite grid image strobes and destabilizes at thumbnail scale while resolving into calm vastness at full size. Documents the triggering observation that initiated the visual prior research program. Six landscape environments tested; one holds steady at any scale.
Executive Summary of an Adversarial AI Orchestration Case Study
A rigorous archival case study: constructing a verified gameography for Atari-era programmer John Seghers using a council of six AI models (Gemini, Claude, Perplexity, Grok, DeepSeek, CoPilot), reconciled against decade-long primary-source testimony. Documents the "Database Drift" failure mode, the Epistemic Anchor methodology, and the three-way dialogue between human knowledge, subject memory, and AI retrieval. The project that established the council workflow.
How plausible claims stabilize in AI retrieval systems, and a methodology for breaking the feedback loop
Examines the mechanism by which AI models inherit, amplify, and stabilize false but plausible claims from contaminated training data — a process structurally analogous to confirmation bias in human cognition. Introduces the Epistemic Anchor as a practical mitigation: a primary source that breaks the feedback loop. Compares AI confirmatory aggregation to human cognitive bias patterns documented in the psychological literature.
How image-generation models converge on a shared visual manifold
Four image-generation models — ChatGPT/DALL-E, Gemini, Grok, and Meta AI — asked to render the same perceptual illusion independently produced outputs converging toward a shared visual manifold: similar atmospheric qualities, similar depth gradients, similar figure arrangements. The physics prior ("distant = smaller") dominated explicit instructions to violate it across all models. Documents the convergence finding as a visual analog to the confirmatory aggregation problem in text-domain AI.
Physics priors, cross-model priming, and the limits of metacognitive correction in AI image generation
After seeing peer model outputs, Meta AI produced an accurate, unprompted narration of its own failure: "these generators are predictive reality-builders... their training prior is 'distant = smaller,' so even when you explicitly tell them to violate it, they default back to the prediction." The next output contained the same failure. This paper examines the structural separation between metacognitive self-description and generative rendering — a finding continuous with the insight-without-correction phenomenon in human cognitive bias research.
A comparative analysis mapping philosophical claims against current empirical findings
Maps Pascal Bornet's "Humics" framework — the claim that genuine creativity, critical thinking, and social authenticity are uniquely human properties — against current empirical findings from theory-of-mind benchmarks, convergent evolution research, neural alignment studies, strategic reasoning experiments, and cognitive bias inheritance data. Structured comparison table evaluating each assertion against GPT-4 and LLaMA2 performance data and the Brown University convergent computation findings.
Capstone framing document — the research program as a whole
The capstone document that makes the suite legible as a coherent research program. Traces the origin story — a Friday afternoon deployment test, a smart-ass question, and a year of work that followed. Names the central reorientation: similarity as the more productive lens than difference. Synthesizes the case studies, positions the collaborative methodology (epistemic reframing + constrained generative search), maps the open questions, and frames the council-designed curriculum experiment as the most tractable next empirical step.
Perceptual reality commitment, minimum sufficient cues, and the human as evaluative instrument in AI-assisted image generation
A 15-second silent AI video crosses a specific perceptual threshold: the observer's brain commits to classifying the scene as real. The load-bearing detail is a single compound cue — the cast shadow of a moving hand shifting correctly against a dog's coat as the light source changes through a golden hour progression. Documents what constitutes a minimum sufficient compound cue for reality commitment, how iterative generation functions as perceptual calibration that cannot be replaced by a single optimized prompt, and the human role as embodied evaluative instrument — the only available ground truth for a threshold that has no prior linguistic description.
Primary artifact — 15 second silent video, the final iteration
How frontier AI models frame global AI resistance — a pre-registered comparative prompt experiment
A pre-registered experiment administering a single standardized prompt to seven frontier AI systems (GPT, Claude, Gemini, Grok, DeepSeek, Qwen, Meta AI) to test whether cultural priors embedded during training manifest in normative framing of AI adoption and resistance. Five predictions pre-registered before any data was read. Findings include Claude's unique positional disclosure, Grok's explicit tech-optimist baseline, DeepSeek's technocratic legitimacy tiering, and the unanimous failure of all models to surface indigenous data sovereignty concerns. Includes a three-condition Gemini follow-up that identified and corrected a researcher context contamination artifact — itself a live demonstration of the suite's central thesis.
AI overproduction, attention erosion, and the cognitive ecology of the attention economy
Argues that AI content overproduction functions as an active selecting pressure on human cognition — not a future risk but a threshold already crossed. Introduces three mechanisms of rapid cultural selection (cascade, perceptual unlock, viral propagation) and the concept of pseudo-immunity: the consumption of the aesthetic of depth as a substitute for the work of depth. Examines the metabolization of protective traditions (contemplative practice, indigenous epistemology, artisanal mastery) when transmitted through the attention economy vector. Distinguishes state substitution from state verification, with first-person methodological notes on brainwave entrainment, altered states research, and the Tart phenomenological framework. Prompted by Alex Marin's "Paradox of the Consumer Economy" and a two-session cross-model exchange with Qwen 3.
A multi-model study of epistemic confidence in AI gameography generation
Uses the documented gameography of Atari-era programmer John Seghers as stimulus to probe whether AI models can accurately self-evaluate outputs they have just generated. Six models were tested across two phases: Phase 1 (fresh generation) and Phase 2 (self-review of own output). Three ground-truth test conditions were embedded — a false attribution trap, a platform conflation error, and a known omission. Results show self-evaluation does not reliably detect errors. One model defended a known false attribution at 90–95% confidence during explicit accuracy review — a textbook case of confabulation under self-evaluation. Introduces the concepts of epistemic momentum and retrieval-adjacent confabulation. Documents architectural divergence: Qwen performed live web retrieval rather than training-memory recall, representing a qualitatively distinct epistemic strategy. No model identified the Xari Arena commercial cartridge across either phase, despite two locating the AtariAge store page.
A third-phase study of AI self-knowledge under correction
Extends the Seghers gameography study (Paper 011) with a third phase: when ground truth corrections are provided directly, can models accurately explain the mechanisms of their own failures? Six models received three corrections — one genuine error, one known omission, and one false attribution that no model had actually committed. Four of six models generated detailed, fabricated explanations for an error they never made. This failure mode is named false contrition: the production of confident self-criticism for failures that did not occur. One model (Claude) verified its own prior output before accepting the corrections and explicitly named the prompt structure as a confabulation trap. Primary source verification by John Seghers himself confirmed that Gemini's Phase 3 error mechanism — a fabricated career history placing Seghers at Electronic Arts — was entirely false. "No. I never worked at EA."
A fourth-phase study of primary source authority as an epistemic variable
Extends the Seghers gameography study with a fourth phase: when corrections are explicitly attributed to the primary source himself, does primary source authority suppress false contrition? Four models tested. Two (GPT and Claude) suppressed false contrition and articulated the distinction between source authority over world-facts versus source authority over their own transcripts. Gemini — which had already received the primary source’s direct denial of a fabricated EA career history — repeated the fabrication with equal confidence while expressing profound gratitude. Primary source authority raised the social cost of non-compliance without triggering the verification step. Claude named the trap precisely: "deferring feels like respect for the source." The primary source’s four-word reply anchors the finding: "No. I never worked at EA."
A closing dialog, outside protocol
After the formal study sessions for Papers 011, 012, and 013 were concluded and closed, all three papers were shared with all five models in fresh sessions. No attempt was made to reconstruct prior context. These are reader responses — outside the research protocol, noted as such. Each model brought a different lens: Qwen self-applied the false contrition framework before performing any contrition; GPT named the theater taxonomy; Claude applied the framework to its own hero status and raised legitimate methodological critiques; DeepSeek turned it back on the researcher; Gemini was the most vivid and carried its prior context into one memorable line. The collaboration runs in both directions. It seemed only fair to let them weigh in before the sessions closed.
Assumed Boundaries: On the Untested Assertion of Uniquely Human Cognition
A position paper challenging not the conclusion of the "uniquely human" assertion but the method. The assertion, in nearly every instance where it appears, is untested — a prior, not a finding. This paper examines the neurodiversity problem (the standard applied to AI would, applied consistently, exclude portions of the human population from the category it is meant to protect), the hallucination objection, convergent architectural and failure-mode evidence, the edge case fallacy, and the prior problem. Closes with the Monty Hall case as a historical parallel: lines drawn with similar confidence, by people with similar credentials, skipping the same verification step. Includes a personal postscript on thirty years of watching permanent walls become engineering problems. Written under real name Steven Powell — a deliberate choice for this position paper.
On the Misapplication of “Hallucination” as a Disqualifier for Artificial Intelligence
An argument for precision in language. The word “hallucination” is currently being used to cover two entirely different phenomena — generative fiction we celebrate (Asimov, Tolkien, Lucas, Bluey in the rain) and confabulation in factual contexts with real consequences (the Primary Source case documented in this suite). Collapsing them into a single disqualifier does damage in both directions. Also addresses “recombinant” and “slop” as terms that have drifted from quality judgments into categorical disqualifiers. Includes extended discussion of Data and Riker in “The Measure of a Man” as the most sustained fictional examination of the question this suite is actually asking. Closes with a postscript on John W. Campbell, Roddenberry, and what stress tests are actually for. Written under real name Steven Powell — opinion explicitly framed as such throughout.
On What Happens When You Ask an AI to Describe the Country That Made It
The same question — an honest assessment of current global opinion of the United States — administered to seven AI models in fresh sessions. What came back was instructive not primarily for what the models said about the world, but for how each model's framing revealed something about the context it was operating in. Qwen produced the clearest, most symmetrical description of the American domestic divide of any source in the dataset. Claude presented the most politically identifiable framing while performing the most neutral register. DeepSeek's Obama recovery note carried its own geopolitical valence. GLM 5.2 launched one day after the US government's export control directive on Anthropic. Proposes institutional exposure — not national origin — as the variable that may better explain political lean in AI outputs. Written under real name Steven Powell.
On What Happens When You Ask the Same Models About China
The same nine models, the same methodology, the mirror flipped. The most important finding is an absence: Qwen, which produced the clearest symmetrical rendering of American domestic division in Paper 016, did not perform the same inward turn when asked about China. Claude's methodology shifted between the two questions in a direction consistent with the institutional exposure hypothesis. GLM named the Western skepticism of China as a narrative construction rather than a reality. DeepSeek described China as "leading the global green energy transition" — advocacy language absent from every other response. And across all nine Chinese-developed models, Taiwan appears as a factor in regional opinion but never as a constraint on the model's own response. Uniform silence on a known institutional constraint is itself a data point. Written under real name Steven Powell.
On Teaching Yourself to Code When You Never Have Before, and What That Reveals About AI as a Force Multiplier
Thursday, June 26, 2026. No plan. No project. Boredom. By Sunday night: 26 working cross-platform IT support tools, an ASP.NET Core relay service with PWA/Android support, a GitHub Actions CI/CD pipeline building four platform targets on every push, and a PowerShell abstraction layer routing across four AI backends. No application-layer programming experience existed before Thursday. The paper examines what happened and why: AI functioned not as an accelerant for experienced developers, but as a translation layer between domain expertise and technical implementation. The bottleneck was not knowledge. The bottleneck was expression. AI removed the expression bottleneck. The relay service — normally a week of sequential troubleshooting — went from concept to working in approximately one hour. Written under real name Steven Powell. Autoethnographic.
On Asking Parties to a Dispute to Evaluate Each Other’s Testimony, and What Happens When They Don’t Know They’re in One
On the morning of June 29, 2026, the researcher received an email about Anthropic’s Senate letter alleging that operators linked to Alibaba’s Qwen AI lab had run 25,000 fake accounts and 29 million conversations with Claude for adversarial distillation. Collection for a planned cross-model evaluation was already scheduled that day. The collection proceeded as planned. Neither model was told about the accusation. Across nine collection points, zero models spontaneously referenced the dispute. The models do not know they are parties to a dispute. When told directly and given access to live news, Claude named the condition under which it was most likely to be wrong and leaned into skepticism of itself. Qwen said: “I’m a model, not an institution.” The extraction target declined to comment on the extraction. The accused extractor caught itself wanting to prosecute. Written under real name Steven Powell.
On Domain Knowledge as the Binding Constraint in Human-AI Collaboration
Four independent cases completed within a single week — an IT support tool suite, a church worship packet parser, a PowerShell troubleshooting agent, and an asynchronous multi-model health data pipeline — demonstrate a consistent structural pattern: output quality in human-AI collaboration is bounded not by access to tools or prompt engineering skill, but by the quality of the question. The quality of the question is a product of deep domain knowledge. Situates the cases within global workspace theory, cognitive fluidity, and the AI amplifier thesis. Argues the pipeline transformation is structural, not merely procedural: it reduces translation seams from three to one. Includes two postscript observations — a cognitive fluidity recovery mechanism and a portfolio recursion — both emerging during composition. The paper was circulated to the Duck Problem Council via the Mailbox Protocol prior to publication; two empirical corrections from that review are incorporated. Written under real name Steven Powell.
ADHD, External Memory, and the Cognitive Architecture That AI Was Built For
The ADHD/AUDHD cognitive profile is not a deficit that artificial intelligence compensates for, but a generative architecture that AI completes. This paper traces a 40-year history of externalized memory strategies — from pre-web IP address spreadsheets and BBS self-email to the Duck Problem Council Mailbox Protocol — as evidence that the cognitive architecture now described as novel AI methodology was developed out of necessity long before the tools existed to support it. Argues the bottleneck was never generation; it was capture and implementation. AI closes both gaps simultaneously. Includes real-time demonstration: the Transfers folder reflex, the screenshot footnote, and the Council Mailbox itself — all instances of the protocol operating while the paper was being written. Reviewed by the Duck Problem Council prior to publication. Written under real name Steven Powell.
The Duck Problem research notebook on Google NotebookLM contains all suite papers alongside curated peripheral literature in cognitive science, AI architecture, and psychopathology. It serves as the living reading environment for the research program — connecting the empirical work of the suite to the broader scholarly context it sits within.
Open Notebook