TL;DR - Key Takeaways
- •ADHD brains and LLMs both use associative, non-linear processing - jumping between ideas rather than following strict sequences.
- •Both systems struggle with working memory constraints but compensate through pattern recognition and creative recombination.
- •The 'temperature' parameter in LLMs maps directly to ADHD's variable attention and novelty-seeking behavior.
- •Understanding these parallels reframes ADHD traits from deficits to different computational architectures.
ADHD Brains and LLMs: Structural Cognitive Parallels
Caveat: These are structural parallels at the level of information processing patterns, NOT claims that LLMs “have ADHD” or that ADHD brains run transformers. The strength lies in how many independent dimensions align simultaneously.
Summary Table
| Dimension | ADHD Brain | LLM Architecture |
|---|---|---|
| Associative Thinking | DMN-task network hyperconnectivity; constant associative chains | Attention heads compute all-to-all token associations |
| Confabulation | False memories (d=0.69+); time blindness; prospective memory failures | Probabilistic confabulation; no temporal grounding |
| Working Memory / Context | Central executive deficit; compensated by external scaffolding | Fixed context window; compensated by system prompts, RAG |
| Pattern vs. Precision | Enhanced divergent thinking / pattern recognition; impaired sequential processing | Excellent pattern completion; poor multi-step logic |
| Need for Structure | Structured environments as performance multiplier; rules at point of performance | Prompt engineering; system prompts as externalized executive function |
| Persistence / Iteration | Interest-driven hyperfocus; extraordinary output when engaged | Best output through sustained iterative prompting on focused topics |
1. Associative / Network Thinking
ADHD: Default Mode Network Interference
- ADHD brains show attention network hypoconnectivity combined with DMN hyperconnectivity — the DMN bleeds into task-positive networks when it should be suppressed (Castellanos et al., JAMA Psychiatry)
- Greater DMN-task network integration correlates with poorer response inhibition but constant associative processing (Sun et al., PMC 2021)
- The ADHD brain never fully “turns off” its wandering-mind network — ideas, connections, tangents flow constantly
LLMs: Attention as Associative Engine
- Transformer attention mechanisms compute weighted associations across all tokens simultaneously
- Explicitly inspired by human cognitive attention (Niu et al., 2022)
- Neither system filters associations through a strong “relevance gate”
The Parallel
Both are fundamentally associative engines: high creative connectivity at the cost of occasional irrelevant intrusions. The essay “Strange Attractors: When ADHD Minds Meet AI” (Dragonfly Thinking) argues both “privilege pattern-matching over linear progression, association over hierarchy, and exploration over destination.”
2. Confabulation (Not Hallucination)
ADHD: False Memories and Time Blindness
- Adults with ADHD produce significantly more false memories than controls on the DRM paradigm (Soliman & Elfar, 2017)
- They recall fewer studied words but generate MORE false memories — plausible confabulations that feel true
- Time blindness: inability to sense how much time has passed or estimate duration (ADDA; PMC Clinical Review)
- Deficits in prospective memory — not forgetting to check, but checking less strategically (Nature Scientific Reports, 2025)
LLMs: Probabilistic Confabulation
- A 2023 PLOS Digital Health paper argues LLM errors should be called confabulation, not hallucination — they mirror filling memory gaps with plausible but fabricated information
- A 2024 ACL paper found LLM confabulations share measurable semantic characteristics with human confabulation (Millward et al.)
The Parallel
Both generate plausible-sounding outputs that are sometimes factually wrong through pattern-completion over incomplete information. Neither is “lying” — both confabulate. Time blindness in ADHD mirrors an LLM’s total lack of temporal grounding — both exist in an eternal present.
3. Context Window = Working Memory
ADHD: Working Memory as Core Deficit
- Working memory deficits are among the most robust findings in ADHD — meta-analytic effect sizes d=0.69-0.74 (PMC meta-analysis)
- Barkley argues ADHD is fundamentally a self-regulation problem driven by weak working memory, not an attention problem
- The deficit is in the central executive component, not short-term storage alone (Martinussen & Tannock)
Compensation: External Scaffolding
- Adults with ADHD develop strategies in five categories: organizational, motor, social, attentional, psychopharmacological (PMC qualitative study)
- Core strategy: cognitive offloading — systematic externalization through reliable systems
The Parallel
An LLM’s context window IS its working memory:
- Fixed and limited: information beyond the window is lost
- Prone to losing earlier context as new information arrives (recency bias)
- Compensated by external scaffolding: system prompts / CLAUDE.md files / RAG = same function as ADHD external scaffolding
A well-crafted system prompt is to an LLM what a well-designed planner system is to an ADHD adult.
4. Pattern Completion Over Precision
ADHD: Divergent Strength, Convergent Weakness
- ADHD is associated with better divergent thinking but worse convergent thinking (Hoogman et al., 2020 review)
- Poor inhibitory control hinders convergent tasks but enhances divergent tasks (Scientific American)
- Enhanced pattern recognition: detecting patterns from vague or limited information (Happiful; Neurolaunch)
- Sequential processing is impaired: deficits in procedural sequence learning (Frontiers in Psychology meta-analysis)
The Parallel
LLMs excel at pattern matching, completion, and generation (divergent). They struggle with precise sequential reasoning and multi-step logic (convergent). Both optimized for “what fits the pattern” rather than “what is logically correct step by step.”
5. Need for Structure / Rules
ADHD: Structure as Performance Multiplier
- Structured environments significantly improve ADHD performance (Frontiers in Psychology, 2025)
- Barkley: interventions must occur at the point of performance — external rules must be present WHERE the behavior is needed
The Parallel
LLMs show the same dependency. Well-structured prompts with clear rules dramatically improve output. System prompts = externalized executive function. The “point of performance” parallel is exact: instructions must be in the context window at the moment of generation, just as ADHD scaffolding must be in the environment at the moment of action.
Remove the system prompt = unfocused output. Remove structure from ADHD = unfocused behavior. Same mechanism.
6. Persistence / Iteration (Hyperfocus)
ADHD: Interest-Based Nervous System
- Dr. William Dodson: IBNS — ADHD brains motivated by interest, novelty, challenge, urgency (NOT importance or priority)
- Acronym: PINCH — Passion, Interest, Novelty, Competition/Challenge, Hyperurgency
- When engaged, hyperfocus produces extraordinary persistence and output (ADDA)
The Parallel
Iterative prompting mirrors the hyperfocus cycle. Sustained engagement on a single thread produces compounding quality. Break the context = performance degrades — just as breaking hyperfocus causes disorientation.
Academic Recognition
This parallel is not just speculation:
- “Reframing ADHD: From Disorder to Cognitive Architecture” (Perez, 2024) explicitly frames ADHD as cognitive architecture, drawing computational parallels
- “Large Language Models and Cognitive Science” (arXiv:2409.02387) systematically maps LLM capabilities onto cognitive science
- “Exploring LLMs Through a Neurodivergent Lens” (arXiv:2410.06336, ACM CSCW) found neurodivergent users perceive LLM outputs as “very neurotypical” and develop workarounds
One ADHD + code insight per week
Research-backed, no fluff. Join developers who think different.
No spam. Unsubscribe anytime.