LLM failure modes map surprisingly well onto ADHD cognitive science. Six parallels from independent research.
By Holidays in Europe / March 22, 2026 / No Comments / Uncategorized
Exploring the Surprising Parallels Between Large Language Model Failures and ADHD Cognitive Patterns
In recent years, large language models (LLMs) such as GPT-4 and others have revolutionized the way we approach automation, AI-driven content creation, and interactive applications. However, these models are not infallible; they exhibit specific failure modes that can sometimes seem familiar, especially to those with certain cognitive profiles like ADHD. As someone with ADHD who has been experimenting with pair programming with LLMs, I noticed that the ways they falter closely mirror common patterns observed in ADHD-related cognition. Intrigued by this, I delved into the cognitive science literature and uncovered six fascinating parallels, supported by independent research, that shed light on both the AI failures and human cognition.
1. Associative Processing and the Brain’s Default Mode Network
ADHD is characterized by a tendency for the Default Mode Network (DMN)—the brain’s resting state—to intrude into task-oriented networks, leading to a more associative, less focused thought process (Castellanos et al., JAMA Psychiatry). Similarly, transformer attention mechanisms in LLMs operate by computing weighted associations across tokens without a strict relevance filter. This results in highly creative connections and random, irrelevant intrusions—much like the associative thinking seen in ADHD brains.
2. Confabulation: Filling Gaps with Plausible Fabrications
Adults with ADHD often produce false memories that feel authentic to them—a phenomenon known as confabulation (Soliman & Elfar, 2017). A recent study in PLOS Digital Health argued that LLM errors should be classified as confabulation rather than hallucination, because both systems fill informational gaps with plausible, pattern-completed content. Research (Millward et al., 2024) further found that LLM confabulations share measurable traits with human confabulation, emphasizing that neither is intentionally lying—both are completing uncertain data with what seems probable.
3. Working Memory as Context Window: The Memory Bottleneck
Working memory deficits are a well-documented aspect of ADHD, with meta-analyses indicating moderate to large effect sizes (d=0.69–0.74). Likewise, an LLM’s context window functions as its ‘working memory’—a fixed-size buffer where information can fall off or become fuzzy over long conversations. Both humans and models optimize their external systems to compensate: people use notes, organizers, and external tools; LLMs utilize system prompts, retrieval-augmented generation (RAG), and document embeddings.
4. Preference for Pattern Completion Over Precise Reasoning
ADHD tends to enhance divergent thinking—generating creative ideas—while impairing convergent thinking, which involves precise, step-by-step reasoning (Hoogman et al., 2020). The same is true for LLMs, which excel at pattern matching and creative completion but struggle with multi-step logical reasoning. Both are tuned to predict “what fits the pattern” rather than produce strictly logical outputs.
5. Structure as a Tool to Enhance Performance
Structured environments dramatically improve both ADHD functioning and LLM outputs. In ADHD, clear routines and constraints serve as external scaffolding to boost focus (Frontiers in Psychology, 2025). Similarly, a well-crafted prompt with explicit instructions enables LLMs to generate coherent, relevant responses. Remove structure, and outputs tend to become unfocused or incoherent, highlighting how both systems rely on external frameworks to succeed.
6. Sustained Engagement and Thread Continuity
Maintaining focus on a single thread of thought leads to compounded, high-quality outputs in both the ADHD brain and LLMs. Interrupting this focus causes a sharp decline in coherence—whether it’s losing track of a complex conversation or a person forgetting where they left off after an interruption. This underscores the importance of continuity for productivity and effective collaboration.
Practical Insights and Reflection
The key takeaway is that individuals with ADHD have, through years of self-management, developed strategies—external scaffolding, pattern-first thinking, iterative approaches—that align remarkably well with effective AI collaboration techniques. Recognizing these parallels not only provides insight into AI behavior but also highlights the value of cognitive strategies for managing both human and machine cognition.
For a deeper dive into the research behind these parallels, I’ve compiled a detailed article complete with citations at thecreativeprogrammer.dev.
Your Thoughts?
Have you observed similar patterns? Do you see parallels between how large language models fail and how your own cognitive processes work? I’d love to hear your experiences and insights.
Note: This article synthesizes research from multiple fields to explore the intriguing intersection of AI failure modes and ADHD cognition. Understanding these parallels can inform better AI design, as well as strategies for managing cognitive diversity in human-AI collaboration.