The legal profession is betting big on AI technologies. The global legal tech market was valued at $31.6 billion in 2024 and is projected to nearly double by 2032 (Fortune Business Insights), while in the U.S., AI adoption among large firms jumped from 16% to 46% in just one year (Lawnext). But a groundbreaking MIT study on AI and memory suggests a counterintuitive truth: lawyers may achieve stronger long-term results by engaging their own brains first, and only layering in AI afterwards. As we rush to integrate AI into daily practice, this raises a pressing question: are we gaining efficiency at the cost of our cognitive abilities, and should we be adopting AI more strategically?
The MIT Study: AI's Impact on the Brain
Researchers at MIT's Media Lab recently published a comprehensive study examining how AI tools affect human cognition over time. The 200+ page study by Kosmnya et al, titled "Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task," followed 54 participants over four months as they completed writing tasks under three different conditions: using AI tools like ChatGPT (“LLM group”), using traditional search engines (“Search Engine group”), or relying solely on their own cognitive abilities (“Brain-only group”). Conditions were maintained across three sessions, with tool-switching in session four (18 participants).
The findings were striking. Researchers used EEG to record brain activity across 32 regions and found that the LLM group (using ChatGPT) had the lowest brain engagement. By contrast, participants of the Brain-only group writing without tools engaged the widest neural networks and carried the heaviest cognitive load, yet achieved the strongest learning outcomes.
Key Research Findings
The study examined the neural and behavioral consequences of using LLMs in essay writing.
- Brain-only group: Though working under greater cognitive load, participants achieved the strongest outcomes: better memory, higher semantic accuracy, and stronger ownership of their work. Their reliance on internal content generation fostered deeper learning and a stronger sense of authorship.
- Search Engine group: Participants balanced internal effort with use of a search engine for research. They showed moderate memory retention and a partial sense of ownership over their work.
- LLM group: Though participants demonstrated efficiency, they exhibited poor memory retention, reduced self-monitoring, and difficulty recognising their own authored work.
Reduced Brain Activity
EEG analysis showed that the Brain-only group had the highest neural connectivity, especially in alpha, theta, and delta bands linked to creativity, semantic processing, and memory load. In contrast, AI users showed significantly less engagement.
Memory Impairment and Ownership Loss
When participants who had used LLMs were later asked to write unaided, 83% failed to recall their own essays in Session 1 (Kosmnya et al p.137), and none could provide accurate quotes. By Session 4, 78% of LLM-to-Brain participants still failed to quote anything, compared with just 11% in the Brain-only group (p < 0.01)(Kosmnya et al, p.140). This impairment reflected weaker alpha and theta activity, suggesting bypassed memory processes.
Persistent Effects
Perhaps most concerning, even after stopping AI use, is that the cognitive effects lingered. LLM-to-Brain participants never reached the same neural pathways developed by the Brain-only group. In contrast, Brain-to-LLM participants, those who first wrote without tools in Sessions 1-3, displayed enhanced connectivity and memory reactivation when later using LLMs. This suggests that unaided effort in early stages strengthens later tool use, while early AI reliance may hinder long-term memory formation and ownership.
Cognitive Debt
Another key finding is the accumulation of cognitive debt among participants who relied heavily on LLMs (Kosmnya et al, p.141). By outsourcing the most effortful parts of essay writing (idea generation, semantic integration, and critical evaluation) these participants deferred cognitive effort in the short term but incurred long-term costs. This debt manifested in weaker memory traces, reduced semantic recall, and a narrower engagement with topics. Though based on a relatively small sample and flagged as preliminary, these specific cognitive debt findings are nonetheless significant, pointing to the risk that heavy reliance on LLMs may create lasting cognitive costs that outlive the short-term gains in efficiency.
In contrast, Brain-only participants built stronger foundational memory networks and displayed greater metacognitive control, allowing them to later benefit from AI assistance without the same deficits. These results suggest that cognitive debt may not simply reflect momentary under-engagement but a cumulative effect, whereby repeated reliance on external systems reduces the capacity for independent synthesis, critical inquiry, and creative ideation. Managing this debt, by sequencing tool use after self-driven effort, may be key to balancing efficiency with long-term cognitive autonomy.
Impact on Legal Professionals
These findings have profound implications for the legal profession. Legal practice fundamentally relies on cognitive skills that the MIT study suggests may be at risk: analytical thinking, memory retention, pattern recognition, and the ability to synthesise complex information.
But here's the crucial question: in delegating tasks to AI, are we delegating our memory?
In many ways, lawyers have allongfunctioned like generative systems, we take masses of information and synthesise it, pulling out key details and formulating arguments from complex data. This process of synthesis, described by educational theorist Howard Gardner as taking disparate pieces of information and weaving them into coherent understanding, is fundamental to how human memory works and what makes lawyers valuable.
When we actively synthesise information ourselves, we strengthen neural pathways and build lasting knowledge. However, if we delegate this synthesis work to AI, it's the AI that does the cognitive heavy lifting, not us. Synthesis is not just important for memory, it is essential for human intelligence itself. Without the capacity for synthesis, lawyers cannot function effectively.
Consider the parallels between the study's essay-writing tasks and core legal work:
- Legal research and analysis require the same deep cognitive engagement that was diminished in AI users
- Brief writing and argumentation demand the creative and analytical thinking that showed reduced neural activity
- Case preparation and client counselling rely on memory retention and the ability to recall and connect relevant information
- Professional competence depends on developing and maintaining expertise through active cognitive engagement
If legal professionals increasingly offload cognitive work to AI systems, we risk creating the type of cognitive debt kosmyna et al observed, i.e. a gradual erosion of the very mental faculties that define excellent lawyering!
Using AI with Intention: Two Practical Approaches
The research does not suggest abandoning AI entirely, but rather using it more thoughtfully. Here are two strategies legal professionals can adopt:
1. Practice the "Think First" approach: Before turning to AI tools, spend time thinking through the problem yourself. Draft initial thoughts, outline key issues, or brainstorm potential arguments before seeking AI assistance. This could be as simple as jotting down a rough draft before asking for a rewrite, or outlining ideas manually before letting an AI polish them up. This ensures your brain remains actively engaged in the cognitive work that builds expertise.
2. Use AI as a collaborator, not a replacement: Rather than asking AI to complete entire tasks, use it to enhance your own thinking. Ask it to challenge your arguments, suggest alternative perspectives, or help you refine your analysis—but ensure the foundational cognitive work remains your own. This approach maintains the active neural engagement that the study shows is crucial for memory formation and skill development.
The Path Forward
The MIT study serves as an important reminder that technological advancement isn't automatically beneficial. As legal professionals, we must balance the undeniable efficiency gains AI provides with the need to preserve and develop our cognitive capabilities.
The goal shouldn't be maximum AI usage, but rather optimal integration, leveraging AI's strengths whilst maintaining the deep analytical thinking, memory formation, and cognitive flexibility that define excellent legal practice.
Your thoughts?
How has AI integration affected your own cognitive processes and professional practice? Have you noticed changes in your memory, analytical thinking, or problem-solving approaches since adopting AI tools?
We'd love to hear from legal professionals about their experiences with AI adoption and any strategies they've developed for maintaining cognitive engagement. Please share your thoughts and experiences: info@deep-lex.com
Disclaimer: The above article is intended for information purposes only and does not constitute legal advice. Please refer to the terms and conditions page for more information.
Sources
1. “Legal Technology Market Size, Share & Growth Report,” Fortune Business Insights, 2024–2032 estimates, citing a valuation of $31.59 billion in 2024 and a projected rise to $63.59 billion by 2032, https://www.fortunebusinessinsights.com/legal-technology-market-109527?utm_source=chatgpt.com
2.Bob Ambrogi, “ABA Tech Survey Finds Growing Adoption of AI in Legal Practice,” LawNext, 7 March 2025: AI use rose from 16% in 2023 to 46% in 2024 among law firms with 100+ attorneys, https://www.lawnext.com/2025/03/aba-tech-survey-finds-growing-adoption-of-ai-in-legal-practice-with-efficiency-gains-as-primary-driver.html?utm_source=chatgpt.com
3.Nataliya Kosmyna, Eugene Hauptmann, Ye Tong Yuan, Jessica Situ, Xian-Hao Liao, Ashly Vivian Beresnitzky, Iris Braunstein, and Pattie Maes. "Your brain on chatgpt: Accumulation of cognitive debt when using an ai assistant for essay writing task." arXiv preprint arXiv:2506.08872 (2025).Published online on 10 June 2025., https://arxiv.org/abs/2506.08872