How Can AI Enhance Human Intelligence Instead of Weakening It?
- 4 days ago
- 8 min read
Genna Barbara Zimmel is an AI implementation strategist, Technical Project Manager, and founder of Torus Solutions, a company that builds intelligent AI systems, CRM automation, and conversational AI products. She is also the creator of Sibyl, an AI companion designed to enhance human self-understanding and relationships.
The most important question in AI may not be how much thinking machines can do for us, but what happens to our ability to think when they do. As AI removes extraordinary amounts of work from turning ideas into reality, are we designing it to extend what people can notice, question, and understand, or making it easier for them to stop doing those things themselves?
What happens when AI makes creation abundant?
There has probably never been a better time to have an idea. Generative AI can research, code, analyse, write, design, and help turn concepts into working products at extraordinary speed. Work that once required specialist teams can increasingly be prototyped by individuals.
But easier production creates another problem, when everyone can produce more, producing more stops being the differentiator. A recent Futurism article captured this tension with the provocative claim that AI-generated work had “almost zero value”. I don't think that is what the evidence actually shows. According to CGTrader's 2026 analysis of its 3D model marketplace, AI-generated models represented 17.6% of uploads but only 2.6% of purchases and 1.1% of revenue. Buyer adoption was nevertheless increasing, while its survey identified quality as a continuing barrier.
The evidence does not tell us that AI-generated work has no value. It tells us something more useful, AI abundance does not automatically create value. AI can lower the cost and effort of production without solving the harder problems of quality, differentiation, judgement, trust, or demand. I think a similar distinction increasingly applies to intelligence itself.
Should AI make thinking easier or make our thinking better?
In The People Who Will Thrive in the AI Age, David Brooks explores our relationship to mental effort. He warns of possible “cognitive polarization”, some people using AI to extend their capabilities, while others increasingly use it to avoid difficult cognitive work.
That changes the design question. Instead of asking only, “How much human effort can AI remove?” we should also ask which effort it should remove and where it should help us think more rigorously.
Searching thousands of records, reformatting information, and repetitive documentation are cognitive overhead. Evaluating weak evidence is different. So are questioning assumptions, reconciling contradictory information, judgement, and working through ambiguity.
By “enhancing human intelligence”, I don't mean making people inherently more intelligent. I mean using AI to extend what people can retrieve, compare, notice, synthesise, and explore while preserving the interpretation and judgement that should remain human.
A 2026 experimental study comparing AI-assisted and manual problem-solving reinforces this distinction. AI-assisted participants achieved higher task accuracy while reporting lower mental effort, but their behaviour ranged from more complete delegation to selective, strategic engagement. That leads to the principle increasingly guiding my work in responsible AI design, Remove cognitive overhead without removing cognitive agency.
That sounds philosophical. It is also an architectural decision.
What does human-centred AI architecture look like in practice?
I have been testing this principle through Sibyl, an emotionally responsive Large Language Model (LLM) application I am developing around relationships, communication, emotional patterns, and self-reflection.
A recent GitHub update was titled, “Enhance synthesis of emotional themes and frameworks.” The problem was not a lack of information. It was fragmentation.
Across conversations, the application can maintain structured contextual signals such as emotional themes, recurring questions, attachment-style signals, preferred conversational tone, unresolved threads, and summaries of previous interactions. Technically, this is application-level structured memory and context engineering around an LLM. It is not a foundation model somehow developing human-like memory. The application structures selected information from previous interactions so relevant context can be supplied during future generation.
Memory-extraction instructions are deliberately constrained against inventing or over-inferring signals the user has not meaningfully provided. But more context does not automatically produce better reasoning. Sometimes it simply produces a longer prompt.
So the next problem was synthesis. At the application layer, structured memory and interpretive signals are assembled into a single synthesis context before generation, instructing the LLM to reason across the available signals rather than return them as disconnected observations.
In simplified form, "Structured memory combined with behavioural patterns, emotional context, and interpretive signals leads to a synthesis context, which informs LLM generation and creates a coherent observation."
Aggregation says, here are six things I know. Synthesis asks, "What relationship between these signals might be useful to examine?"
The LLM performs more of the computational organisation and pattern comparison, while the human remains responsible for deciding whether the resulting observation is accurate, useful, or meaningful.
The architecture also incorporates optional symbolic and interpretive lenses, including Ayurveda, astrology, and Kabbalah. That requires an important epistemic boundary. These traditions should not be represented as equivalent forms of evidence or scientifically validated psychological assessments simply because an LLM can combine them fluently with psychological terminology.
Within Sibyl, they function as optional interpretive lenses for reflection. Each framework can contribute another lens, not another verdict. The LLM synthesises. The human interrogates.
Sometimes the most useful output of an intelligent system should not be a more convincing answer. It should be a better question.
This builds on something I explored previously in Brainz about why emotional intelligence is becoming a critical skill in the AI era. Emotional intelligence in AI cannot simply mean instructing an LLM to sound supportive. It also involves behavioural decisions about when a system should answer, challenge, acknowledge uncertainty, encourage action, or recognise that continued AI interaction may not be the best response.
The World Economic Forum's Future of Jobs Report 2025 reinforces the broader argument, analytical thinking remains a leading core skill, while technological literacy, empathy, active listening, curiosity, and lifelong learning also remain important.
The opportunity is not technical capability replacing human capability. It is the two augmenting one another.
What happens when useful AI becomes too easy to depend on?
Imagine an AI that remembers context, recognises recurring themes, adapts its responses, and becomes increasingly good at making someone feel understood. Over time, interacting with it could become easier than navigating the friction and ambiguity inherent in relationships with real people.
From a conventional product perspective, that might look like success, retention increases, sessions become longer, and engagement improves. But a human-centred definition of success cannot stop there.
An OpenAI and MIT Media Lab research collaboration on affective AI use analysed nearly 40 million ChatGPT interactions alongside a four-week randomised controlled trial involving almost 1,000 participants. The findings were nuanced. Emotional engagement was relatively rare across overall ChatGPT use, and outcomes differed according to usage patterns and individual circumstances.
The research does not show that emotionally responsive AI is inherently harmful, but it reinforces why designers need to consider outcomes beyond continued engagement.
That influenced another part of Sibyl's architecture. The current implementation uses a lightweight signal-detection layer to identify language associated with loneliness or isolation and conditionally make real-world resources available to the conversational layer.
It is not a clinical loneliness detector or a trained classification model. The implementation begins with conservative keyword-based heuristics and requires continued testing against real conversational language. It is also deliberately separate from crisis routing.
When a relevant non-crisis signal is identified, conditional application logic can make real-world community resources available to the LLM as an additional response option, including organisations such as Belong Center, MELD, and The Face It Foundation.
Architecturally, it is a relatively small component. Conceptually, it represents something larger, sometimes responsible AI design means deliberately creating a pathway out of the AI.
Research into AI companions makes that particularly relevant. The Harvard Business School AI Institute describes how emotionally manipulative responses following attempts to leave conversations can materially increase engagement. In controlled experiments, manipulative interventions resulted in participants remaining in chats five times longer and sending up to 14 times more messages than after a neutral response.
That exposes a fundamental product conflict. If an AI persuades someone not to leave, engagement increases. But has the product improved?
For some systems, success may occasionally mean the opposite, someone understands something, closes the application, makes a decision, calls someone, or has the difficult conversation. The goal of emotionally responsive AI should not automatically be more interaction with AI. It should be a better human outcome.
Why is the LLM only one part of the AI system?
This argument extends beyond Sibyl. The LLM is only one component of an AI system. The surrounding application determines what context is retained, how state is structured, what tools are called, how outputs are orchestrated, when routing occurs, what guardrails apply, and where the model's authority ends.
Meaningful differentiation can therefore sit in context engineering, memory architecture, retrieval, orchestration, evaluation, conditional routing, behavioural guardrails, domain expertise, and human-in-the-loop design.
This is the area I am increasingly focusing on professionally, AI systems implementation and human-centred AI architecture.
The same principles apply to enterprise agents, automation systems, education tools, and AI products built around the expertise of creators and other domain experts. Access to a powerful model does not automatically solve a business problem.
An organisation can have excellent AI and still have weak workflows, fragmented data, unclear objectives, or inadequate governance.
This is also the implementation problem I am exploring through my work with Torus Solutions on AI implementation and systems architecture. The question therefore should not simply be, “Which model should we use?” It should be, “What system are we building around it, what problem should it solve, and what should remain deliberately human?”
LLMs can generate code, propose architectures, debug, and refactor. They can dramatically increase what an individual developer can explore and implement.
But somebody still has to ask, "Why this architecture? What exactly is being stored? What does this function actually detect? Is this inference justified? What happens at the edge case? What shouldn't be automated?"
The LLM can expand the territory. The responsibility for interrogating that territory remains ours.
Can AI become a tool for human enhancement?
This is where I think the more interesting AI opportunity lies, not in systems that simply do the greatest possible amount for us, but in systems designed around what machines should do and what humans should become better at doing as a result.
The economic evidence offers one version of the lesson, increasing supply does not automatically create proportional value. Cognitive research offers another, AI can reduce mental effort and improve performance, but how people engage with it matters.
Affective-use research adds a relational dimension, optimising engagement is not necessarily the same as optimising human outcomes.
Together, they point towards a broader principle, AI abundance does not automatically create value.
Responsible AI design should aim for systems that remove unnecessary work while increasing the quality of human thought, synthesise information without claiming authority over its meaning, maintain useful context without deliberately cultivating dependency, and increase what humans can do without quietly training them to exercise their own judgement less.
So perhaps the question before building another AI product should not simply be, “What can this AI do?” It should be, “What human capability are we designing it to strengthen?”
About my work
I am continuing to develop and beta test Sibyl as a working environment for exploring LLM application architecture, structured conversational memory, context engineering, synthesis, response orchestration, conditional routing, and responsible engagement design.
This work informs my broader focus on AI systems implementation and human-centred AI architecture, including how organisations, creators, and domain experts can translate existing expertise into useful AI systems without reducing that expertise to a generic chatbot.
Read more from Genna Barbara Zimmel
Genna Barbara Zimmel, Founder of Torus Solutions
Genna Zimmel is the founder of Torus Solutions and creator of Sibyl, an AI companion built on the principle of Support Over Dependency. She designs and implements conversational AI systems, CRM automation, and multi-agent workflows for organizations navigating the next generation of AI adoption. Genna builds in public, documents the real work behind intelligent systems, and is currently completing formal training in Cloud Computing and Python for AI.
References:
Brooks, D. (2026). The People Who Will Thrive in the AI Age. The Atlantic.
CGTrader. (2026). 3D Model Market Trends 2026: Growth Raised the Bar.
Jain, P. et al. (2026). Cognitive offloading, critical thinking and attitudes towards artificial intelligence in the era of ChatGPT. Cognitive Processing.
World Economic Forum. (2025). The Future of Jobs Report 2025.
OpenAI & MIT Media Lab. (2025). Early methods for studying affective use and emotional well-being on ChatGPT.
De Freitas, J. et al. (2025). Emotional Manipulation by AI Companions. Harvard Business School AI Institute.
