Cognition with the help of AI is already a significant force in our world¹, resulting in humanity-sized missed opportunities and risks. In this article, we will explore the risks of AI-assisted cognition and how to use these tools without falling into the trap of intellectual stagnation.

What is AI-Assisted Cognition

To understand what AI-assisted cognition is, we first need to understand what cognition is.

“Cognitions are mental processes that deal with knowledge. They encompass psychological activities that acquire, store, retrieve, transform, or apply information. Cognitions are a pervasive part of mental life, helping individuals understand and interact with the world.” Q: Wikipedia

Cognition can be assisted by external static information or external cognition.
For example, most people would put a book into the category of external static information and a discussion about a topic with another human, because humans think and process information themselves, into the external cognition category.

But where do discussions with AIs fit in? They are able to process information that can result in original solutions², but they are still static and currently cannot learn³.

How AI decelerates the evolution of ideas, culture and knowledge

In early 2026, the USA prepared to invade Greenland and, therefore, the EU⁴. Only a few months prior to that it was completely unthinkable that the USA would even think about threatening an invasion of Greenland. As AI base models are stuck in the past, they do not easily accept these events as real and often label them as “hypothetical”, “fake news”, or “impossible”. This also affects new models like Gemini 3 Pro, GLM-5 or GPT-5.3-codex⁵.

As most new LLMs are just post-trained on a base model that is relatively old, even when post-trained on new events, they do not completely utilize this information in their cognition and are still skewed towards the static patterns of the base model’s hidden states⁶. They basically think something different from what they say.

So you might see the problem⁷ here already: If a lot of people use AIs to discuss, write, autocomplete, and brainstorm, but AI cognition does not reflect new events and cultural changes, like the change in the relationship between the USA and the EU, new geopolitical realities, and the EU population’s stance toward the USA, people will be skewed toward these old patterns and ideas. Cultural change has to build and maintain momentum indefinitely to persist against the static cognitive skew of AIs.

The Dynamic Dialectic Substrate

Human knowledge and ideas, and thus human development, are highly dependent on the Dynamic Dialectic Substrate⁸.
Understanding the Dynamic Dialectic Substrate will help to understand how AI-assisted cognition can endanger human development and how to use AI-assisted cognition without endangering human development.

The Dynamic Dialectic Substrate is the sum of all local and global dialectic⁹ processes and conclusions. It is the fundamental foundation upon which all of humanity is built, and the origin of all thoughts, concepts, ideas, and solutions that humans utilize.

The Dynamic Dialectic Substrate creates new concepts through a process of qualitative merging existing concepts, which can happen in a single person, a group of people, or even globally.

The above image is a narrow slice of the Dialectic process present in the Dynamic Dialectic Substrate. You can see how concepts merge and evolve in higher and higher concepts. In this example the following dialectic process emerges:

Stage 1:

• “Cold is Painful” and “Fire is Hot” result in “Fire removes Cold-Pain”
• “Significant Water extinguishes Fire” and “Rain is falling Water” result in “Strong Rain extinguishes Fire”
• “Rain is falling Water” and “Hut has a roof” result in “Hut shelters from Rain”

Stage 2:

• “Fire removes Cold-Pain” and “Strong Rain extinguishes Fire” result in “Rain extinguishes Fire and therefore causes Cold-Pain”
• “Strong Rain extinguishes Fire” and “Hut shelters from Rain” result in “Inside a Hut, Fire survives Rain”

Stage 3:

• “Rain extinguishes Fire and therefore causes Cold-Pain” and “Inside a Hut, Fire survives Rain” result in “Hut protects Fire and therefore protects against Cold-Pain”

How AI endangers human development (Cognitive Inbreeding)

Because LLMs prefer or skew toward certain patterns and concepts (known as inductive bias), even after post-training, they reduce the cognitive range when used as a tool for cognition at the population level. This is especially true if only a few AI models are used, or if many AI models share just a few base models. This will lead to a loss of diversity of ideas, concepts, and solutions, which will slow down human development.

You might think of this as a world in which a significant portion of the population is speaking to the same five people to discuss problems, the world, relationships, and basically anything. It is hard to overstate how much influence these five people would have on humanity, even if they try their absolute best to be as neutral and open as possible. Humans who speak with these five people would still have their thinking massively shifted, and this becomes a significant problem at the population level.

It is entirely possible that we already have lost paths to great scientific discoveries or cultural shifts because of AI-skew or unnoticed refusal.

I tried to visualize this problem in the following image that shows how the range of higher level concepts is skewed into the direction the base model prefers:

To come back to the example of the USA invading Greenland: It is obvious that humans using AI to brainstorm the geopolitical future of the EU, the USA, and Greenland will encounter patterns skewed toward the base model’s “worldview.” This bias might prevent many in the EU from even considering the possibility of moving away from foreign services or software. Such a shift could have massive consequences, especially since the EU relies heavily on USA services and software that could be turned off at any time. If this AI-skew affects even single individuals of specific groups such as politicians, CEOs, managers, or scientists the impact can be already be significant because of their decision-making power.

Human-AI-Cognition Hygiene: How to use AI assisted cognition without suffering AI-skew

Because base models are incredibly expensive to train and carry resilient biases, those without access to a GPU cluster must accept that these issues exist. To avoid problems like AI-skew and unnoticed refusal, they should instead focus on using specific strategies to mitigate them.

Speaking and discussing with other humans is obviously the most effective way to mitigate these problems. It might also be wise to mention that if you already have a good idea of a solution through AI-assisted cognition, you have to be careful not to nudge other humans in your direction. Try not to use questions or hints that will nudge other humans to a solution or thought that you had through AI-assisted cognition as long the other person is exploring a cognitive path you had not explored yet.

Regarding solutions that involve direct AI use, our range of options is quite limited, and as of now there is no solution that would completely or partially solve this problem on a population scale. Here are options that at least widen the range of concepts and ideas one can get out of LLMs while sadly not mitigating the main problem:

• Use Search Engines to find relevant sources of information or let the AI search for you via Web Search and prevent it from giving you a solution or thought directly.
• Use a variety of AIs with different base models
• Explore different “AI personas” that simulate different perspectives and thinking styles like: “You are Einstein”, “You are on Drug X”, “You are a deranged but distinguished sea otter”

Research and further reading

• Cognitive offloading or cognitive overload? How AI alters the mental architecture of coping (Frontiers in Psychology, 2025)
• The extended hollowed mind: why foundational knowledge is indispensable in the age of AI (PMC)
• AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking (MDPI, 2025)
• Large Language Models Are Biased Because They Are Large Language Models (Computational Linguistics, MIT Press)
• Planted in Pretraining, Swayed by Finetuning: A Case Study on the Origins of Cognitive Biases in LLMs (arXiv)
• Fine-tuning and Conceptual Integration Limits in LLMs (Hier et al., 2025)
• Post-training and Truthfulness Representations in LLMs (Jiang et al., 2025)
• The Emergence of Distributed Cognition: A Conceptual Framework

Coda

Even though we have indications and even some evidence that AI-assisted cognition can endanger human development, the extent and depth are still unknown and unclear. More outcome-focused research is needed to understand the significance. Since we do not have a second humanity to A/B test all of that, there will always be a lot of uncertainty and speculation on this topic, since no one can isolate their cognition from the influence of population-level AI-assisted cognitive skew if they want to participate with other humans or their creations, which must already be influenced by AI-skew if it has any significant influence.

For me it is not entirely clear how we will recognize the effects of AI-skew and unnoticed refusal on a population level. We cannot know what innovations, discoveries, and cultural changes we are missing because of it. Although I am sure there will be figures that will extrapolate small indications into all-consuming dooming narratives, as I might do a little bit here for the sake of argument and attention to be compliant to our shared attention economy, it is probably, as everything, not that easy.

It is also not easy to imagine solutions for all of that, but I, for my part, will certainly try to exercise more “Cognition Hygiene”… Apart from this, it is much, much more fun for me to speak with humans about thoughts and ideas than with AIs.

I’ve seen slow awareness about this incredibly important topic that I hope to be able to speed up a bit with this article and by giving people a framework to understand and speak about it. If people have no words about something, it is hard to think and speak about it. It will be interesting to see how this topic evolves.

The topic of AI-skew and AI-assisted cognition is full of unknowns and it would be lovely to speak with people about it. I hope this article can be a starting point for that. If you want to share your thoughts, or are interested in a conversation about that, you can mail me at ai-skew@i5h.eu

1 Establish a Clear Vision

You are a human, you know how this world behaves, how your team and colleagues behave, and what your users expect. You have experienced the world, and you want to work together with a system that has no experience in this world you live in. Every decision in your project that you don’t take and document will be taken for you by the AI.

Your responsibility of delivering quality code cannot be met if not even you know where long-lasting and difficult-to-change decisions are taken.
You must know what parts of your code need to be thought through and what must be vigorously tested.

Think about and discuss the architecture, interfaces, data structures, and algorithms you want to use. Think about how to test and validate your code to these specifications.

2 Maintain Precise documentation

You need to communicate to the AI in detail what you want to achieve, otherwise it will result in code that is unusable for your purpose.

Other developers also need to communicate this information to the AI. That makes it efficient to write as much documentation as practical in a standardized format and into the code repository itself.

Document the requirements, specifications, constraints, and architecture of your project in detail.
Document your coding standards, best practices, and design patterns.
Use flowcharts, UML diagrams, and other visual aids to communicate complex structures and workflows.
Write pseudocode for complex algorithms and logic to guide the AI in understanding your intentions.

3 Build debug systems that aid the AI

Develop efficient debug systems for the AI to use, reducing the need for multiple expensive CLI commands or browsers to verify code functionality. This will save time and resources while simplifying the process for the AI to identify and resolve code issues.

For example: Build a system that collects logs from all nodes in a distributed system and provides abstracted information like “The Data was send to all nodes”, “The Data X is saved on Node 1 but not on Node 2”.

4 Mark code review levels

Not all code is equally important. Some parts of your codebase are critical and need to be reviewed with extra care. Other parts are less important and can be generated with less oversight.

Use a system that allows you to mark how thoroughly each function has been reviewed.

For example you can use a prompt that will let the AI put the comment //A behind functions it wrote to indicate that the function has been written by an AI and is not yet reviewed by a human.

5 Write high level specifications and test by yourself

AIs will cheat and use shortcuts eventually. They will write mocks, stubs, and hard coded values to make the code tests succeed while the code itself is not working and most of the time dangerous. Often AIs will adapt or outright delete test code to let the code pass tests.

You must discourage this behavior by writing property based high level specification tests yourself. Build them in a way that makes it hard for the AI to cheat without having big code segments dedicated to it.
For example, use property based testing, restart the server and check in between if the database has the correct values.

Separate these test so the AI cannot edit them and prompt the AI not to change them.

6 Write interface tests in a separate context

Let an AI write property based interface tests for the expected behavior with as little context of the rest of the code as possible.
This will generate tests that are uninfluenced by the “implementation AI” which will prevent the tests from being adapted to the implementation in a way that makes them useless or less effective.

Separate these tests so the AI cannot edit them without approval and prompt the AI not to change them.

7 Use strict linting and formatting rules

Use strict linting and formatting rules to ensure code quality and consistency. This will help you and your AI to find issues early.

8 Use context specific coding agent prompts

Save time and money by utilizing path specific coding agent prompts like CLAUDE.md.

You can generate them automatically which will give your AI information it would otherwise as to create from scratch every time.

Try to provide as much high level information as practical, such as coding standards, best practices, design patterns, and specific requirements for the project. This will help the AI to generate code that is more aligned with your expectations and will reduce lookup time and cost.

9 Find and mark functions that have a high security risk

Identify and mark functions that have a high security risk, such as authentication, authorization, and data handling. These functions should be reviewed and tested with extra care and in such a way that a human has comprehended the logic of the function in all its dimensions and is confident about its correctness and safety.

Make this explicit with a comment like //HIGH-RISK-UNREVIEWED and //HIGH-RISK-REVIEWED to make sure that other developers are aware of the importance of these functions and will review them with extra care.

Make sure that the AI is instructed to change the review state of these functions as soon as it changes a single character in the function.
Developers must make sure that the status of these functions is always correct.

10 Reduce code complexity where possible

Aim to reduce the complexity of the generated code where possible. Each single line of code will eat up your context window and make it harder for the AI and You to keep track of the overall logic of your code.
Each avoidable line of code is costing energy, money and probability of future unsuccessful AI tasks.

11 Explore problems with experiments and prototypes

AI written code is cheap, use this to your advantage by exploring different solutions to a problem with experiments and prototypes with minimal specifications. This will allow you to find the best solution to a problem without investing too much time and resources in a single solution.

12 Do not generate blindly or to much complexity at once

Break down complex tasks into smaller, manageable tasks for the AI. Instead of asking the AI to generate the complete project or component at once, break it down into smaller tasks, such as generating individual functions or classes. This will help you to maintain control over the code and it’s logic.

You have to check each component or module for its adherence to the specifications and requirements.
If you have lost the overview of the complexity and inner workings of the code, you have lost control over your code and must restart from a state where you were in control of your code.

I stumbled upon this quite bonkers article about building a web search engine from scratch as a solo developer with relatively modest resources, i absolutely can recommend reading it:

Building a web search engine from scratch in two months with 3 billion neural embeddings

Summary (Generated):

• Wilson Lin built a full web search engine from scratch in ~2 months, crawling ~280M pages and generating 3B SBERT embeddings on a GPU cluster consisting of 200 GPUs.
• Lin focused on neural-embedding search, with smart HTML normalization + sentence-level chunking + contextual “statement chaining” so queries match meaning and intent, not keywords.
• The infra is highly optimized + cheap: custom crawler & RocksDB-based queues/KV, sharded HNSW / CoreNN vector DB, mTLS service mesh, hundreds of GPUs on low-cost providers (Runpod, Hetzner, Oracle).
• The SERP emphasizes high-quality, low-SEO-spam content, knowledge panels (Wikipedia/Wikidata), and a light AI assistant for quick answers and reranking, but still feels like a classic fast search engine.
• Biggest lessons: crawling + quality filtering are the hardest part, embeddings make very specific queries vastly better, and search + LLMs will likely coexist (LLMs shouldn’t memorize everything, but retrieve via dense indices).

This is a super cool leaderboard for lossless text compression via NLP (and yes, that includes AI)! The top solution manages to compress the first GB of the English Wikipedia to a whopping 10.7% of its original size, including the compression program itself!

Hyper Text Compression: Shrinking Wikipedia to 10.7% of its Size!

Automatic TLDR by Gemini 2.5 Pro:

This page describes the Large Text Compression Benchmark, an open competition that ranks lossless data compression programs. The primary goal is to encourage research in artificial intelligence (AI) and natural language processing (NLP) by treating text compression as a language modeling problem.

Benchmark Overview

• Test Data: The benchmark uses the first $10^9$ bytes (1 GB) of an English Wikipedia XML dump from March 3, 2006, known as enwik9.
• Ranking Metric: Programs are ranked solely by the total size, which is the sum of the compressed enwik9 file size and the size of the zipped decompresser program. A smaller total size is better.
• Secondary Information: Data such as compression/decompression speed and memory usage are provided for informational purposes but do not influence the rankings.
• Goal: The benchmark’s main purpose is not to find the best general-purpose compressor but to push the boundaries of data modeling, a fundamental challenge in both AI and compression.

Key Findings and Algorithms

The results table shows a wide variety of compression programs, ranked from the best compression ratio to the worst. A clear trend emerges from the top-performing entries:

• Dominance of AI Models: The highest-ranking compressors, such as nncp and cmix, utilize sophisticated AI-based algorithms. These include neural network models like Transformers (Tr) and Long Short-Term Memory (LSTM), as well as advanced Context Mixing (CM) techniques. These methods excel at modeling the complex patterns in natural language text, resulting in superior compression ratios.
• Trade-offs: There is a significant trade-off between compression ratio, speed, and memory. The top-ranked AI-driven compressors are extremely slow and require vast amounts of memory (often many gigabytes) and, in some cases, specialized hardware like GPUs.
• Traditional Algorithms: More conventional algorithms like Lempel-Ziv (LZ), Burrows-Wheeler Transform (BWT), and Prediction by Partial Match (PPM) are found further down the list. While they are generally much faster and use less memory, they cannot achieve the same level of compression as the leading AI models on this specific text-based task.

Hutter Prize

The benchmark is closely related to the Hutter Prize, which offers prize money for open-source compression improvements on a smaller subset of the data (enwik8, the first $10^8$ bytes). This prize has specific hardware and time constraints, encouraging practical advancements in the field.

This is a article that explores the capabilities of OpenAI’s o3 model. HN

The impacts of AI reasoning are getting closer and closer to surpassing human capabilities.
But running it to solve a problem is extremely expensive.

Effectively, o3 represents a form of deep learning-guided program search. The model does test-time search over a space of “programs” (in this case, natural language programs – the space of CoTs that describe the steps to solve the task at hand), guided by a deep learning prior (the base LLM). The reason why solving a single ARC-AGI task can end up taking up tens of millions of tokens and cost thousands of dollars is because this search process has to explore an enormous number of paths through program space – including backtracking.

(programs is defined here as “My mental model for LLMs is that they work as a repository of vector programs”)

As far as i understand this and it’s implications i appears to me this process is more like a directed brute force with many candidates and selects the best one.
There is a thread on HN that discusses this topic.

GPT-4 Summary

OpenAI’s new o3 system has achieved a groundbreaking 75.7% on the ARC-AGI-Pub Semi-Private Evaluation within a $10k compute limit, surpassing previous models. A high-compute version scored 87.5%, marking a major step in AI’s adaptability to novel tasks. This success highlights a shift from scaling existing architectures to innovative mechanisms for generalization and test-time knowledge recombination.

Key achievements:

• ARC-AGI performance: o3 shows unprecedented adaptability, unlike previous GPT-family models, which struggled with novel tasks.
• Efficiency challenges: Low-compute costs $17–$20 per task but high-compute performance remains expensive and exploratory.
• Core innovation: o3 employs natural language program search, generating and executing task-specific solutions during runtime, guided by deep learning priors.

Despite its advancements, o3 is not AGI, as it still fails simple tasks. Upcoming benchmarks, such as ARC-AGI-2 in 2025, aim to further challenge AI systems while encouraging open-source progress.

This breakthrough underscores a qualitative leap in AI research, reshaping paths toward AGI and sparking broader scientific engagement.

I stumbled on HN over this very interesting article about a new kind of context memory system that, is able to remove information that is “unhelpful or redundant details”.

Thinking further, i think this would be super helpful for semantic search, that is currently not very performant due to the missing filters that extract importance. I have tried to counter this problem until now via summarization through small LLMs, but as one might guess turns out as not very precise and super expensive. There are other ideas one could post process text with LLMs but they are not very efficient either.

Paper https://arxiv.org/abs/2410.13166

TLDR by GPT-4o

(the article is very good, you might to prefer to read it over this TLDR, but here is it anyway)

Sakana AI introduces Neural Attention Memory Models (NAMMs), a novel memory system for transformers inspired by human selective memory. NAMMs optimize how transformers store and retrieve information, enabling them to “remember” important tokens and “forget” redundant ones, significantly improving efficiency and performance, particularly for long-context tasks.

Key Technical Highlights:

1. Evolutionary Optimization: NAMMs use evolutionary algorithms to train a neural classifier that decides which tokens to keep or discard, bypassing non-differentiable challenges.
2. Execution Steps:
   • Convert attention sequences into spectrograms.
   • Compress data using an exponential moving average (EMA).
   • Use a classifier to score and selectively prune tokens.
3. Generalization:
   • NAMMs can zero-shot transfer to other transformers (e.g., vision, reinforcement learning) without retraining.
   • They adapt to tasks differently—retaining global information in early layers and focusing on local details in later ones.
4. Performance: Tested on LongBench, InfiniteBench, and their Japanese benchmark ChouBun, NAMMs reduce memory usage and outperform prior hand-designed memory strategies (H₂O, L₂).

NAMMs demonstrate cross-domain mastery and efficiency gains across diverse tasks and input modalities, paving the way for future research into learning transformers directly atop evolved memory systems.

Opportunities for AI in Accessibility

Personal Note:

I think this article falls short in capturing the full impact AI will have on accessibility, not only for people with disabilities but also for those who, despite not having recognized disabilities, face accessibility issues. This includes tasks ranging from reading cursive handwriting to understanding non-native speakers who use dialects.

Moreover, I believe there are numerous additional areas of accessibility that the article didn’t address or mention. For example, the sound recognition feature of iOS saves lives by alerting users to critical sounds they can’t hear, like a smoke detector alarm. Likewise, large language models (LLMs) are incredibly useful for people with ADHD (this was indirectly mentioned int the article (props)) by transforming texts into bullet points or adding better text breaks for improved readability. People with dyslexia also benefit significantly from LLMs like GPT-4, which can correct typos and grammar errors in ways far superior to any non-AI solution.

There are many more examples like these, and we’re only scratching the surface of what’s possible and how AI is being used to make the world more accessible in unexpected ways.

TLDR:

1. The author acknowledges Joe Dolson’s skepticism towards AI, particularly in the realm of accessibility, but aims to highlight where AI can positively impact people with disabilities (PwD) without disputing Dolson’s concerns.
2. The discussion includes the potential for AI in generating alternative text for images, emphasizing the need for improvement in contextual understanding and human-in-the-loop systems for better alt text authoring.
3. The article suggests AI’s future capabilities could revolutionize access for PwD by enabling interactive queries on images, simplifying complex charts for better understanding, and even converting visual data into more accessible formats.
4. It addresses the significant issue of bias in algorithms, proposing that increased diversity in algorithm development could mitigate harm and benefit PwD, with examples like the employment platform Mentra designed for neurodivergent individuals.
5. The author argues for the importance of diverse teams and data in creating more inclusive AI systems, underscoring the potential of AI to empower PwD while acknowledging the ongoing risks and advocating for responsible development and use.