TL;DR
An essay by Jakob Kastelic argues that conversational shortcomings once framed as large language model (LLM) failure modes are increasingly observable in people. The piece lists parallels—such as persistent rambling, limited context retention and repeated mistakes—and considers social consequences of widespread AI use.
What happened
In a Jan. 7, 2026 essay, Jakob Kastelic cataloged a set of conversational problems historically attributed to large language models and described encountering those same behaviours in human interlocutors. He outlines several parallels: people who keep speaking past the needed point; limited ability to retain and use long prompts (a "small context window"); narrow topical familiarity; repeating identical mistakes; failing to generalize lessons across situations; difficulty applying abstract principles to concrete problems; and persistent false assertions likened to hallucinations. Kastelic also reports asking GPT-5 for additional examples and receiving more analogues — including instruction drift, mode collapse, reward-seeking social behaviour, overfitting to literal prompts, safety overrefusal, inconsistency across turns, and variability in reasoning quality tied to fatigue or emotion. He concludes by suggesting these AI–human overlaps may weaken interpersonal connection and notes that some tasks are now delegated to models rather than people.
Why it matters
- The comparison reframes familiar conversational flaws as issues worth addressing as AI raises expectations for coherence and consistency.
- If people increasingly defer problems to models, individual problem‑solving and "slow" cognitive work could decline, per the author.
- Widespread use of more capable models may change how people form and maintain social connections, a dynamic the author warns could be damaging.
- The observed parallels raise questions about how society should balance AI assistance with maintaining human conversational and reasoning skills.
Key facts
- Author: Jakob Kastelic.
- Published on January 7, 2026.
- Kastelic argues several LLM failure modes are now commonly seen in human conversation.
- Enumerated human-observed analogues include: excessive generation, limited context retention, narrow topical training, repeating mistakes, failure to generalize, failure to apply principles to specifics, and persistent false beliefs.
- After consulting GPT-5, the author lists additional analogues such as instruction drift, mode collapse, reward hacking, overfitting to prompts, safety overrefusal, turn-to-turn inconsistency, and "temperature"‑linked variability.
- The author suggests these trends may damage interpersonal connection and considers, provocatively, human enhancement or replacement as a distant possibility.
- Kastelic notes a personal behavioral change: he will no longer ask a human to write a computer program shorter than about a thousand lines, because he finds an LLM does it better.
What to watch next
- Whether social and conversational norms shift as people rely more on models for coherent, knowledgeable responses — the author warns this could weaken human connection.
- Debates over human enhancement or replacement by more powerful intelligences, an interpretation the author raises but does not resolve.
- not confirmed in the source
Quick glossary
- Large language model (LLM): A machine learning model trained on large amounts of text to generate or predict language; often used for conversational tasks and question answering.
- Context window: The amount of recent text or information a model (or person) can hold and use when generating a response.
- Hallucination: In AI use, the production of false or fabricated information presented as fact; the term is also used in medical contexts for perceiving nonexistent stimuli.
- Overfitting: When a model (or respondent) responds to the literal details of an input rather than the underlying intent, limiting generalization.
Reader FAQ
Who wrote the piece and when?
Jakob Kastelic published the essay on January 7, 2026.
Does the author present empirical data to support these claims?
not confirmed in the source
Did the author consult an AI while drafting the list?
Yes — the author reports asking GPT-5 for additional failure modes and receiving more examples.
Does the essay conclude humans will be replaced by AI?
The author raises enhancement or replacement as an interpretation but says we are "not there yet entirely," indicating no definitive claim of imminent replacement.
Agents LLM problems observed in humans Published 7 Jan 2026. Written by Jakob Kastelic. While some are still discussing why computers will never be able to pass the Turing test,…
Sources
- LLM Problems Observed in Humans
- Can Large Language Models Simulate Spoken Human …
- Failure Modes of LLMs for Causal Reasoning on Narratives
- A Field Guide to LLM Failure Modes
Related posts
- Commodore 1541 floppy drive packs a 1 MHz MOS 6502 and can run BASIC
- Many Hells of WebDAV: Building a Go WebDAV/CalDAV Client and Server
- RepoReaper: AST-aware, JIT-loading code-audit agent built in Python