AI models are getting dumber over time
Why People Believe It
This perception often stems from "drift," where users notice a change in how a model responds to a specific prompt they have used successfully in the past. When a model provides a shorter answer or refuses a request it previously fulfilled, users interpret this as a loss of general intelligence or a decline in capability.
Additionally, the high visibility of "edge case" failures on social media creates a confirmation bias. When a user discovers a new way to "break" a model that seemed robust a month ago, it is often framed as a regression in intelligence rather than a discovery of a pre-existing limitation or a result of updated safety guardrails.
Why It's False
Empirical benchmarks and standardized testing do not support the claim of general cognitive decline. Across industry-standard evaluations—such as MMLU (Massive Multitask Language Understanding) and HumanEval—newer iterations of frontier models consistently show improvements in reasoning, coding proficiency, and factual accuracy.
The perceived "dumbness" is typically a result of targeted optimization rather than a loss of parameters or capability. For example, when a developer implements RLHF (Reinforcement Learning from Human Feedback) to reduce hallucinations or toxicity, the model may become more cautious. This increased caution can be mistaken for a loss of creativity or a decreased ability to follow complex instructions, but the underlying latent space of the model remains as capable as before.
Furthermore, the phenomenon of "catastrophic forgetting" is largely mitigated in modern LLM deployment. Since most commercial models are served as static snapshots (frozen weights), the model does not "forget" how to reason over time; rather, the system prompts and filtering layers surrounding the model are what evolve.
What's Actually Happening
What users are experiencing is a trade-off between "helpfulness" and "harmlessness." As AI providers refine models to be safer and more concise, they often apply stricter constraints. This can lead to "over-refusal," where the model declines a benign task because it triggers a safety heuristic, giving the impression that the model has lost its ability to understand the context.
There is also the impact of "quantization" and "distillation." To reduce latency and operational costs, providers may deploy smaller, compressed versions of a model. While these versions maintain most of the original's performance, they may struggle with highly nuanced logic or very long-context windows compared to the full-parameter version, leading to a perceived drop in quality.
Ultimately, AI development is moving toward specialized mixture-of-experts (MoE) architectures. Instead of one monolithic "brain," the system routes queries to specific sub-networks. If the routing mechanism is tuned to prioritize brevity or safety, the output may feel less sophisticated, even though the total computational intelligence of the system has increased.