Is GPT4 getting dumber?

Is GPT-4 getting dumber? 97% to 2% accuracy drop

Recent studies suggest is GPT-4 getting dumber remains a significant concern for users facing decreased output quality. Understanding these shifts helps users adjust their expectations and prompting strategies to maintain productivity. Recognizing performance changes prevents frustration and ensures more effective interactions with AI tools during complex technical or creative tasks.

The Great GPT-4 Debate: Is the AI Actually Getting Dumber?

The short answer is yes, GPT-4s performance on certain tasks has measurably declined since its early 2023 release, though not in every dimension. This perception of is GPT-4 getting dumber is backed by hard data from Stanford and UC Berkeley, which showed its ability to solve math problems (specifically prime number identification with step-by-step reasoning) collapsed from 97.6% accuracy in March 2023 to just 2.4% in June 2023.^[1] (reference:0)(reference:1) However, the models behavior is complex, and the reality involves a mix of LLM drift, architectural changes, and shifting priorities at OpenAI.

The Evidence: What the Numbers Say About GPT-4's Decline

Lets cut to the chase. The most dramatic evidence comes from the GPT-4 Stanford Berkeley study summary. When asked to identify prime numbers with step-by-step reasoning, GPT-4s accuracy fell from 97.6% to 2.4% between March and June 2023.^[6] (reference:2) Code generation capabilities also took a hit, with the share of directly executable code dropping from 52% to 10% during the same period.(reference:3) More recently, a 2025 study from the University of Essex found GPT-4s creative performance dropped by 43-49% over 18-24 months.(reference:4)

The model also became more cautious. Its response rate to sensitive questions plummeted from 21% to 5% in just one month (May to June 2023).^[5] (reference:5) Users quickly noticed, flooding forums with complaints about why is GPT-4 lazy—leaving placeholder text, refusing to complete code, and giving vague answers.(reference:6) I’ve personally spent hours re-prompting the model to get it to finish a script it would have written fully a year ago. Its frustrating.

How GPT-4's Performance Changed (March 2023 vs. June 2023)

The table below summarizes the key performance drops identified by the Stanford/UC Berkeley research.

Why Is This Happening? Unpacking the Theories Behind the 'Dumbing Down'

So, what’s the culprit? Its not one simple thing. The leading theory points to OpenAI quietly changing the underlying model architecture to cut costs. Many believe they shifted GPT-4 to a Mixture of Experts (MoE) model, where a network of smaller, specialized models is activated per query instead of the whole massive system.(reference:7) This makes the model faster and cheaper to run, but some argue it sacrificed quality. In other words, the ship of Theseus was gradually replaced, and we’re now sailing on a different vessel.

Another major factor is OpenAI model drift explained and aggressive safety alignment. As OpenAI fine-tuned GPT-4 to refuse harmful requests, they inadvertently made it overly cautious and lazy. The RLHF (Reinforcement Learning from Human Feedback) process, meant to make the model helpful, may have also rewarded shorter, more generic answers to reduce computational load.(reference:8)(reference:9)

Architectural Changes vs. Safety Alignment

Heres a look at the two primary drivers of the perceived decline.

GPT-4 Performance: March 2023 vs. June 2023

The table below summarizes the key performance drops identified by the Stanford/UC Berkeley research.

Task: Math Problem Solving

Only 2.4% accuracy on the same questions, showing a massive decline in reasoning.

97.6% accuracy on prime number identification.

Task: Code Generation

Executable code generation dropped to just 10%, with frequent formatting errors.

52% of generated code was directly executable.

Task: Sensitive Questions

Answer rate fell sharply to 5%, indicating much stronger safety filters.

Answered 21% of a set of 100 sensitive/controversial queries.

A Startup's Struggle with a 'Lazy' GPT-4

A small software startup in San Francisco relied on GPT-4 to auto-generate SQL queries from natural language prompts. In early 2023, the model worked like magic, outputting perfectly formatted, complex queries 90% of the time. The team was ecstatic, and their development velocity doubled.

By the summer, everything fell apart. The same prompts started yielding half-finished code with placeholder text like '...add your own logic here.' The team was furious. They were still paying for the API, but the model's output quality had tanked. One engineer spent three hours debugging a simple script because the AI refused to write more than 50 lines.

The founder later read about the 'lazy GPT' theory online and realized what had happened. The model they had built their product on had been fundamentally changed. The startup was forced to spend weeks adjusting their entire pipeline to work with the less reliable model, a costly lesson in the risks of relying on a black-box AI.