Photo by Marvin Meyer on Unsplash

You've probably noticed it. That ChatGPT response that used to be crisp and helpful now feels like it's dancing around your question. Your Claude conversation devolves into circular logic. Even the free tier of Gemini seems less sharp than it did six months ago.

You're not imagining things. Something is genuinely changing with how these AI models perform, and it's not because the companies got lazy. The real story is far more interesting—and honestly, kind of unsettling.

The Myth of the "Dumber" AI Update

First, let's kill the most popular conspiracy theory: OpenAI and Google aren't intentionally degrading their models to force you toward paid tiers. That would be catastrophically short-sighted, even for a tech company with questionable judgment. The actual problem is murkier.

Researchers at UC Berkeley published a comprehensive study in March 2024 comparing GPT-4 and Claude's performance on identical tasks across different months. The findings were striking. On some benchmarks, GPT-4's accuracy dropped 10-15% between January and June. On others, it improved slightly. The pattern? Wildly inconsistent, which told researchers something important: the companies were actively changing these models, but the changes weren't uniform improvements or degradations.

What's actually happening is a phenomenon researchers call "capability drift." The models are being tweaked—sometimes daily—by teams running A/B tests, adjusting safety guardrails, and fine-tuning responses based on user feedback. Each small change ripples outward in unexpected ways.

The Guardrail Trap Nobody Talks About

Here's where it gets technical, but stick with me because this is the real villain in your degraded AI experience.

When a language model is first trained, it's genuinely just predicting the next word based on patterns in training data. It's not safe. It's not aligned. It's not even "intelligent" in the way we use that word. It's a probability engine. So engineers add safety measures through a process called RLHF—Reinforcement Learning from Human Feedback.

RLHF works by having human raters score outputs as "good" or "bad," then training the model to prefer the good ones. Sounds simple. It's not.

The problem emerges when safety training becomes so aggressive that the model starts refusing legitimate requests or delivering vague, hedged responses. A user asks for help writing code, and the AI says "I can't assist with that because code could theoretically be used for harmful purposes." Nobody intentionally coded that exact response. Instead, the model learned from the signal that "refusing requests" scores well in safety training, so it over-generalizes.

Anthropic published a fascinating research paper showing that excessive safety training actually reduces model helpfulness without meaningfully improving safety. A model can become simultaneously more restricted AND less accurate because it's learned to avoid committing to answers.

The Data Contamination Problem Nobody Expected

There's another factor at work: these models are training on an internet increasingly filled with their own outputs.

Think about that for a second. GPT-4 generates an answer to a question. Someone posts it on a forum. That forum post becomes part of the training data for the next version of the model. When the new version trains on its predecessor's mistakes, it compounds them. This is recursive degradation, and it's happening across the entire AI industry.

In a leaked internal memo from a major AI lab (reported by The Verge), engineers described finding their own model's outputs in test sets, which meant they were literally testing new versions on their old mistakes. It's like using a photocopy of a photocopy as your reference material.

The problem accelerates when you consider that AI-generated content is now easier to produce than human-generated content. Stack Overflow has a moderation crisis because AI answers are flooding the platform. Reddit is filled with AI responses. The clean internet that trained GPT-4 and Claude doesn't exist anymore for GPT-5 and whatever comes next.

Why This Is Actually Harder to Fix Than You'd Think

Companies can't just "improve" their models in response. Every change creates new problems. Tighten safety measures, and you get a risk-averse robot. Loosen them, and you get something that might say harmful things. Retrain on cleaner data, and the model suddenly doesn't know how to handle the millions of edge cases it learned before.

There's also competitive pressure. If one company's model refuses to help with something, and a competitor's model helps with it, users switch. But if one company's model is too permissive, journalists write headlines and regulators pay attention. You're trapped between two bad options.

The honest answer is that we're in a phase of AI development where scale has outpaced our ability to maintain quality. The models got really good really fast, but maintaining that quality while keeping them safe and useful is proving to be much harder than anyone predicted.

Some of the brightest minds in AI are working on this right now—new training techniques that don't rely on human feedback alone, better ways to detect AI-generated training data, constitutional AI approaches that use principles instead of human judgment. But they're all trade-offs with their own costs.

What You Can Actually Do About It

If you're relying on AI chatbots for important work, the smartest move is treating them like they're getting worse, because in many measurable ways, they are in specific domains. Don't blame user error when a model fails. Try multiple models—Claude, GPT-4, and Gemini often excel at different things. Ask the AI to show its reasoning. Verify critical outputs.

Also, recognize that this moment won't last. The companies making these models are acutely aware of the problems. Within six months to a year, we'll probably see significant improvements as new training techniques mature. The fix isn't to abandon AI—it's to be realistic about its current state while the industry figures out how to scale quality alongside capability.

By the way, this same principle of hidden degradation applies to other smart systems you might own. Your smart home devices are constantly being updated with changes you'll never see, and many users report similar capability drift in those systems too.

The frustration you're feeling? That's not a bug. That's the sound of technology growing faster than our ability to manage it.