AI Progress in [2026]: What Has and Hasn't Changed

Artificial intelligence has been through an extraordinary period of public attention since ChatGPT's release in November 2022, with claims ranging from "AI will end work" to "it's all hype." Here is the honest assessment of what has genuinely changed, what the technology can and can't do, and what the near-term trajectory looks like.

What Has Actually Changed

Large language models (LLMs) — the underlying technology behind ChatGPT, Claude, Gemini, and similar systems — represent a genuine qualitative leap in AI capability for language-related tasks. The ability to produce coherent, contextually appropriate text across an enormous range of topics and styles, to reason through multi-step problems, to write and debug code, and to engage in sophisticated dialogue represents capabilities that didn't exist at this quality level before 2022. The speed of improvement since 2022 has been remarkable — the capabilities that seemed impressive in GPT-3 are substantially exceeded by models released two years later.

Specific domains where AI has produced measurable, practical impact: software development (AI-assisted code completion and generation is now used by a majority of professional developers, with meaningful productivity effects), customer service (AI handling a significant proportion of customer interactions that previously required human agents), content creation (AI assists or accelerates writing, image generation, and video creation across creative industries), and scientific research (AlphaFold's protein structure prediction being the clearest example, with other AI research tools becoming standard in biology and chemistry research).

What Hasn't Changed

AI systems in 2026 remain tools that require skilled human direction rather than autonomous agents that replace human judgment across complex domains. The "hallucination" problem — AI systems confidently producing plausible but factually incorrect information — hasn't been fully solved, which limits deployment in domains where accuracy is critical without verification. AI systems are also notably worse at novel physical world manipulation, long-range planning, and genuine creativity (as distinct from sophisticated pattern recombination) than the most optimistic claims suggest.

The economic impact of AI, despite genuine productivity effects in specific domains, hasn't yet produced the broad macroeconomic transformation that the most dramatic predictions projected. Productivity statistics show some improvement attributable to AI adoption but not the step-change disruption that "AI will replace X% of jobs" predictions implied for the 2023-2026 period. This could reflect slow adoption diffusion (productivity effects from major technologies often appear in the statistics years after adoption), or it could reflect that the genuine capability gains haven't translated to broad economic productivity in the way the more optimistic scenarios projected.

The Near-Term Trajectory

The most credible near-term AI developments: continued improvement in model capabilities (each new generation of frontier models has demonstrated meaningful improvement), expansion of multimodal capabilities (AI handling text, images, audio, and video in integrated ways), agent systems (AI that can take sequences of actions to complete multi-step tasks with less human intervention), and diffusion of current capabilities into products that make them accessible to users who aren't AI specialists. The more speculative claims — artificial general intelligence within the decade, AI that surpasses human performance across all cognitive domains — remain genuinely uncertain.

My honest take: The AI capability gains since 2022 are real and significant — not hype. The macroeconomic transformation hasn't yet appeared in the statistics, possibly due to adoption lag. AI remains a tool requiring skilled human direction in complex domains, not autonomous replacement. The near-term trajectory is more capability + broader diffusion; the far-term trajectory remains genuinely uncertain.