December 2023, Revisited: The Year AI Assistants Became Normal

Our Three-Year Retrospective closes out 2023 with December — a month that mattered less for any single release than for what it confirmed. Twelve months earlier, AI coding assistance had been a novelty with an asterisk. By the end of December 2023, with Copilot Chat reaching general availability in the closing days of the year, it was simply part of the toolchain, and the interesting conversations had moved from whether to how.

Chat arrives in the editor for everyone

GitHub Copilot Chat went GA in late December 2023, available in VS Code and Visual Studio and included for both individual and business subscribers, by then running on GPT-4. After a year of staged betas, the Copilot X promises from that March were finally, mostly, real product.

The shift from autocomplete to chat was bigger than it looked. Ghost-text completion had kept the developer firmly in charge: the AI proposed, you accepted or kept typing. Chat introduced a different grammar — explain this code, write tests for this function, why does this fail — that pulled the assistant into design and debugging conversations, not just transcription. It also introduced a different failure mode: a wrong autocomplete is obviously wrong in context, but a wrong explanation arrives wrapped in confident prose. Teams were just beginning to learn that reviewing AI reasoning is a distinct skill from reviewing AI code.

The year of the first AI-usage policy

December 2023 was also when a critical mass of engineering organizations finished — or finally started — their first formal AI-usage policies. The drafts we saw that winter shared a shape: which tools are approved, what code and data may be sent to them, who is accountable for AI-assisted output, and what disclosure is expected in review.

Three forces made the policies suddenly writable after a year of stalling. First, vendor indemnification had arrived — Microsoft's Copilot Copyright Commitment that September, with similar moves from other vendors — taking the worst-case legal scenario off the table for paying customers. Second, business-tier products now had real controls: data retention commitments, the public-code matching filter, admin policy enforcement. Third, and most honestly, the bottom-up reality was undeniable: developers were using these tools with or without permission, and an unwritten policy just meant an unmanaged one.

The good policies from that era treated AI output exactly like any other untrusted contribution — it goes through review, the human who ships it owns it. The bad ones tried to enumerate permitted prompts. The distinction has held up.

The trust questions that stayed open

What December 2023 did not resolve is worth remembering precisely. Nobody had good answers yet on measurement: leaders wanted to know if the spend was working, and the available metrics — acceptance rates, mostly — measured enthusiasm rather than outcomes. Nobody had settled the junior-developer question: if assistants handle the apprentice work, where does judgment come from? And the review-load problem was just becoming visible: code was getting cheaper to write and no cheaper to review, an asymmetry whose consequences the industry would spend the next two years absorbing.

It was also obvious that December — with Google having just announced Gemini and the model market visibly accelerating — that "AI assistant" would not stay synonymous with any single product. The organizations that wrote tool-agnostic policies saved themselves several rewrites.

Looking back from June 2026

The normalization was permanent; the specifics were not. Copilot is now one assistant among many, the chat paradigm has been joined by agents that plan and execute multi-step work, and the 2023-era policy documents have mostly been revised twice — once for agents, once for the governance and audit expectations that followed enterprise adoption. The questions that were open in December 2023 are the ones that still earn their keep: how to review machine-generated work honestly, how to grow junior engineers, how to measure whether any of it helps. Tools changed; judgment did not.

Those durable questions are the spine of AI-Assisted Software Engineering Fundamentals, which teaches the working practices that survive each tooling generation. And for the people who have to write the policies rather than just follow them, Responsible AI for Engineering Leaders covers governance, measurement, and risk with three more years of evidence than anyone had in December 2023.