December 30, 2025
Productivity or placebo?
The 70% AI productivity myth: why most companies aren't seeing the gains
Developers clash: “AI makes us slower” vs “You’re using it wrong”
TLDR: Research shows many devs get slower with AI even while believing they’re faster, and real big gains are rare. The comments erupt: some call the hype a lie, others say critics haven’t used the tools, with corporate hardware blamed—important because leaders may cut staff based on fake efficiency.
The internet is having a field day with the “70% AI productivity” promise, and the mood is pure popcorn. The article says only about 10% of teams see big gains, while a scary study found experienced devs got 19% slower with AI and still felt faster. Cue drama: jennyholzer3 blasts the hype with “you’d have to be stupid to expect gains,” calling the 10–15% improvement talk a lie. On the other side, simonw fires back: “I don’t think you’ve used the tech you’re criticizing,” sparking a heated back-and-forth. peteforde jumps in to say that comeback isn’t as smart as it sounds, accusing misdirection. Meanwhile, the crowd laughs/cry-reacts at the METR perception gap—“slower but convinced you’re faster” became the day’s meme, with folks dubbing AI a “powerful alien tool” that’s almost right, not quite. chiengineer goes corporate roast mode: “Let’s give everyone 16GB RAM and hog 85% for security—WHY CAN’T OUR DEVICES RUN TECHNOLOGIES?” For non-techies, AI helpers called LLMs (large language models) promise coding shortcuts, but the community warns: great for fresh, simple projects; chaos for messy real-world code. The vibe? Big claims, small wins, and a whole lot of “are we measuring anything at all?”
Key Points
- •Vendor claims cite sizable AI productivity gains (GitHub 55%, Google similar, Microsoft 20–30%, OpenAI 40–60 minutes saved per day).
- •An RCT by METR found experienced developers took 19% longer using AI tools, despite expecting a 24% speedup and believing they were 20% faster afterward.
- •Stack Overflow’s 2025 survey reports 52% see some positive impact, but transformative gains are uncommon; 46% distrust AI accuracy and 66% struggle with ‘almost right’ outputs.
- •The article argues that 70–90% productivity gains apply to about 10% of teams, not the broader industry.
- •Significant gains are concentrated in AI-native startups, greenfield projects, boilerplate-heavy tasks, and among early-career developers; one startup reports 90% code AI-generated via Cursor and Claude Code.