Reliability, Quality & Capability

Does AI work well enough? The honest answer to every accuracy and quality concern.

AI is brilliant at the line —

and blind to the whole picture.

Every line is still ours to answer for.

AI cannot be blindly trusted. Every answer below rests on the same principle — AI accelerates, the expert decides.

Every answer on this page reduces to one sentence: the AI's output is never trusted — the engineer who reviews it is. That is also why the pitch is "accelerated expertise," not "AI magic."

Evidence

METR 2026 update and original paper: experienced developers were slower with AI on complex familiar repos.
Faros AI research: AI adoption can raise throughput while increasing review and quality pressure.
DORA 2025: AI amplifies existing engineering strengths and weaknesses.

Reliability, Quality & Capability

AI requires verification on technical topics

Hallucination is a bug — we design as if AI will sometimes be confidently wrong

AI is weakest on the niche edge cases we get paid for

Risk of shipping something half-baked and landing in the 'slop bucket'

Cleaning up AI's mistakes can erase the time it saved

Experienced experts can get slower with AI on complex work

Models may be near their ceiling

AI is poor at pure algorithmic problems

AI loses its guardrails and context over long sessions

AI-generated code accrues tech debt that surfaces versions later

Evidence

On this page