Why AI Hasn't Replaced Human Experts (Yet)

Two specific signals stand out:

• Advanced technical questions on Stack Overflow have doubled since 2023. The naive prediction in 2023 was the opposite - that ChatGPT and friends would reduce the need for community Q&A to zero. The opposite happened. People shifted the easy questions to AI and brought a much higher proportion of genuinely hard ones to humans. The volume of those hard ones doubled.
• 75% of developers turn to another human for clarity when they don't trust an AI-generated answer. Not Google. Not a different AI. A person - a colleague, a stranger in a forum, a domain expert. The trust shortcut for an AI answer that feels wrong is, overwhelmingly, going back to other humans.

Read together: AI handled a chunk of routine work and freed people up to ask harder questions - and the answer-of-last-resort for the hard ones is still each other.

Because the same shape applies to using AI for anything consequential - your finances, your health questions, a tricky bit of conveyancing, planning a long renovation, choosing between two pension options.

The mental model that matches both the Stack Overflow data and ordinary lived experience: AI is excellent at the questions where being roughly right is fine. It's wobbly on the questions where being precisely right matters. And the trust shortcut for the wobbly answers is a human who has actually done the thing, not a different AI assistant giving you a second confidently-wrong opinion.

Three rough heuristics, in order of how often they bite:

1. Stakes. If being wrong has a small, fixable downside (a recipe substitution, a tweak to a draft email, a quick fact you'd otherwise google), AI is fine. If being wrong has a large or irreversible downside (medication, money, legal, anything time-locked like a fixed-rate deal), don't trust the AI by itself. Use it to draft the question to ask the human.

2. Specificity. AI is much better at general principles than current, jurisdiction-specific, version-specific particulars. "What is compound interest?" - fine. "What's the actual rate on a UK fixed-rate ISA at NatWest right now?" - check the live page. "What does Section 21 do in English law?" - high-level fine. "Can my landlord serve me with notice this week given my exact circumstances?" - solicitor.

3. Confidence-without-citations. A confidently-asserted specific number, name, date, or quote from an AI without a verifiable link to a primary source is the single most reliable warning sign. Real research-grade sources almost always trip up the citation step. If the AI sounds extremely sure and you can't find the source on its first attempt, the source probably doesn't exist.

Possibly. The Stack Overflow finding is from 2026, three years into the post-ChatGPT era. AI models keep getting better at the hard questions, especially when paired with retrieval. The trust gap might narrow, the doubled volume of hard questions might level off, and in five years the shape may look very different.

But the structural reason humans hold the high end isn't pure capability - it's accountability. A person you ask has reputation on the line in a way that no current AI does. They can be wrong in front of you, walk it back, refine their answer based on your follow-up question, and the next time you ask them something the previous wrongness is part of the relationship. AI vendors are trying to engineer accountability features into the products (citations, confidence scores, source links), but it's still being bolted on to a system that fundamentally generates likely-sounding strings of words.

Until that gap closes - and it isn't close to closed - the right move is to keep using AI for the part it's brilliant at and stay in the habit of asking humans for the part it isn't.