Ask an AI tool to write the report, fix the code, or draft the pitch, and it will hand back something polished in seconds. Ask the person who submitted that work to explain why it’s structured that way, what assumption it’s built on, or what happens if one input changes, and the answer often falls apart. AI vs human explanation of work has quietly become one of the sharpest tests in offices, classrooms, and interview rooms in 2026, and a growing stack of research suggests a lot of people would fail it.
The Interview Question That Exposes This Directly
The clearest, most literal version of this test now happens in job interviews themselves. Meta began rolling out AI-enabled coding interviews in late 2025, handing candidates an assistant like Cursor, Copilot, or Claude Code instead of a blank editor, and companies including Shopify, Rippling, LinkedIn, Canva, and Uber have adopted similar formats through 2026. The AI is allowed and expected. What’s being graded isn’t whether the candidate can produce working code; it’s whether they can direct, verify, and explain it.
The failure pattern shows up consistently across these interviews: at Canva, interviewers pause after each AI generation and ask candidates to walk through what the code does, and a Rippling candidate was explicitly rejected for relying too heavily on AI even though their initial approach was correct, because they let the AI make the implementation decisions instead of directing them. Getting the right answer stopped being the bar. Being able to explain and defend the answer became the actual bar.
Botshitting: Shipping Work You Can’t Defend
There’s now a specific term for the pattern this creates outside interview rooms, too. A 2026 report from Glean found that nearly 7 in 10 AI users admit to what researchers call botshitting, shipping AI-generated work they haven’t reviewed, don’t fully understand, or couldn’t defend if asked, with heavy AI users, Gen Z employees, men, and managers most likely to do it. The same report found digital workers now spend nearly a full workday every week botsitting, checking outputs, debugging mistakes, and fixing confidently wrong answers, a workload that itself seems to be pushing people toward skipping the review step entirely rather than adding more of it.
When the Work Fails, Who Actually Takes the Blame
The gap between doing the work and explaining it has a psychological cost that researchers can now measure directly. When AI-generated work fails, 40 per cent of workers blame the AI, while only 29 per cent admit it was their own fault, a pattern researchers describe as moral disengagement, the gradual process by which people stop holding themselves accountable for outcomes they had a hand in. A related Boston Consulting Group experiment found something that makes this worse rather than better: when AI was framed as an employee rather than a tool, workers felt less accountable for what it produced and reviewed its output less carefully. The more AI gets treated like a colleague, the less anyone feels obligated to explain what it actually did.
The Accountability Gap Shows Up at the Company Level Too
This isn’t only an individual habit problem. It’s showing up in how organisations track their own AI-generated output. Lanai’s 2026 AI Labour Report found that 92 per cent of leaders say their organisation tracks the financial and efficiency impact of AI-generated work, but only 2 per cent say more than half of that work is actually recorded as a measurable business outcome, a pattern researchers call AI labour orphaning. Separately, Grant Thornton’s 2026 AI Impact Survey found that 78 per cent of business executives lack strong confidence they could pass an independent AI governance audit within 90 days. If leadership can’t reconstruct how an AI system reached a decision, asking an individual employee to explain and defend it becomes an almost impossible position to put them in.
Frontline Workers Are Explaining Decisions They Didn’t Make
A July 2026 Harvard Business Review study, based on multi-year fieldwork across banking, recruitment, and biotechnology, found that frontline employees are increasingly expected to communicate, justify, and defend AI-generated decisions they neither created nor fully understand. The researchers found workers respond in one of three ways: they relay the AI’s output verbatim without really owning it, they mask it by presenting it as their own reasoning even when it isn’t, or they amplify and complement it by adding genuine judgment on top. Only the third pattern actually closes the gap between AI vs human explanation of work in any meaningful way, and it’s also the pattern that takes the most deliberate effort to practice.
The Same Pattern Is Showing Up in Classrooms
This isn’t strictly a workplace phenomenon either. The interview trend at Meta, Canva, and Rippling has a direct parallel in how universities and bootcamps have started grading assignments: rather than banning AI outright, instructors increasingly allow it and then grade the student’s ability to walk through and defend what the tool produced, on the logic that this is now the actual skill being hired for. Reports from hiring managers running technical interviews describe a consistent tell: a candidate who produces an optimal solution unusually quickly but can’t walk through even one worked example, or fails when asked pointed questions about specific lines of their own submitted code. Several interviewers have said plainly that when a candidate can’t explain their own logic, it doesn’t matter whether AI was involved or not: the explanation gap alone is the disqualifying factor.
How to Actually Close the Gap
None of these point toward using AI less. It points toward a specific habit that’s easy to skip under deadline pressure: before submitting anything AI-assisted, being able to state in one or two sentences why it’s structured the way it is, what it assumed, and what would break if a key input changed. That single check is close to what Canva’s interviewers are actually testing for, and it’s the difference between directing a tool and simply forwarding its output. Teams that build this into a review habit, not a one-off audit, tend to catch the confidently wrong answers before they ship rather than after a client or a manager asks a question nobody can answer.
It also helps to treat any AI-generated draft the way a good editor treats a first draft from a junior writer: useful as a starting point, not trustworthy as a finished product. Reading the output slowly enough to notice one specific claim, number, or design decision worth questioning takes a few extra minutes. Skipping that step is exactly what the bot-shitting data above describes, and it’s the gap that eventually surfaces, usually at the worst possible moment, in a client meeting, a code review, or a follow-up question nobody expected to need an answer for.
Conclusion
AI vs human explanation of work isn’t a philosophical debate anymore; it’s showing up as a literal grading criterion in job interviews, a measurable accountability gap inside companies, and a quiet psychological habit of blaming the tool when something goes wrong. AI can produce the report, the code, or the decision faster than any person could. Whether that work holds up under a single good question still depends entirely on whether a human actually understands what was produced, and that understanding has to be built deliberately, because it clearly isn’t happening by default.
