Doing More With AI, Feeling Like I Did Less

Seventy-one percent of CEOs say AI has made them feel like an imposter in their own job. Two-thirds of other senior leaders, in the same 2026 Korn Ferry survey of 10,000 workers and executives, say the same thing. That’s not a stat about junior staff getting replaced. That’s the people with the most experience in the room quietly wondering if they still know what they’re doing.

I know that feeling. Not from a boardroom. From a tech audit.

A year or so ago I built a skill in Claude Cowork that runs a full technical SEO audit – pulls the crawl data, checks the schema, flags the page speed issues, structures the whole thing into a brief. It took months to get right. Every time it produced something wrong I’d go back in, adjust the process, tighten the rules. When it finally worked, it felt genuinely great. A shiny new thing I’d built with my own hands, doing real work.

Then I used it for the fortieth time. And somewhere around then, the feeling changed. I wasn’t proud anymore. I was just running it. And I started asking myself a question I didn’t expect to be asking after fifteen years in this industry: did I actually do that?

The trade nobody’s pricing in

Here’s the thing about that question – it’s not really about the tool. It’s about a specific psychological mechanism, and there’s a name for it.

A 2025-2026 study in the peer-reviewed journal Behavioral Sciences traced the exact loop I fell into. Repeated exposure to fast, high-quality AI output shifts where you locate the reason for your own success. Instead of crediting your judgment, you credit the tool. That shift erodes self-efficacy – Bandura’s term for your belief in your own capability to do a thing. And here’s the part that makes it a trap rather than a one-off cost: lower self-efficacy predicts more AI dependency, not less. The tool succeeds for you, you credit the tool, your confidence in yourself drops, you need the tool more next time.

But it isn’t just the tool doing this. Talk to anyone running a tech audit for a living and you’ll hear the same complaint: a proper audit takes a day. Do it properly – crawl the site, walk it manually, cross-check what the automated tools flagged against what’s actually happening on the page – and a day is realistic. Now imagine you’ve been told you have three hours, because that’s what the AI is supposedly for, because the team’s been cut, because that’s how the tool got sold to the business in the first place. That’s the actual damage. Not AI existing. The pressure to compress a day of judgment into three hours and call the gap “efficiency.”

A separate 2025 study out of Frontiers in Psychology found something worth sitting with here too – it’s not the least confident people who feel AI’s threat hardest. It’s the ones in the middle. Moderate self-efficacy, not low or high, amplifies the sense that AI is coming for what you do. Which tracks. If you’re bad at something, AI replacing it is relief. If you’re excellent, you trust your judgment over the tool’s. It’s the people who are good but not yet certain who feel it most.

Losing the muscle you didn’t know you were using

I still run analytical work. I still do creative work. Those skills feel intact. What’s gone soft is something else entirely.

I take AI-free days on purpose – a full day where I don’t touch the tools for whatever I’m working on. Recently I sat down to do a tech audit the old way, no Cowork, no skill, just me and the site. I spent the first thirty minutes trying to remember how to start. Not the technical knowledge – I still know what schema is, what a bad Core Web Vitals score looks like. What I’d lost was the sequence. What do I check first? What questions am I even asking about this site? The muscle memory for the process had gone quiet, because the skill I built handles that sequencing for me now. I hand it a URL. It handles the “what do I look at, in what order, and why.”

There’s a thirty-year precedent for exactly this pattern, and it comes from aviation. Cockpit automation research going back to the 1990s found a specific split in pilots: their manual “stick and rudder” skills stayed intact under heavy automation, but their critical-thinking skills – situational awareness, diagnosing a problem as it unfolds, navigating without the automated system’s help – measurably degraded. A 2010 FAA study found pilot error was a factor in nearly two-thirds of crashes, with automation complicity a documented driver, largely through complacency: over-trusting the automated system, under-monitoring it.

The mechanical skill survives. The judgment underneath it goes quiet first. That’s not new to large language models. It’s what happens any time a profession hands its thinking, not just its labour, to a system that’s good enough not to be questioned.

Not every part of what I’ve automated fits that pattern, though. Writing is a genuine weakness of mine – always has been. So the process I built for content isn’t “hand it over and walk away.” It’s an interview. I get asked a set of questions, I dictate my answers, and if something’s not landing I stop and redirect the question. That gets compiled into a brief. A writer does a second interview based on that brief. Nothing moves to a final draft without my approval at each stage. AI is doing the part I’m bad at – assembling my scattered thinking into clean prose. I’m still doing the part that’s actually mine: the research, the position, the judgment calls, the course-correcting. That’s the difference between offloading a task and offloading a skill, and it’s the difference the tech-audit example doesn’t have.

The private version of imposter syndrome

I’ve questioned whether I’m good enough at this job since long before AI showed up – whether I’m good enough at SEO, at marketing, at photography. That’s not new. What’s new is the specific accelerant.

Recent research on sense of agency backs this up more precisely than I expected. Higher automation does not produce a stronger felt sense of control over your own work – it produces the opposite. Task load and psychological ownership both mediate a measurable negative relationship between how automated a process is and how much agency a person reports feeling over the outcome, even when the outcome itself improves. Heavier automation reduces how involved you feel in your own process, and that drop shows up as reduced belonging to what you produced. There’s a term for the pattern that follows: “power displacement” – you keep the formal authority and responsibility for the work, while your felt control over it quietly drops.

For me this stays private. It doesn’t show up with clients. It shows up in my own head, late at night, wondering if I’m keeping pace with an industry I used to feel certain about. What’s kept it from spiralling is one habit: when I hand a task to a skill I’ve built, I still go through the output afterward. I check it. I question it. That review step is what keeps “AI did this” from turning into “so I must not be needed” – because I’m still the one deciding whether the output is right.

What happens when nobody checks

I had a client who went all in on programmatic AI content – scale, speed, no guidelines, no human review of any of it. The writing read like keyword-stuffing from 2012. Some of it wasn’t even accurate. For a while, traffic held. Then, over the following year, it collapsed to near zero. The site’s overall quality signal had cratered, and Google noticed before the client did.

That’s not an isolated horror story. It’s the individual-scale version of a documented phenomenon at the scale of the entire model ecosystem. Ilia Shumailov and colleagues published peer-reviewed research on what happens when generative models train on the output of earlier generations of themselves: model collapse. The model’s view of reality narrows. Rare, accurate detail vanishes first. Output drifts toward a bland, confident average. As of April 2025, over 74 percent of newly created web pages contained AI-generated text – which means the training data for the next generation of models is increasingly made of exactly this kind of narrowing, unchecked output. The snake starts eating its own tail, and it’s already most of the way through its meal.

Worth being precise here, because this is where a lot of AI-slop commentary gets lazy: Ahrefs’ own data across 600,000 pages found close to zero correlation between a page being AI-written and being penalised in search. The problem was never that AI wrote it. The problem is scaled, unchecked, low-quality output – which AI happens to make catastrophically easy to produce at volume. Same tool, opposite outcome, depending entirely on whether a human is still checking the work.

The bit nobody prices in

One more thing, quickly, because it changes how “cheap and efficient” actually reads. AI data centres are projected to draw around 565 terawatt-hours of power in 2026. A hundred-word AI prompt uses roughly half a litre of water for cooling. None of that shows up on the invoice for the tool you’re using, but it’s real, and it’s growing faster than the efficiency gains are shrinking it. AI looks cheap right up until you price in what it actually costs to run.

The rule I actually use

If I had to reduce this to one rule, it’s three parts, and none of them are “use less AI.”

Guardrails first – the skill has rules, not free rein. Approval always – nothing moves forward without me signing off, every time, not just when I remember to check. And AI only ever does part of a task, never the whole thing. If you hand over the entire task, ask yourself honestly what you’re actually still doing. You built the skill once. That’s real work. But building it isn’t the same as doing the thing it now does for you, over and over, without you.

Treat AI the way you’d treat a capable junior on your team. You’d hand a junior the parts of a task that are repetitive, time-consuming, and don’t need years of judgment behind them. You wouldn’t hand them the whole strategy and walk away. Same rule applies here.

And take AI-free days. Start small – title tags, a rough content edit, whatever’s low-stakes – and do it without the tool for a day. Then two. It’s not about proving you can do without AI forever. It’s about making sure the muscle you built over fifteen years is still there when you reach for it, and noticing early if it isn’t.

The point was never to reject the tool. I use it every day, across Claude, Gemini and ChatGPT, and some of it makes the work genuinely better – the same notification-driven compulsion loop that hijacks your attention through variable rewards is exactly what’s hijacking people through AI efficiency now, and naming the mechanism is the first step to not being run by it. The point is that the industry conversation is entirely about how much more you can produce, and almost nobody’s asking what you’re trading to produce it. Attention was always the scarce resource, not time – and it turns out the feeling of having actually done something is scarce too, once you start handing that part away as well.

Where to from here? Probably: pick one thing you’ve fully handed over, and go do it by hand once this week. See what’s still there. See what isn’t. That’s the actual data you need, not another productivity stat.

FAQ

What is self-efficacy, and why does using AI erode it?

Self-efficacy is your belief in your own capability to do something well – a concept from psychologist Albert Bandura. Research shows that repeated success using AI shifts the credit for that success away from your own ability and onto the tool, which measurably lowers self-efficacy over time and, counterintuitively, increases reliance on the tool rather than reducing it.

What is “sense of agency,” and why does it drop even when AI improves the result?

Sense of agency is your felt experience of controlling your own actions and their outcomes. Research on human-AI interaction has found that heavier automation reduces this feeling even when the final output is objectively better, because task involvement and psychological ownership of the work both drop as automation increases.

Is AI-driven imposter syndrome a real, documented pattern?

Yes. A 2026 Korn Ferry survey of 10,000 workers and executives found AI use contributed to imposter syndrome in 71 percent of CEOs and around two-thirds of other senior leaders. Some researchers describe this as “competence vertigo” – experienced professionals suddenly doubting real, previously solid expertise once AI is visibly doing part of their work.

What is AI model collapse, and why does it matter for content quality?

Model collapse is a degenerative process, described in a 2024 peer-reviewed Nature paper, where AI models trained on the output of earlier AI models gradually lose the richness of real human data – rare and accurate detail disappears first, and output drifts toward a bland average. With over 74 percent of new web pages now containing AI-generated text, the problem is compounding: each generation of models trains on more of the last generation’s output.

How do I use AI without losing my own skills?

Keep three things in place: guardrails on what the tool is allowed to do, mandatory human approval before anything ships, and a hard limit that AI only ever completes part of a task, never the whole thing. Treat it the way you’d treat a junior team member – hand over the repetitive parts, keep the judgment.

Is AI actually saving people time, or just adding more work?

The data is mixed by design. Roughly 3 in 4 knowledge workers say AI genuinely helps their individual output, but only 13 percent of organisations report performing significantly better as a result, and 67 percent of 2025 AI adopters ended up working more hours by year’s end, not fewer. Individual productivity and organisational benefit are turning out to be two different questions with two different answers.

What’s a good rule of thumb for how much of a task AI should handle?

If you can’t answer what you personally contributed to a piece of finished work, AI handled too much of it. A useful test: could you explain, out loud, the specific judgment calls you made on this task that the tool didn’t make for you? If the answer is nothing, the task, not just the labour, has been handed over.