AI Leadership Disease: Why Groth’s Satire Tracks the Data

Paul Groth, Development Lead at Anthologie Perth, has named a workplace condition the AI industry has been quietly measuring for two years: AI Leadership Disease. His July 9, 2026 essay “Too much computer” on Oton Technology wraps the warning in the language of a pandemic briefing. The diagnosis lands because the failure mode he describes, where AI saves you from thinking and the thinking atrophies, now has its own peer-reviewed and survey backing.

A 2025 Microsoft and Carnegie Mellon study of 319 knowledge workers found that the more confident users were in AI’s ability to complete a task, the less critical thinking they reported applying to the output. The Q2 2026 BairesDev Dev Barometer, drawn from 1,569 developers across 77 countries, found that only 16% of senior engineers say their junior colleagues fully understand the AI-generated code they submit. Both findings sit on the same fault line Groth is naming, between the speed AI delivers and the comprehension that disappears with it.

The Diagnosis

Groth, a software engineer by trade, opens with the framing of a public-health bulletin. The AILD pandemic has spread to 193 countries, he writes, infecting an estimated 43 percent of the population worldwide. The symptoms follow: using AI to answer most questions and pasting the output into group chats; using AI to write social posts and sometimes entire articles, possibly through a fully automated process; not reading any of it, and expecting other people to read it and applaud.

The escalation, what Groth calls the AILD-SS or Super Spreader variant, reaches the point of deploying AI-built code to public servers and enforcing minimum AI usage quotas inside an organisation. He writes the variant plainly as a workplace phenomenon he has watched spread from individuals to policies. The satire tracks because the symptoms he lists are already recognisable in most offices.

What makes AI a great tool, is also what makes it a disaster. That is, responding to questions and tasks that save you time, or allowing you to explore things outside of your expertise in detail.

Groth wrote that sentence in the essay “Too much computer,” published July 9, 2026 on Oton Technology.

Why the Same Property Is the Problem

The mechanism Groth names is mechanical. AI returns text that looks complete, so the user stops reading it the way they would read a peer’s draft. The output takes the shape of an answer and inherits the trust of one. Two of his examples show the asymmetry between the sender and the recipient.

First, sending AI-generated text about a topic outside your area of expertise to a person who has expertise in that topic. The recipient, who can spot a wrong claim, is now in the position of auditing an unsolicited document the sender has not read. Groth’s verdict on that pattern: you have had too much computer.

Second, asking AI to scaffold a project document, sections and brief descriptions of what to write in each, and sending it to a team to fill in. The result, in his experience, makes the work take many times longer than no AI was involved at all. The structure brings unnecessary parts, omits necessary ones, and breaks concepts into an illogical order. The cost of AI’s speed is paid by the people who then have to undo its structure.

The AI Leadership Disease symptoms Groth lists in his essay map onto the same asymmetry:

• Using AI to answer most questions, often copying the output straight into group chats.
• Writing social posts, sometimes entire articles, on a fully automated loop and not reading any of it.
• Building a website on a personal computer, deploying it proudly, and asking colleagues to marvel at the work.
• Adding minimum AI usage quotas to performance reviews from a position of organisational power.
• Sharing AI-generated text on a topic outside your expertise with someone who has that expertise.

What the Software Process Hides

Groth is a software engineer, and his version of “AI did the work” sits inside a process the public never sees.

He uses AI to help diagnose bugs and sometimes to scaffold new code. He modifies the output manually to fit his team’s code standards, runs automated tests, manually checks it for performance and security issues, and verifies it meets requirements and compliance. A second engineer reviews it. Non-engineers and the client test it in other environments. Only when all of that passes does it go to production. Once live, the system is monitored in real time by automated tools and by real people at set intervals.

The chain he describes is standard practice in software work, apart from the AI in it, which is too new for any industry-standard practice to have taken hold. Walking through the chain shows what the public never hears when someone says “an AI-built website.”

So what happens when someone without that expertise skips the chain and ships the AI output anyway? Nothing good. Vulnerabilities at the server level, malware and scams arriving from anywhere else in the world simply because the website exists, and an ever-growing list of other issues. His side note is sharp: did you check the code you deployed publicly for any sensitive information or API keys that might have accidentally been left in there? Whether AI has all the context it needs or not, it cannot be trusted to generate valuable output. Feeding it everything it needs to know is probably worse, because you gave it sensitive information and still cannot trust it.

The Industry Is Now Measuring It

Groth published the essay as a software engineer’s first-hand take. The data backing it has been piling up since 2025.

The Microsoft and Carnegie Mellon study, titled “The Impact of Generative AI on Critical Thinking,” surveyed 319 knowledge workers who self-reported how they used generative AI. The pattern: the more confident a worker was in the AI’s capability for a task, the less critical thinking they applied to the output. Workers with access to generative AI also produced a less diverse set of outcomes for the same task, a finding the authors read as a deterioration of independent thinking. Microsoft Research publishes the 319-knowledge-worker study on AI and critical thinking on its research site.

BairesDev’s Q2 2026 Dev Barometer split a 77-country sample of 1,569 developers into 1,059 juniors and 510 seniors, and asked both sides the same questions. The result that has stayed in the headlines is the 16% figure: only 16% of senior engineers say junior colleagues fully understand the AI-generated code they submit. Juniors, on the same survey, said AI had improved their understanding of software development at a rate of 85%, with 24% admitting outright that writing code from scratch is the task they feel least confident doing without AI help. The same survey found 72% of senior engineers naming critical thinking as the foundational skill they expect from juniors over the next three years. Oton Technology’s own reporting on the survey covers the 16% senior-engineer verdict on junior AI code in full.

The two studies, run with different methods on different samples, describe the same gap from opposite sides. Groth’s diagnosis and the data converge on one sentence: the speed AI delivers is paid for by the comprehension it removes.

When AI Code Goes Public

The cost moves from thinking to infrastructure when AI-built code is shipped without the review Groth describes, and his specific question about leaked API keys is the pre-deployment audit that gets skipped when AI output goes straight to production. The 2026 numbers on what happens when it does are now available across several independent reviews.

IBM’s 2026 X-Force Threat Intelligence Index, as cited in the 2026 review of AI-generated code vulnerabilities on Cycode, found that over 300,000 ChatGPT credentials were discovered in infostealer malware in 2025. The same review reported 81% of organisations now lack visibility into how AI is actually used inside their own codebases. Shadow AI, the use of AI tools not authorised by IT, is now seen as a definite or probable challenge by 76% of organisations, up from 61% the year before. IBM’s Cost of a Data Breach Report, cited in the same review, found that shadow AI incidents raise the average cost of a breach by about $670,000.

Groth’s prompt to non-engineers to check their deployed code is not abstract. The 2026 data on AI-generated code shows the pattern he warns about, in numbers:

• 300,000 ChatGPT credentials discovered in infostealer malware in 2025, per IBM’s 2026 X-Force Threat Intelligence Index.
• 76% of organisations now see shadow AI as a definite or probable challenge, up from 61% the year before, per Cycode.
• 81% of organisations lack visibility into how AI is actually used inside their codebases, per the same review.
• $670,000 average increase in breach cost tied to shadow AI incidents, per IBM’s Cost of a Data Breach Report.

The People Who Get the Output

The thread that ties Groth’s diagnosis to the security numbers is the recipient. The cost of unread AI output lands on whoever has to read it, evaluate it, or clean up after it. In his first example, that is the expert on the receiving end of an unsolicited AI summary. In the second, that is the team asked to fill in a document whose structure was set by a model with no context for the project. In the Super Spreader variant, that is the user visiting a website whose code was never reviewed, or the colleague asked to marvel at the result.

Groth’s single takeaway is a directive: before you hit the share button on something generated by AI, think how it might affect the recipients. Maybe hit delete instead. He pairs the instruction with a tongue-in-cheek plea to use AI in moderation and to seek help if you think you have AILD. The warning is not against the tool. It is against the loop of generating without reading, shipping without checking, and asking other people to absorb the cost of saved time.

Frequently Asked Questions

Who is Paul Groth?

Paul Groth is the Development Lead at Anthologie Perth, an Australian independent impact studio. He wrote the July 9, 2026 essay “Too much computer” on Oton Technology, which introduced the term “AI Leadership Disease.”

What is AI Leadership Disease?

AILD is Groth’s satirical name for the workplace habit of using AI to answer, write, and decide, then sharing the output without reading or verifying it. The escalation he calls AILD-SS, or Super Spreader, covers deploying AI-built code to public servers and enforcing minimum AI usage quotas from a position of organisational power.

What did the Microsoft and Carnegie Mellon study find?

The 2025 study “The Impact of Generative AI on Critical Thinking” surveyed 319 knowledge workers. Higher confidence in AI’s ability to complete a task correlated with less critical thinking applied to the output, and AI users produced a less diverse set of outcomes for the same task than non-users did.

What did the BairesDev survey show about AI-generated code?

The Q2 2026 Dev Barometer surveyed 1,569 developers across 77 countries. 85% of junior developers said AI had improved their understanding of software development. Only 16% of senior engineers said juniors fully understand the AI-generated code they submit, and 72% of seniors named critical thinking as the foundational skill they expect from juniors over the next three years.

How does AI Leadership Disease differ from regular AI criticism?

Groth’s framing focuses on the loop of generating without reading and shipping without verifying, and on the recipient who absorbs the cost. Most public AI criticism centres on job loss, model bias, or AI capabilities. Groth’s argument is operational: the failure mode is the unread, unchecked, unsupervised handoff of AI output to another person or to a live system.