I’ve been thinking a lot lately about why AI sometimes makes work harder instead of easier.
Not wrong. Not broken. Just… harder. Like you asked a clear question and got a five-paragraph essay back that somehow left you more confused than when you started.
Turns out there’s research on exactly this. A peer-reviewed study published in March 2026 in ACM Transactions on Computer-Human Interaction put 34 finance professionals in a room with ChatGPT and a complex financial valuation task. They analyzed nearly 1,200 conversation exchanges. The results are worth understanding if you use AI tools at work.
The finding that surprised me
AI-generated content did help. Professionals who used more of it in their work outputs produced higher quality results. That part matched expectations.
But here’s the thing that stopped me: the cognitive burden from cluttered AI responses — what researchers call extraneous cognitive load — had a negative association with quality roughly three times larger than the task’s inherent complexity. Three times.
So the AI was helping and hurting at the same time. Whether you came out ahead depended almost entirely on how focused the conversation stayed.
What cognitive load theory actually means for accountants
Cognitive Load Theory has been around since the late 1980s. Educational psychologists developed it to explain why some instructional designs work and others don’t. The core idea is simple: your working memory is limited. When it fills up with irrelevant processing, performance drops — not because you’re not capable, but because you literally can’t hold everything at once.
There are two kinds of load that matter here. Intrinsic load is the inherent complexity of the task — a lease accounting analysis under ASC 842 is just hard, and that’s not going away. Extraneous load is the burden caused by how information is presented. The same analysis becomes even harder if the materials are disorganized, redundant, or constantly pulling your attention in different directions.
For decades this theory was applied to textbook design and e-learning. The insight was always the same: even correct content, presented poorly, tanks comprehension. We’re now learning it applies to every AI conversation you have on the job.
The biggest culprit: task switching you didn’t ask for
The study looked at a range of AI behaviors and ranked them by how much damage they did to output quality. The clear winner — and not in a good way — was unsolicited task switching.
This is when the AI decides, on its own, to introduce subtopics you didn’t ask about. You’re working through a specific question about revenue recognition and suddenly the response pivots to something adjacent. The AI meant well. But now you have to stop, evaluate whether that tangent matters, decide whether to follow it, and then try to get back to where you were.
That context switch consumes exactly the working memory you needed for the actual judgment call. The study found this was more damaging than verbose responses and more damaging than incomplete answers. The active misdirection is what does the most harm.
What made this worse: once a conversation got cluttered, it stayed cluttered. Neither the user nor the AI naturally reset. Users didn’t reorganize their next prompt after a sprawling response. The AI didn’t simplify after going off on tangents. Both parties just kept going on their respective trajectories, and the damage compounded across turns.
Two things this means for accounting firms
The first is about how we use AI for research and drafting. When you ask AI to help with a complex standard — an IFRS disclosure question, a tax position, a lease classification — a focused response is genuinely useful. But if the AI proactively introduces adjacent topics you didn’t ask about, you’re now triaging relevance instead of doing the actual work. More output from the AI does not mean better output from you.
The second is about who’s most at risk. The study found that less experienced professionals face a quality penalty from AI-generated cognitive overload roughly three times larger than what their experienced colleagues face. And they don’t automatically compensate by leaning more on AI content the way experienced professionals do.
Think about what that means. Firms that are deploying AI to level up junior staff may be inadvertently exposing their most at-risk users to the most harm. A senior manager navigating a cluttered AI conversation has years of domain knowledge helping them filter the noise. A first-year associate doesn’t. If the AI introduces ten adjacent topics when they needed two, they don’t necessarily know which eight to ignore.
Key Takeaways
- AI content helps output quality — but extraneous cognitive load from cluttered AI responses hurts it 3x more than the task’s inherent complexity. The gain and the cost operate independently.
- Unsolicited task switching is the most damaging AI behavior. When AI introduces subtopics you didn’t ask about, it forces a context switch that consumes the working memory you need for actual judgment calls.
- Less experienced staff face the largest penalties and don’t naturally compensate. Firms should be deliberate about AI training for early-career professionals — don’t assume they’ll navigate the cognitive overhead on their own.
- Practically: keep prompts focused on one task at a time, restart conversations when they drift, and be skeptical when AI introduces topics you didn’t ask about.
Want the CPE credit? Take the full lesson on EverydayCPE and earn 0.2 CPE credits: [lesson link]
Leave a Reply