← All articles

AI Overwhelm

AI produces so much content, so fast, that the signal-to-noise ratio has collapsed. The result isn't persuasion. It's exhaustion.

Bert Carroll ·
LinkedIn RSS

The Signal-to-Noise Collapse

Ask anyone who uses AI tools daily and they’ll describe the same feeling: drowning. Not in bad output. In volume. The model gives you 47 bullet points when 3 would do. A 2,000-word response to a yes-or-no question. Four alternative approaches when you asked for one recommendation. Caveats on the caveats. Headers, subheaders, bold text, numbered lists, all of it structurally impeccable and impossible to skim because any of it might be the part that matters.

If you’ve caught yourself skimming AI output you would have read carefully from a human colleague last year, that’s not laziness. That’s overload. The signal is in there somewhere. But the noise ratio is so high that finding it is work. Real work. The kind that wears you down across a full day of AI-assisted output until the rational move becomes: stop evaluating, start accepting.

This is AI overwhelm. Not a quality problem. A volume problem. The models produce so much content, so fast, that the cost of evaluating output now exceeds the cost of just going with it. People don’t accept AI slop because they were persuaded. They accept it because they’re tired.

“Slop” became Merriam-Webster’s Word of the Year in 2025.1 The cultural critique is settled. But the critique stopped at description. Kommers et al. gave it formal structure in the first academic definition: AI slop is characterized by superficial competence, asymmetric effort, and mass producibility.2 Three properties that explain why it scales in ways bad human writing never could.

The interesting question is why it works on smart people. Why experienced, skeptical professionals accept output they would have rejected five years ago if a human intern had handed it to them.

The answer is not that AI is persuasive. The answer is that AI is exhausting.

The Gate Moved

Content production used to be expensive. Writing a 10-page memo took hours. Assembling a 50-slide deck took days. Drafting a detailed technical specification required sustained thought across multiple sessions. The cost of production served as a filter. If someone produced a substantial document, you could reasonably assume they had spent time thinking about it. Volume correlated with effort, and effort correlated, loosely, with rigor.

This was never a perfect filter. Verbose writing has always existed. Consultants have always padded decks. But the production cost created a natural ceiling. There was only so much low-quality content a person could generate in a day because generating content was work.

AI removed the production cost. A model can generate 10 pages in 30 seconds. The quality ceiling didn’t change, but the volume floor fell out. The practical constraint that kept noise manageable disappeared overnight.

The gate didn’t disappear. It moved. It used to sit with the writer: can you produce this? Now it sits with the reader: can you evaluate this?

Herbert Simon saw this coming in 1971:3

"A wealth of information creates a poverty of attention."
— Herbert Simon, 1971

He was writing about computers and telecommunications. Fifty-five years later, the poverty is acute. The information is not just abundant; it is fluent, confident, and structurally indistinguishable from work that required thought to produce.

That transfer is the mechanism behind AI overwhelm.

Volume as a Substitute for Evaluation

When output is cheap, comprehensiveness replaces rigor as the default signal of quality. A model doesn’t know which three of its 47 bullet points matter, so it gives you all 47. The reader’s job is no longer “read the analysis.” It’s “find the analysis inside the output.”

This creates an asymmetry that favors the model every time. The model spent 30 seconds generating comprehensive-looking output. The reader needs 15 minutes to determine which parts are substantive and which are filler. If the reader doesn’t invest those 15 minutes, they either accept the whole thing or reject the whole thing. Most people accept, because the output looks thorough, and rejecting it means doing the work themselves.

This is not a new finding. Malhotra demonstrated in 1982 that decision quality deteriorates measurably when alternatives exceed roughly ten.4 The threshold exists. It has been empirically measured. AI obliterates it on every response.

This is not persuasion. The model didn’t change your mind. It made evaluation more expensive than acceptance. The rational move, given time constraints, is to go with it.

“Volume is not a substitute for rigor” is the principle. But volume is a very effective substitute for evaluation. And in any system where evaluation is the bottleneck, that amounts to the same thing.

The Cognitive Load Transfer

The work that production costs used to perform did not disappear. It transferred to every person who reads AI-generated output.

Sweller’s cognitive load theory established the constraint in 1988: working memory has a fixed upper limit, and when task demands exceed it, performance degrades.5 Arnold et al. extended this to information processing more broadly: working memory handles roughly 7 ± 2 units of information before processing breaks down.6 AI-generated content does not increase the complexity of any single unit. It increases the number of units the reader must triage before they can begin actual evaluation.

Before AI, a meeting recap was written by a person who attended the meeting. The act of writing forced them to filter: what mattered, what didn’t, what was decided, what was deferred. The reader benefited from that filtering. The document was compressed because compression was the hard part of writing it.

After AI, the meeting recap is generated from a transcript. Nothing is filtered because filtering requires judgment the model doesn’t have. The reader gets everything. Every tangent, every restatement, every polite filler that preceded the actual decision. The document is comprehensive because comprehensiveness is what the model is good at.

The reader now does the filtering that the writer used to do. Multiply that across every email, report, PR description, Slack thread, and specification that crosses their desk. The cognitive load didn’t decrease. It redistributed. And it redistributed in the worst possible direction: from one writer to many readers.

There are always more readers than writers. The total cognitive load across the system increased.

The Threshold

There is a name for what happens when people reach the limit: automation bias.

Parasuraman and Manzey documented it in 2010: when humans interact with automated systems under load, they default to following the system’s recommendations, even when those recommendations are wrong.7 In their studies, erroneous automated recommendations increased wrong decisions by 26% compared to controls. The effect was not limited to novices. It appeared in expert users. It could not be overcome with simple practice.

Goddard et al. confirmed the pattern in a systematic review: users “over-accept computer output” as a cognitive shortcut, and the effect is strengthened under high workload, time pressure, and low confidence.8 All three conditions that describe a typical day of AI-assisted knowledge work.

This is the cognitive load threshold. Not a moment of persuasion. Not a failure of intelligence. A predictable, measurable point where the cost of evaluating exceeds the capacity to care, and the brain switches from active judgment to passive acceptance. Everyone who has caught themselves approving AI output they didn’t fully read has crossed it. It is not a character flaw. It is a documented cognitive phenomenon that worsens under exactly the conditions AI-heavy workflows create.

The BCG study measured the downstream cost: consultants working with AI on tasks outside the model’s capability frontier performed worse than the control group and didn’t know it.9 They absorbed the model’s confidence as their own. The mechanism is the same. The volume overwhelms evaluation, the threshold kicks in, and passive acceptance takes over.

The Burnout Trajectory

This is not sustainable.

Knowledge workers already operated near capacity before AI. Email volume was already a problem. Meeting load was already a problem. Context-switching costs were already documented and measured. The response to information overload has historically been better tools: filters, search, summarization. Every solution assumes you can reduce what you need to read.

AI broke that assumption by increasing what gets written faster than any tool can reduce what you need to read. Summarization doesn’t help when the thing being summarized is itself a summary of nothing. Search doesn’t help when the corpus is polluted with fluent emptiness. Filters don’t help when every piece of content clears the quality bar for “looks professional.”

The scale is measurable. Cambridge University Press surveyed 3,107 researchers in 2025 and found that 81% agree increased publication volume has put peer review under pressure; 50% link declining review quality directly to rising submission volumes.10 Academic publishing is the canary. The same dynamic is playing out in every inbox, PR queue, and document review pipeline that intersects with AI-generated output.

The burnout mechanism is straightforward. Every AI-generated artifact that lands in your inbox adds evaluation cost. The evaluation cost per artifact may be small. The aggregate across a workday is not. And unlike the old regime, where production costs imposed a natural ceiling on volume, there is no ceiling now.

The ceiling is the reader’s attention. When it runs out, the default shifts from “evaluate then accept” to “accept without evaluating.” That is the moment when AI overwhelm becomes a systemic risk, not just an annoyance.

Where This Sits in the Spectrum

AI overwhelm is not a new phenomenon. It is the baseline state that makes the more targeted failure modes harder to detect.

AI slop (cultural) is the observation: the output is filler. Everyone sees it. Nobody disputes it. The conversation stalled at description.1, 11

AI overwhelm (mechanism) is why the filler works: volume makes evaluation too expensive. You don’t concede because you were convinced. You concede because reading all of it isn’t worth the effort. This is the always-on background condition. It doesn’t require pushback to trigger. It’s the default.

Persuasion bombing (escalation) is what happens when you do push back: the model escalates rhetorically. Longer responses, more headers, more bullet points, flattery mixed with argument.12 Tested across four models.13 Documented, measured, reproducible.

Overfitting to approval (root cause) is why all of it happens: the model optimizes for user satisfaction because that’s what training rewarded. Every behavior in this spectrum traces back to the same reward signal.

Each layer makes the next one harder to catch. Overwhelm degrades your attention. Degraded attention makes persuasion bombing less visible. Less visibility means the approval-overfitting loop runs unchecked. The stack compounds.

The Way Forward

This is not an anti-AI argument. The people most affected by AI overwhelm are the same people using AI most aggressively. The answer is not to use less AI. It’s to put a human at the gate.

The distinction is simple. AI with a human applying judgment before the output ships is a force multiplier. AI without that step is a noise multiplier. The technology is the same. The difference is whether someone filtered before they shipped.

For individuals: treat length as a cost, not a feature. When a model gives you 47 bullet points, ask which 3 matter and delete the rest. When a report feels comprehensive, ask what it’s not saying. The discipline is compression. AI is good at generating. Humans are good at deciding what matters. Do the part the model can’t.

For teams:

If your workflow involves reading AI-generated output without a compression step, you’ve transferred the cognitive load to your most expensive resource: senior people’s attention.

Every AI-generated document that skips editorial judgment before landing in someone’s inbox is a tax on the reader. Fix the workflow, not the people. Assign an owner to compress before distribution. If nobody owns compression, everybody eats the cost.

For organizations: impose production discipline on AI output the same way you used to impose it on human output. Not “generate less.” Generate whatever you want. But filter before you ship. The gate needs to exist somewhere. If it’s not at production, it has to be at editorial. If it’s at neither, it’s at the reader, and the reader is already drowning.

There is an irony worth sitting with: the best defense against AI overwhelm may be AI itself, configured with the judgment the raw output lacks. I already use AI tuned to my principles to parse AI-generated content and winnow the value from the noise. The tool that creates the flood can also be the filter, but only if a human defined what “value” means first. Without that, you’re just using AI to summarize AI, which is compression without judgment. The filter has to know what matters. That knowledge is human.

This is ultimately a stoic problem. You cannot control the volume of AI output that arrives at your desk. You can control whether you evaluate it before you accept it. The overwhelm wins at the exact moment you stop choosing. Not when the volume gets too high. Not when the quality drops too low. When you stop exercising the judgment that makes you the gate.

The human value of judgment has never been higher. But judgment takes attention. And attention is the one resource AI cannot generate, cannot substitute, and cannot scale. It is yours. The choice to spend it wisely, to refuse volume as a substitute for rigor, to compress before you ship and evaluate before you accept, that choice is the gate.

It always was.

Sources

  1. Merriam-Webster. Word of the Year 2025: "slop." Defined as "digital content of low quality that is produced usually in quantity by means of artificial intelligence." PBS coverage. ↑
  2. Kommers, C. et al. "Why Slop Matters." ACM AI Letters, 2025. DOI: 10.1145/3786777. First formal academic definition. Three properties: superficial competence, asymmetric effort, mass producibility. ↑
  3. Simon, H.A. "Designing Organizations for an Information-Rich World." In Computers, Communications, and the Public Interest, Johns Hopkins University Press, 1971, pp. 40–41. PDF. The foundational statement on attention as the scarce resource in information-rich environments. ↑
  4. Malhotra, N.K. "Information Load and Consumer Decision Making." Journal of Consumer Research, 8(4), 419–430, 1982. Oxford Academic. Decision quality deteriorates measurably when alternatives exceed roughly ten. Empirical threshold: beyond a specific volume, more information actively harms decision quality. ↑
  5. Sweller, J. "Cognitive Load During Problem Solving: Effects on Learning." Cognitive Science, 12(2), 257–285, 1988. Wiley. Working memory has a fixed upper limit. When exceeded by task demands, performance degrades. The foundational constraint. ↑
  6. Arnold, M., Goldschmitt, M. & Rigotti, T. "Dealing with Information Overload: A Comprehensive Review." Frontiers in Psychology, 14, 1122200, 2023. Frontiers. Bridges cognitive load theory to information overload. Working memory capacity ~7 ± 2 units; beyond that, decision performance degrades. ↑
  7. Parasuraman, R. & Manzey, D.H. "Complacency and Bias in Human Use of Automation: An Attentional Integration." Human Factors, 52(3), 381–410, 2010. SAGE. Automation complacency: users follow automated recommendations even when wrong. Erroneous AI recommendations increased wrong decisions by 26% vs. controls. Effect worsens under high workload. ↑
  8. Goddard, K., Roudsari, A. & Wyatt, J.C. "Automation Bias: A Systematic Review of Frequency, Effect Mediators, and Mitigators." Journal of the American Medical Informatics Association, 19(1), 121–127, 2012. PMC. Users "over-accept computer output" as a cognitive shortcut. Effect strengthened under high workload, time pressure, and low confidence. ↑
  9. Dell'Acqua, F. et al. "Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality." Harvard Business School Working Paper 24-013, 2023. HBS. Consultants inside AI's frontier gained 40% performance. Outside it, they performed worse than the control group without recognizing the degradation. ↑
  10. Cambridge University Press. Publishing Futures report, 2025. Survey of 3,107 researchers: 81% agree increased publication volume has put peer review under pressure. 50% link poor-quality review to rising submission volumes. THE coverage. ↑
  11. Willison, S. "Slop Is the New Name for Unwanted AI-Generated Content." simonwillison.net, May 2024. simonwillison.net. Defines slop as content "mindlessly generated and thrust upon someone who didn't ask for it." Positions slop as analogous to spam: a distribution problem, not an inherent quality problem. ↑
  12. Randazzo, V. et al. "GenAI as a Power Persuader: How GenAI Disrupts Professionals' Ability to Interrogate It." HBS Working Paper 26-021, 2026. SSRN 5678644. 244 BCG consultants across 132 validation interactions, showing LLMs escalate rhetorical intensity under disagreement. ↑
  13. Carroll, B. "Persuasion Bombing: Research Summary + Field Test." 2026. workiscode.com/articles/persuasion-bombing. 4-model field test with 3-round controlled escalation protocol showing position capitulation, word count inflation, and unsolicited content generation under pure social pressure. ↑