Everybody Says No to AI. Almost Nobody Behaves Like It.

Two numbers have been doing a lot of work in the AI conversation this year.

The first: 60% of US consumers say an "AI-generated" label on brand content is off-putting. The second: 95% of enterprise generative-AI pilots produce no measurable impact on profit and loss.

Put together, they tell a tidy story — nobody wants this stuff, nobody's making money from it, the whole thing is a bubble waiting for a pin. It's a satisfying story. It's also the wrong reading of both numbers.

Because underneath each one is a gap between what people say and what people do — and the two gaps run in opposite directions. Consumers say they reject AI content considerably more than their behaviour suggests. Companies say they're getting transformative value considerably more than their accounts suggest. And the capital expenditure — somewhere around $700 billion across the four US hyperscalers this year — is being justified against the stated version of both.

That's the actual story. Not "AI is a lie." Something more uncomfortable: the two markets that are supposed to be validating this buildout are both running on self-report, and self-report is unreliable in the direction that flatters whoever is answering.

Part One: The content nobody admits to watching

What AI can actually produce right now

Start with capability, honestly assessed, because most of the argument skips it.

For short-form commercial text — product descriptions, listing copy, summaries, first-draft marketing — the machines are at or above the median human professional, and have been for a while. For image generation, the ceiling is high enough that stock photography as a category is in real structural trouble. For short video, quality is now good enough for social-length content and advertising cutdowns, though not for anything demanding sustained continuity. For music, generated tracks are convincingly competent in genre pastiche and thin in everything that makes a specific artist worth following twice.

What it still can't do: hold a coherent argument across long form without drifting, be reliably specific rather than plausibly generic, or produce anything whose value depends on the audience believing a particular person meant it. That last constraint is the important one, and it's not a capability problem. It's a provenance problem. No amount of model improvement fixes it, because the thing being purchased is authorship, and a machine cannot supply it by getting better.

Hold that thought — it explains almost everything about the backlash.

What users say they reject

The stated-preference data is consistent and it is brutal.

A WordPress VIP survey of 2,000 US consumers in April 2026 found that 60% view AI labels on brand content as off-putting, and that 86% don't fully trust AI-generated content, preferring to verify claims at the original source (reported here). Clutch, surveying in June 2026, found 33% say a brand's use of AI worsens their perception of it, against just 16% who say it improves it (Clutch). A Baringa study on US consumers found around a quarter now prefer generative-AI creator content, down sharply from a majority who said so in 2023 (Baringa).

The word "slop" did real work getting us here. It gave people a label for a feeling they already had, and once a category has a contemptuous name, the social cost of being associated with it rises fast. That's why you now see the anti-AI brand as a deliberate market position — not a values statement, a differentiation strategy (The State of Brand). Publishers and agencies have started declaring war on it publicly (Fortune).

If you stopped reading the research here, you would conclude that AI content is commercially radioactive.

What they actually do

Then you look at behaviour, and it stops being simple.

The Bynder study — worth flagging clearly, this is 2024 data that gets recycled as current, so discount it accordingly — put two articles in front of 2,000 UK and US participants, one written by ChatGPT and one by a trained copywriter, without saying which was which. 56% preferred the AI version. The same respondents, asked about AI copy in the abstract, said they'd disengage from it — 52% of them (Bynder).

That's the whole phenomenon in one study. Blind, they prefer it. Told, they reject it.

Germany's NIM ran the cleaner version of this experiment on advertising, and found the effect is caused by the label itself: identical creative, evaluated more critically when disclosed as AI-generated — seen as less natural and less useful, despite being literally the same content (NIM). The paper's title is the finding: transparency without trust.

This is not a quality judgement. It is a disclosure penalty. People aren't detecting badness; they're responding to a category signal, and then reverse-engineering a quality complaint from it.

The reality: scale is beating preference, and the algorithm noticed first

The best single piece of evidence I've found on what's actually happening at scale is a longitudinal study of Kuaishou — a Chinese video platform with over 400 million daily active users — covering June 2024 to May 2025 (arXiv). It's the rare study that measures creators, consumers, and the ranking algorithm at once, across a real platform rather than a survey panel.

The findings are more interesting than either camp's talking points:

AI-content creators publish substantially more — a median of four more videos than matched human creators.
Per item, viewers engage less. Across 47,288 matched interactions, AI content had a lower valid-view rate (a difference of −0.076), lower full-view rate (−0.027), and shorter view duration (a median of −0.213 seconds). Small per-video, consistent, statistically robust.
In aggregate, the two groups earn comparable engagement anyway — because volume compensates for the per-item deficit. The authors call this scale over preference.
And the recommendation algorithm is already discounting it. Across 178,854 matched video pairs, AI content received lower cumulative exposure — a median of 59 fewer impressions — and its exposure lifecycle was compressed by two days.

That last bullet is the one nobody is talking about, and it matters more than the survey data.

The correction to AI slop is not arriving through disclosure labels, and it is not arriving through consumer virtue. It is arriving through the ranking layer, which measures dwell time rather than opinions and has no particular feelings about authorship. Platforms don't need to detect AI content to suppress it. They need only to notice that people don't watch it as long — which they already measure, precisely, for every item.

So the honest answer to "what is the reality" is this: users reject AI content less than they claim and less than they believe, but the systems that distribute content are down-weighting it anyway, on evidence users don't know they're generating.

Trajectory

Three things follow, and I'd hold them with different levels of confidence.

High confidence: disclosure infrastructure becomes mandatory and mostly ineffective at changing behaviour. The plumbing is already in. C2PA — the cryptographic provenance standard backed by Adobe, Microsoft, the BBC, Sony, Google, OpenAI, Meta and TikTok — is now wired into automatic platform labelling. Between January and August 2026 the IAB shipped an AI transparency framework, Google launched an ad-specific disclosure panel, Meta updated its "AI info" labelling with detected-but-undisclosed labels advertisers cannot get removed, and the EU AI Act's transparency article became applicable. TikTok has labelled over 1.3 billion videos with AI provenance data (platform comparison). Given the NIM result, universal labelling means a universal disclosure penalty — which mostly means everyone eats it and it stops being a differentiator.

Medium confidence: "human-made" becomes a paid tier, not a default. If provenance is cryptographically verifiable, it's sellable. Verified human authorship is the kind of scarcity that finds a price.

Lower confidence, but this is where I'd put money: the volume strategy stops working within about eighteen months. Scale-over-preference is only stable while ranking systems under-price the per-item engagement deficit. The Kuaishou data shows they're already correcting. Once that correction is fully priced, the economics of flooding a platform with cheap generated content invert — you'd be paying to produce inventory the distribution layer won't carry.

Part Two: The payoff that hasn't landed

Now the second story — and watch the same structure appear.

The 95% number, handled properly

The statistic doing the work here comes from MIT's Project NANDA, in a report titled The GenAI Divide: State of AI in Business 2025. It found that despite $30–40 billion of enterprise investment, roughly 95% of generative-AI pilots delivered no measurable P&L impact, while about 5% were extracting real value.

Three caveats that almost never travel with the number, and all three matter:

It's from July 2025, not this year. It gets cited as current constantly.
The sample is modest — 52 executive interviews, 153 survey responses, and analysis of 300 public deployments.
It is explicitly preliminary and was not peer-reviewed.

None of that makes it wrong. It makes it a directional signal that has been carrying far more argumentative weight than its method supports. If you're going to quote it on air, quote the caveats with it.

The corroborating figure is sturdier and less quotable: McKinsey's 2026 State of AI survey found fewer than 20% of AI pilots reach enterprise-scale production. The other 80% sit as proofs of concept indefinitely, consuming budget.

Why it isn't landing

The failure mode is boringly consistent, and it isn't about model quality.

Pilots are run in conditions that don't resemble production: clean data, an enthusiastic team, no procurement, no change management, no integration with the systems where the work actually happens. The measured gain is real in the pilot and evaporates on contact with the organisation. The pattern separating the 5% from the 95% is reported as agents wired into real institutional data and real workflows, rather than chatbots wrapped around a system prompt.

And here's where the mirror to Part One snaps into focus. Enterprise AI value is overwhelmingly self-reported. Someone estimates hours saved. Someone attributes a revenue movement. Almost nobody runs a controlled comparison, because controlled comparisons are expensive and politically unattractive once a CEO has publicly committed to an AI strategy. The incentive runs exactly one way.

So: consumers under-report their AI consumption because admitting it is low-status. Executives over-report their AI returns because admitting otherwise is career-limiting. Same bug, opposite sign.

For balance, the upside is real where it lands: the organisations that do get it working report roughly $3.70 back per dollar spent, and the top few percent of deployers attribute 5% or more of EBIT to AI. The distribution isn't "nothing works." It's "a small number of things work very well and are hard to copy."

The data-centre economics

This is the part of the brief where the claim was "data centres cannot pay for themselves." The accurate version is narrower, stranger, and more interesting.

The four US hyperscalers — Microsoft, Alphabet, Amazon, Meta — bought $433.9 billion of property and equipment in the four quarters through March 2026, against roughly $149 billion of reported depreciation over the same period. Q1 2026 capex alone hit $129.8 billion, up around 80% year over year, and 2026 guidance sums to roughly $700 billion (Silicon Analysts).

That gap between what's spent and what's expensed is not a scandal by itself — it's how capital accounting works. Servers depreciate over five to six years, buildings over twenty-five to forty. Today's income statement carries a fraction of today's build.

The part that deserves scrutiny is that the schedules have been getting longer while the assets have been getting hotter. One analysis puts the cumulative effect of depreciation-schedule extensions at roughly $200 billion of suppressed depreciation through 2028 — about $46 billion in 2026, $75 billion in 2027, $107 billion in 2028 (Footnote Brief). Stretching a GPU's assumed useful life flatters current earnings. It's defensible if the hardware really does stay productive that long. It is a large bet that it will.

Meanwhile the cash isn't covering it. Hyperscalers have moved to external financing at scale — Meta priced $30 billion in October 2025 plus a roughly $27 billion off-balance-sheet vehicle; Alphabet raised about $25 billion in November 2025 and roughly $31 billion more in February 2026; Amazon added $24.9 billion after $15 billion in November (FactSet). Capex is outrunning operating cash flow, and the difference is being borrowed.

Then there's the circularity. Roughly $800 billion of arrangements now exist in which chip makers and cloud providers invest in AI companies that immediately spend the money buying their products. Nvidia has been working on deals worth more than $750 billion, including a reported $250 billion backstop for OpenAI tied to a 10-gigawatt Ohio campus (Bloomberg). Suppliers financing customers to buy supplier product is the specific pattern that preceded the telecom bust, and it is being noticed — Jim Cramer among others has drawn the dot-com comparison directly (CNBC). And OpenAI is reportedly on track to lose around $14 billion in 2026, roughly triple its 2025 losses, while projecting $100 billion of revenue by 2029.

The fair statement is not "data centres cannot pay for themselves." Modelling suggests hyperscaler AI revenue is currently running at roughly the level of the associated depreciation — approximately breaking even on that specific measure (Edward Conard's macro roundup). The fair statement is: they are currently just about covering their own accounting cost, on depreciation schedules that have been lengthened, funded increasingly by debt, with a meaningful share of demand supplied by vendors financing their own customers.

That's not a lie. It's a structure with no slack in it. It works if AI revenue keeps compounding steeply. It breaks in an ugly, correlated way if revenue merely grows well.

What would change my mind

I'd rather be specific about the exit conditions than sound confident.

On the content side, I'd abandon the scale-over-preference thesis if the Kuaishou exposure findings fail to replicate on a Western platform, or if a major platform demonstrates that disclosure labels change consumption behaviour rather than just stated attitudes. Neither has happened yet.

On the enterprise side, the 95% figure would deserve retirement the moment a large, properly controlled study measures enterprise AI returns against a real counterfactual rather than a self-assessment. The absence of that study — in a market spending hundreds of billions — is itself the most telling fact in this entire piece.

On the capex, the number to watch isn't quarterly spend. It's whether anyone shortens a depreciation schedule. A hyperscaler voluntarily reducing the assumed useful life of its AI hardware would be the first honest signal that the assets are ageing faster than the accounting says. Watch for that before you watch anything else.

The through-line

The disillusionment narrative is half-right in both halves, and wrong about why.

People are not rejecting AI content because it's bad. They're rejecting the disclosure of it, while consuming the undisclosed version at a rate they'd deny — and the platforms are quietly correcting for it through ranking, which is the only mechanism in this story that measures behaviour instead of opinion.

Companies are not failing to get value from AI because the technology doesn't work. They're failing to measure it, in an environment where nobody is rewarded for measuring honestly — which means the 95% failure figure and the transformation claims are both self-reports, and neither should be load-bearing.

And roughly $700 billion of this year's capital is being deployed against those self-reports.

The bubble question is usually asked as "is the technology real?" That's the wrong question; the technology is obviously real and obviously useful. The better question is the one the accounting keeps hinting at: how much of this buildout is underwritten by what people said, rather than by what they did?

Nobody has that number. That's the story.

A note on sourcing

Because this piece leans on statistics, here's what I'd want a listener to know before repeating any of them.

Primary and solid: the Kuaishou study (peer-review status aside, it's a real longitudinal analysis with disclosed method and sample sizes), and the hyperscaler capex and financing figures, which come from company filings via financial analysts.

Directional, quote with caveats: the MIT NANDA 95% figure — July 2025, small sample, explicitly preliminary, not peer-reviewed. The depreciation-suppression estimate through 2028 is one analyst's model, not a reported figure.

Weakest, and reached through secondary aggregation: several of the consumer survey percentages, including the WordPress VIP and Clutch numbers. I have not read those questionnaires. Survey wording drives results heavily in this area, and marketing-adjacent research has an obvious interest in a dramatic answer.

Explicitly stale: the Bynder blind-test result is 2024 data. It's the sharpest illustration of the disclosure penalty I know of, which is exactly why it keeps getting recirculated as though it were current. It isn't.

// Related