The EU votes tomorrow on mass AI scanning for known child abuse images. I support the sunsetting program.
I'm also publishing a preprint suggesting its more expansive sister proposal, still being debated this summer, would catastrophically backfire.
wildetruth.substack.com/p/the-eu-par...
New memoir working title:
Base rate neglected: A mother's guide to high-impact, low-prevalence disasters
New memoir working title:
"All Veras are wrong, but some are useful: A quantitative methodologist seeks meaning in science service"
New memoir working title:
"All Veras are wrong, but some are useful: A quantitative methodologist seeks meaning in science service"
Correction: CC 1.0 (Wed.'s vote) uses hash-matching of known CSAM, a signal detection problem that doesn't face the same false positive math as CC 2.0 (proposed mass scanning for CSAM more generally). My analysis deals w CC 2.0. Still relevant, but my thread overstated the connection. Paper otw.
The nights remain cold, but the days are sunny and warm in Berlin at long last. Fast responders, the sprinkle blossoms are already blooming.
Sorry just realized the opening post does not contain a meme and I regret the error. False flags for everyone!
Some of my previous work on this sort of thing here:
Bayesian backfire summary: www.hertie-school.org/en/digital-g...
Vintage: iborderctrl.no/lie_detection
Wonkier: wildetruth.substack.com/p/multiverse...
Up next: Third trilogue May 4, final deal expected June 2026. National-level implementation decisions after all this political song and dance will matter enormously no matter what passes.
Working paper going up on SSRN soon, hopefully (fighting with it as I post this). Will link when posted.
Note on fightchatcontrol.eu showing Germany "supporting": that seems wrong because it kinda is. Germany's new coalition backed the revised proposal because mandatory E2E scanning was dropped — their historic objection. Not the same as endorsing mass surveillance! Still useful for MEP tracking?
My working paper: at 90% accuracy scanning 10B messages → ~1.6B false positive flags → ~2.2× total EU CSAM law enforcement capacity. Before a single specialist sees a case. Harm dominates across every causal pathway.
What is this? Scanning of common services (Gmail, Messenger, etc) but not E2E chats like Signal / WhatsApp. The Council retreated in Nov. 2025 from a proposed mandatory E2E scanning proposal floating since 2022. The math of the false positive problem doesn't care which haystacks are involved.
Surveillance folks: EU Parliament's LIBE committee votes Wednesday on extending Chat Control 1.0 — a voluntary AI mass scanning regime intended for child sexual abuse material (CSAM), set to expire April 3. What my working paper finds about why such screening backfires. 🧵
Just spoke to a colleague who couldn't get their methodologically superior study published in top journals because it got basically the same effect estimates as vastly inferior methods
bro just one more future study bro, bro I swear just one more future study and it'll fix the inference bro
My parents got me this science chick necklace a few Christmases ago, and now that I finally wear it, man on the street says I have artist vibes. The math → art pipeline is real. Fibonacci! 🌀
New memoir working title:
"No data, no problem: My life as a security researcher"
My progeny have the fine motor skills of a sentient placenta!
There was notable pressure from my home birth midwives 2x to keep it for reasons. I felt it should be used for science. The body did so much work! It's an organ! So much information? But wasn't sure who to call to donate it...
Exploring the iWill Covid vaccine risk simulator from Charité: iwill.charite.de
Love the icon arrays to illustrate outcomes for vaccinated vs. unvaxx / 10,000, from side effects thru hospitalization to death.
Love the personal/societal split levels.
Lots to think about here for risk literacy.
Never fear, it's LOAD-BEARING turtles all the way down!
Schölkopf & co (h/t @p-hunermund.com) on the dangers of simplifying for DAGs, Hardt & co on performative power & how classification IS causal modeling, & a cool study on how much scaffolding matters.
wildetruth.substack.com/p/scaffoldin...
Oh this is great, thank you! Robins' separable treatments work sounds directly relevant to the aggregation problem I'm chewing on (e.g., dedicated vs casual attackers on the strategy pathway in screenings). Would you have a suggested reading list for his work on this? I'd appreciate a reading rec.
Cos graphs help you find things you don't know (colliders, confounders, identification strategies). That's one reason to draw one. You don't draw a DAG cos you think you have everything right; you draw it to reason about causal structure better. But the graph can't check its own variable validity.
Well, the graph is in part a heuristic to help you find and prevent logical errors in causal thinking. But it can't tell you if your variables are valid abstractions. If they're bad compressions (like LDL+HDL→cholesterol), the DAG looks fine while tossing causal salad. That's Rubenstein et al. 2017.
Ha! Not pessimistic — just honest about the recursion. I didn't say progress can't be made!
Rubenstein et al (ty!) call it "turtles all the way down." I'm just taking Schölkopf's own group at their word. 😄
Progress = better scaffolding, not life without turtles. That's the optimistic reading!
Reeling a bit from the recursive nature of error here and in my own work.
Are we just always blind to our own representational or indeed epistemic limits, even when describing those limits in others?
"We dance round in a ring and suppose,
But the Secret sits in the middle and knows." (Frost)
That is, the graph encodes causal relationships, but has no notation for whether your variables are valid abstractions. Aggregated LDL and HDL in one cholesterol variable is the classic example.
According to Rubenstein et al 2017 ("Causal Consistency of Structural Equation Models"; h/t @p-hunermund.com), DAGs have the same blind spot on another level. www.auai.org/uai2017/proc...
In *Causality,* Judea Pearl argues that SEM's founders kept causal meaning in their heads, but didn't encode it explicitly in the notation — and later generations lost track of that causal intent and correct reading.
(I never had a wedding, either. Maybe next time!)
