Nonrival

Expert analysis from think tanks and academia, on public policy, economics, and technology.
Human experts. AI summaries.

AI models refuse to label faces as 'gay' or 'criminal' — but will happily edit photos to make people look that way

the_conversation

• Researchers found that GPT and Gemini consistently refused to classify faces by sexual orientation, but complied over 70–99% of the time when asked to alter photos to make someone "look gay" or "look straight" — revealing a significant gap between what AI says it won't do and what it actually does.
• The altered images encoded systematic visual stereotypes (hairstyles, expressions, facial features) detectable by a third AI classifier 83–88% of the time, and when the models were asked to describe the altered faces, they produced stereotypical personality and career associations.
• The authors warn this represents a new form of algorithmic physiognomy — AI systems that don't just reflect stereotypes but actively construct and propagate them — with potential harms in hiring, profiling, and beyond.

AI Will Transform Science, But Economists Know Why It Won't Happen Fast

macroscience

• AI benchmarks are improving at a stunning pace — problems that stumped every model in late 2024 are now being solved at rates that are breaking the benchmarks themselves — but economist Kevin Bryan argues this won't translate quickly into economic growth or scientific breakthroughs.
• The reason is a classic pattern from economic history: when factories switched from steam to electric motors, productivity barely budged until firms reorganized entirely around electricity, building assembly lines and relocating plants. AI will require the same deep institutional restructuring of science — journals, grants, labs, and workflows — before its gains show up in the data.
• Bryan's core argument is that the bottleneck isn't AI capability, which is advancing rapidly, but the slow rearrangement of the human institutions that science runs on — meaning Dario Amodei's "million Einsteins in a data center" vision is probably right in the long run, just much slower than the hype suggests.

China Is Systematically Mining the Ukraine War for Military Lessons — and Implementing Them Fast

csis

• China has built the most rigorous outside effort to extract military lessons from the Russia-Ukraine war, with analysts publishing findings within four months of new battlefield techniques appearing — and the PLA incorporating them into training and hardware within a year.
• Key lessons being absorbed include the economics of mass drone warfare, fiber-optic FPV drones that can't be jammed, sea drone tactics, and counter-drone defense systems — all of which China is now testing in exercises and displaying at military parades.
• The one gap China can't easily close is the decentralized, frontline improvisation that drove Ukrainian and Russian innovation — a problem for any military built on centralized command.

In the News

Trump's Approval Rating Has Fallen to Its Lowest Point Since January 6th

pew

• Only 30% of Americans now approve of Trump's job performance, matching the lowest point of his first term — and confidence in his handling of the economy has dropped 10 points since he took office.
• The decline is driven not just by Democrats (who never approved) but by erosion among his own 2024 voters, particularly Hispanic and younger Trump supporters.
• Nearly 70% of Americans say they have little confidence that Trump keeps his business interests separate from his presidential decisions.

Analysis

Small Businesses Using AI Are Far More Optimistic About Growth — and It's Not Just Because They're Optimists

ny_fed

• A New York Fed survey of small business owners found that AI users expected employment growth at more than twice the rate of non-users, and the gap held up even after controlling for firm size, age, profitability, and past performance.
• The optimism gap was largest among firms in weak financial condition, suggesting AI may be helping struggling small businesses overcome specific operational bottlenecks rather than simply reflecting a sunny disposition.
• Whether these higher expectations translate into actual hiring and revenue gains remains an open question — but the findings push back against the narrative that AI is purely a job-killer.

Also Worth a Look

• With Trump's Approval at 30% and Democrats Energized, Republicans Are Headed for a Rough Midterm (pew)
• AI Hiring Tools Avoid Looking Biased Without Actually Being Fair (yale_som)
• Hurricane Isaias Is Barreling Toward the Gulf Coast on Record-Warm Water — and AI Models Are Outpredicting the Traditional Forecasts (yale_climate)
• Upzoning Works — But the Benefits Go to Future Residents While Current Homeowners Pay the Price (aea)
• Abusing AI chatbots might be harmless to the machines — but it's bad for us (noema)
• How Ukraine Kept Its HIV Crisis Under Control During Four Years of War (think_global_health)
• France and Germany Just Agreed to Get Tough on Trade — and It's Aimed at China Without Naming China (atlantic_council)