New working paper with Pengxiang (Shawn) Zhou and @GoliAlii : "LLMs as Gatekeepers: Source Concentration, Factual Quality, and Political Slant in Information Search." (https://t.co/1UMoaXoTcT)
More and more, people ask ChatGPT, Perplexity, Gemini, Claude, or Grok instead of searching. That means these tools increasingly decide which news outlets and media sources people actually see. Focusing on the news and media ecosystem, we looked at how AI compares to regular search — and what happens as publishers start blocking AI crawlers.
A few things we found:
1. AI platforms send traffic to a far more concentrated set of sources than organic search (3–8× higher HHI), but currently of higher average factual quality and comparable political slant.
2. The publishers blocking AI crawlers are disproportionately the largest, most factually reliable, and centrist-to-left outlets. Their exit shifts the available pool toward higher concentration, lower factual quality, and rightward slant.
3. Blocking Perplexity's crawler cuts a domain's referral traffic by 46–48%. That traffic flows back to the largest remaining outlets, reinforcing concentration rather than diluting it.
4. The result: redistribution reinforces concentration while failing to undo the shift toward less factual and more right-leaning publishers.
In the case opposing them to Francesca Gino, HBS is now moving to "compel non-privileged deposition testimony" from Gregory Burd and Ben Edelman.
HBS believes they helped Gino fabricate evidence "proving" that a dataset was not fabricated.
If Computer Science can publish consequential/impactful research in ONE round, every other discipline can too. It's a matter of time before the increase in submissions makes the journal model unsustainable.
Yes. Computer science has the best research culture in my experience.
At this point, its impact speaks for itself.
Other disciplines like economics and the political economy of development could learn a lot to reach their ultimate potential.
How did they pull it off? Thread brought to you from .@mlxdoing .@DevEconX.
@ipeirotis If there is automated reviewing, there should be automated responding too. Why would I need to get involved after the first submission? Pretty sure two agents, one at each side, can figure things out and agree on something.
My dad, brother, and sister have a Ph.D. My grandfather and aunt have a Ph.D. My sister-in-law has a Ph.D. My great-aunt, great-uncle, and their daughter (who's my age) have PhDs. Her husband has a Ph.D.
AI review tools are now being built for both authors and journals. What happens when both sides of peer review use the same AI? Could review become AI-to-AI? If correctness is easier to verify, do journals mainly filter relevance? Some thoughts here: https://t.co/tTVKtLXZZZ
AI is reshaping how Americans buy homes, get mortgages, and rent apartments. A thread 🧵
1/ Conversational search is changing home discovery. Redfin's AI search showed users nearly 2x as many listings and made them 47% more likely to book tours. Zillow followed with its own AI Mode in March 2026.
2/ AVMs are powerful but fragile. The LA wildfires were a stress test no algorithm could pass — prices in the Eaton fire zone dropped 62%. An algorithm trained on comparables can't detect that a house burned down.
3/ AI mortgage underwriting cut approval times from days to under 60 minutes. But speed ≠ fairness. The MA AG settled a fair lending case in 2025 against a lender whose AI produced disparate outcomes by race. AI can infer protected characteristics from permissible variables.
4/ Tenant screening AI is generating lawsuits. Algorithms disproportionately deny Black and Latino renters — often based on incorrect or outdated data.
5/ The structural story: Rocket acquired Redfin for $1.75B → 14 petabytes of property data across search, brokerage, and mortgage. This AI arms race heavily favors data-rich incumbents.
Full piece: https://t.co/r9K7wjh9x6
Flying back from Cape Town, where I raced the 2026 @CapeEpic. 8 days. 700km, 16,000 meters of climbing. Done. What an absolutely brutal, beautiful race. Grateful for every climb, every descent, and an incredible teammate. Already thinking about next year 🚵♂️ #CapeEpic#MTB
New research from R.E.A.L. maps commuting fuel costs at the census block group level across the four largest U.S. metros — and the differences are striking.
🔑 Key findings:
• Los Angeles households spend a median of ~$26/week on commuting fuel — nearly $1,360/year — driven by long commutes and $4.55/gallon gas.
• New York sits at just ~$10/week, thanks to its transit network cutting car dependence.
• Dallas and Chicago land in the middle at ~$16/week each.
• The hardest-hit? Middle-income households ($50K–$100K). They bear the highest absolute fuel costs — they drive more than low-income households but commute farther than high-income ones.
With geopolitical tensions pushing fuel prices higher, these numbers are likely a floor, not a ceiling. If LA prices return to their 2022 peak, median annual commuting costs there would jump to ~$1,940.
The U.S. car dependency problem isn't just an urban planning issue — it's a household finance issue.
📖 Full analysis: https://t.co/isY7p8KzEC
Where grocery stores are located is not random.
In this R.E.A.L. post, we map grocery store locations across California and link them to neighborhood demographics. Premium grocers cluster in high-income, highly educated renter areas, while store closures disproportionately affect lower-income communities.
A small data look at how retail geography reflects urban inequality.
https://t.co/MSISPSYMza