Not looking good for #OpenAI. I’m guessing they will slow things down, especially if investors get spooked. Each incident disclosed is worse than the previous one. www.nytimes.com/2026/09/25/t...
Who's Who of AI
Rafael M Batista
Why they matter
Directory member with public evidence across AI research.
- AI signals
- 13 past 30d
- Sources
- 11 distinct domains
- Discusiones
- 4 past 30d
- Latest signal
- 10d ago
Behavioral Scientist.
Lately, I've been thinking (and posting) about: AI+Psych, Personal Finance, Consumer Behavior. Civically engaged, so I occasionally post about that too.
What they're sharing
nytimes.com
AI systems out-persuade expert humans arxiv.org
AI agents are checking the scientific literature — and spotting decades-old errors nature.com
Declaration — Math and AI mathandai.org
Opinion | Bernie Sanders: A.I. Is a Public Resource. You Should Own Half of It. nytimes.com
Separating AI’s Technological Problems From its Capitalism Problems techpolicy.press
theatlantic.com
ft.com
View on Bluesky →
A Flaw in ChatGPT’s Mac App Could Have Let Hackers Grab Sensitive Data wired.com
The efficiency-gain illusion: People underestimate the rate of AI use and overestimate its benefits on simple tasks arxiv.org
A light-touch AI literacy intervention helps protect against AI political persuasion arxiv.org
Attackers Steal METR API Key and Consume AI Credits Worth About $600,000 thehackernews.com
Articles & links
Rafael M Batista reposted
Interesting pre-print on persuasive capabilities of LLMs. arxiv.org/abs/2606.16475
AI Weekly's analysis
→
- Across 18,978 conversations with 6,923 people, AI systems reliably out-persuaded expert humans, including world championship debaters and professional canvassers.
- Experts chose their topics, researched in advance, went through hours of structured practice, and were paid £1,000 cash bonuses, and still lost to AI.
- In a live fundraising test for Save the Children, AI was nearly 3x more effective than professional canvassers at raising real donations.
Read full analysis →
View on Bluesky →
Rafael M Batista reposted
AI agents are checking the scientific literature and spotting decades-old errors www.nature.com/articles/d41...
AI Weekly's analysis
→
- A Zhejiang Lab chemist's AI predicting boiling points clashed with a 75-year-old reference database; manual checks showed the database, not the model, was wrong.
- The same AI spotted further mistakes in older papers and reference books, including a typo and incorrect values of century-old boiling-point measurements.
- Researchers caution AI fact-checkers are not reliable on their own because the models make mistakes like humans do and still need manual oversight.
Read full analysis →
View on Bluesky →
We shouldn't expect to achieve "alignment" with #AI. There's an infinite set of community norms + values to align to, and the recent mathematics achievements promoted by OpenAI illustrate how misaligned the human engineers already are to the experts
Sovereign wealth fund for #AI is a good idea. @sanders.senate.gov proposes one here but Donald Trump has made a similar pitch. As have Sam Altman, Dario Amodei, and Elon Musk. Dividends from AI should return to the people. www.nytimes.com/2026/06/01/o...
Rafael M Batista reposted
Nathan Sanders and Bruce Schneier say we must distinguish AI's technological problems from its capitalism problems. Developers are solving the first, they argue, while leaving the second to market incentives. Structural reforms will be necessary to steer industry toward the pu…
AI Weekly's analysis
→
- Sanders and Schneier argue AI's most-cited harms are really capitalism harms, and the same model can help or harm depending on market incentives.
- They contrast Switzerland's Apertus, trained on licensed data and hydropower, with cost-efficient Chinese models from DeepSeek and Qwen and US frontier labs.
- Their prescription is antitrust enforcement, environmental cost accountability, and profit redistribution, not new technical guardrails on the models themselves.
Read full analysis →
View on Bluesky →
More on #OpenAI 's culture in @theatlantic.com this week www.theatlantic.com/technology/2...
Rafael M Batista reposted
“Like nuclear energy, AI must be at the service of all and of the common good. Decisions about technology must never be separated from conscience and responsibility.” www.ft.com/content/1231...
In the near future—as in now—I think the bigger #AI risk is not coming from the truly remarkable models but from the same-as-always humans 🙈 OpenAI has forever changed the world. But as a firm, in *this* society, it’s still got its training wheels on www.wired.com/story/a-flaw...
AI Weekly's analysis
→
- OpenAI's ChatGPT desktop app for macOS stored every user conversation on disk in plain text, outside Apple's app sandbox.
- Developer Pedro José Pereira Vieito disclosed the flaw and built a separate app that read the chat files with a single click.
- OpenAI shipped an update that encrypts stored conversations; the app is distributed via OpenAI's website, so Apple's sandbox rule does not apply.
Read full analysis →
Across three pre-registered studies (n=2691), people underestimate the rate of AI use and overestimate its benefits on *simple* tasks. Still people use AI despite its inefficiency. arxiv.org/abs/2605.22687
AI Weekly's analysis
→
- Across three pre-registered studies with 2,691 participants, people frequently chose AI for simple tasks even when it saved no meaningful time or effort.
- Users underestimated how often they turned to AI and overestimated how much time and effort it saved on cognitively simple work.
- Within a session, prior AI use led to further AI adoption and entrenched users' miscalibration about time savings, pointing to an overreliance feedback loop.
Read full analysis →
Rafael M Batista reposted
🚨New WP: Protecting users from AI persuasion🚨 🔸A 1-paragraph AI literacy treatment (explaining AIs can be told to pursue non-accuracy goals/to persuade) cuts AI dialogue political persuasion by ~half! 🔸No sig effect on general genAI trust arxiv.org/abs/2609.16432 w/ @rorchinik…
AI Weekly's analysis
→
- A brief pre-conversation warning cut belief change from persuasive LLMs by 48.1% (95% CI -59.5% to -36.8%) across two experiments.
- The intervention was tested on 3,208 Americans across GPT-4.1 (housing policy, N=1,992) and Grok 4.5 (15 political topics, N=1,216).
- Warned participants showed the reduced persuasion effect without a significant drop in overall trust in generative AI.
Read full analysis →
View on Bluesky →
thehackernews.com/2026/09/atta...
AI Weekly's analysis
→
- An attacker stole an API key from AI safety nonprofit METR and consumed roughly $600,000 in model credits over three weeks before detection.
- The entry point was a researcher's personal EC2 instance running a 'vibe-coded' agent dashboard with a fail-open authentication bug.
- A separate May incident saw automated agents probe METR's infrastructure; an independent researcher found a SQL data-exposure bug before attackers did.
Read full analysis →
Their own posts
Recent commentary
If I’m a reviewer on a paper that uses Pangram to claim something is or is not AI-generated, I’ll tell you right now that it’ll be a tough review.
I've been back on X, for about a month now. Much of my feed (partly by construction, partly by selection on who's remained active on X) is discussing #AI. My assessment 1 mo back: I think there is genuinely some foresight on X and it *is* where you need to be participate in the conversation.
I love em-dashes and I love footnotes. Both are such useful writing tools and they make for clearer writing. LLMs have ruined em-dashes for all of us. If they decide to get into the footnote game, we’re toast.
Recent trick I learned to improve my writing with #AI: First, I log out of Claude / ChatGPT / Gemini. Then I pick up a book, preferably one written before LLMs made it onto the scene.
Sensible reminder in @economist.com on #AI starting to ‘build itself’, what’s known as “recursive self-improvement” (RSI)
OpenAI and Anthropic will by some accounts have access to superintelligence within a matter of years if not months. And yet the two continue to release virtually the same suite of features. Does that not seem odd?
Curious about a research position at leading #AI lab, I uploaded cv to pre-fill the app. Every presentation I've given at a university was logged as a separate degree 😑 and this is at a frontier AI firm why do we judge AI capabilities by their ceiling and not their floor?
What in the world… you’re telling me METR is gifted hundreds of thousands of dollars worth of API credits by some unnamed AI company that they may or may not be “independently” evaluating?
I'm probably not the right audience for the benchmark papers because I really don't understand them. If I saw psychologists popping out scales out as quickly as LLM benchmarks, I'd be even more skeptical than I am of them today. What's the idea here?
Does #AI leave consumers more informed? Has anyone seen good evidence of this? Even suggestive evidence? I can imagine consumer-side experiments. Or suggestive evidence being something like since ChatGPT release there's more competition in certain consumer products.
Their network
In Rafael M Batista's orbit
Center = Rafael M Batista. Left = members they follow (green edges). Right = members who follow them (blue edges). Top = mutual follows (orange edges, slightly larger). Drag any node to reposition; click to open that profile.
Are you Rafael M Batista? Show it.
Add the Who’s Who of AI badge to your site or bio. It links back to this profile.
Markdown: [](https://aiweekly.co/whos-who/person/rafmbatista-bsky-social)