Brandon Roberts
LightNews LightNews
banner
bxroberts.org
Brandon Roberts
@bxroberts.org
8.8K followers 270 following 87 posts
Investigative journalist. News Apps Developer at @ProPublica.org. Occasional AI/ML researcher. Using data+code to hold power to account.
Web: https://bxroberts.org
Signal: brandonrobertz.01
Posts Replies Media Videos
bxroberts.org
I feel like "generate new ideas" is pretty debatable here 🤣
October 7, 2025 at 4:01 AM
bxroberts.org
I thought this was funny. "AI adoption is expected to yield significant benefits ... Realizing these benefits depends on continued AI innovation and public adoption."

Kind of a chicken and egg logic here. As if AI isn't good enough for many tasks because we're not all using it yet? 🤔
AI adoption is expected to yield significant benefits, including greater efficiency, improved safety and reliability, expanded access to services, and enhanced economic competitiveness. Realizing these benefits depends on continued AI innovation and public adoption.
September 25, 2025 at 10:47 PM
bxroberts.org
Office of Science and Technology Policy is requesting federal rules/guidelines/etc "that unnecessarily hinder" the development, deployment, and adoption of AI.

The list of "not appropriate" assumptions is interesting because AI experts will tell you many of these are critical to responsible AI use.
Quote from the Federal Register document "These assumptions include, but are not limited to: • Decision-Making and Explainability – Decisions are made, documented, and explained, in ways where the processes and rationale are traceable to a human actor. • Liability and Accountability – Allocation of legal responsibility and remedial frameworks rests with human actors or clearly identifiable organizational decision points. • Human Oversight and Intervention – Prescriptive requirements for human oversight, review, intervention, or continuous supervision in operational processes. • Data Practices – Data collection, retention, provenance, sharing, and permitted uses cases that do not account for the scale, reuse, or training dynamics characteristic of AI systems. • Testing, Validation, and Certification – Approaches to testing, approval, and post-market oversight designed for static products or human-delivered services, rather than adaptive or continuously learning systems." Quote from the Federal Register document: This RFI seeks to identify Federal regulations that hinder AI development, deployment, or adoption, particularly due to rules established before current AI capabilities were anticipated. OSTP is especially interested in regulations that, while serving important purposes, contain requirements or assumptions incompatible with how AI systems function or could function. Respondents are encouraged to identify regulations across all sectors where the underlying assumptions, technical requirements, or compliance frameworks may create unnecessary barriers to beneficial AI applications, even if the core policy objectives remain valid
September 25, 2025 at 10:47 PM
bxroberts.org
FDA's Digital Health Advisory Committee are accepting comments in advance of their meeting around "Generative Artificial Intelligence-Enabled Digital Mental Health Medical Devices" aka AI therapists (cringe every time I write that)
Quote from the federal register There is an increasing demand for mental health services in the US and insufficient access to mental health care providers. These new devices may be one way to help address this gap in care for people, potentially improving outcomes and access.
September 11, 2025 at 11:31 PM
bxroberts.org
In case you needed more evidence you can't just blindly copy and run code written by AI:

I just asked a llama3.2 model how to sort by create date and it gave me a command which would have deleted every file from the directory (after sorting) 😅
Screenshot of a LLM session. LLM was asked "wow to sort a directory in linux by create date?" Response: "find /path/to/directory -type f -print0 | sort -t '-' -k 1,1n -r | xargs -0 rm" This would have deleted all the files in the dir
September 10, 2025 at 9:44 PM
bxroberts.org
For reference, a 2024 VoiceBrain CRADA with Dept of Homeland Security was focused around capturing, transcribing and doing AI analytics around airport radio traffic for TSA in Oakland.

www.prnewswire.com/news-release...
VoiceBrain's AI software captures millions of daily voice communications and converts them into actionable data in real-time. The VoiceBrain platform can be deployed in less than 15 minutes and integrates seamlessly with existing two-way radios, requiring no material modifications to hardware or infrastructure. As VoiceBrain analyzes transmissions, it extracts key content and delivers alerts and notifications about critical situations as they arise. This transforms radio communications into a valuable dataset that enhances decision-making, operational efficiency, safety, and security. Additionally, VoiceBrain's capability to analyze thousands of hours of data in seconds empowers customers to conduct queries, generate summaries, and prepare reports with unprecedented speed and accuracy, optimizing time and standardizing processes.
September 9, 2025 at 10:59 PM
bxroberts.org
US Coast Guard partnering with AI voice transcription + analytics company VoiceBrain in a Cooperative Research and Development Agreement (CRADA). "Eliminating human error and fatigue" but nothing about the new machine errors it will introduce.

public-inspection.federalregister.gov/2025-17361.pdf
Additionally, VoiceBrain's integration into existing and future multicast communication systems provides an opportunity to streamline and enhance the process of maritime response by eliminating human error and fatigue.
September 9, 2025 at 10:59 PM
bxroberts.org
The AI fake journalist saga continues ... Business Insider alone has retracted 40 articles now wapo.st/4p9E7hz

Editors who've published a pitch from someone they never talked to nor saw and paid via PayPal have been put on notice 😅
Quote from the story: Nine essays with the byline Tim Stevensen also contain contradictions. In one article, he claims his daughters are in their twenties and his son is a teenager; in another one published this year he says they’re 11, 13 and 15. Quote from story: Wired only discovered that piece was fraudulent because the person who submitted the work was unwilling to provide information sufficient for its payment system, insisting to be paid through PayPal or check. Wired wrote that it has taken steps to ensure this doesn’t happen again, without specifying what they were.
September 6, 2025 at 11:38 PM
bxroberts.org
Proposed Dept of Education rule change around "Advancing Artificial Intelligence in Education". Is this backed by anything other than vibes?
www.federalregister.gov/public-inspe...
From a Department of Education proposed rule change: "AI tools have the potential to support personalized instruction, increase classroom engagement, and improve student outcomes."
July 18, 2025 at 10:14 PM
bxroberts.org
We have several iterations of the script. In what appears to be an early one they used gpt-4-turbo (superseded by OpenAI GPT-4o) and, in a more recent script, o3-mini was used with the default "reasoning" setting and a timeout of 30s.
A screenshot of AI analysis model instantiation using gpt-4-turbo A screenshot of AI analysis model instantiation using o3-mini
June 6, 2025 at 8:15 PM
bxroberts.org
10/ When we reached out to the VA press secretary, he told us that contract decisions go through multiple reviews by employees, and also the VA would not cancel crucial contracts without a contingency plan:
A quote from the ProPublica article reads: The VA is reviewing all of its 76,000 contracts to ensure each of them benefits veterans and is a good use of taxpayer money, he said. Decisions to cancel or reduce the size of contracts are made after multiple reviews by VA employees, including agency contracting experts and senior staff, he wrote. Kasperowicz said that the VA will not cancel contracts for work that provides services to veterans or that the agency cannot do itself without a contingency plan in place. He added that contracts that are “wasteful, duplicative or involve services VA has the ability to perform itself” will typically be terminated.
June 6, 2025 at 2:36 PM
bxroberts.org
8/ The VA is also standing behind its use of AI to identify “munchable” contracts, calling it “a commonsense precedent.”

Documents released to ProPublica indicate they may even be looking at using it for other VA operations, such as processing disability claims.
A screenshot from the ProPublica article reads – The VA is standing behind its use of AI to examine contracts, calling it “a commonsense precedent.” And documents obtained by ProPublica suggest the VA is looking at additional ways AI can be deployed. A March email from a top VA official to DOGE stated: Today, VA receives over 2 million disability claims per year, and the average time for a decision is 130 days. We believe that key technical improvements (including AI and other automation), combined with Veteran-first process/culture changes pushed from our Secretary’s office could dramatically improve this. A small existing pilot in this space has resulted in 3% of recent claims being processed in less than 30 days. Our mission is to figure out how to grow from 3% to 30% and then upwards such that only the most complex claims take more than a few days.
June 6, 2025 at 2:36 PM
bxroberts.org
4/ Even Sahil Lavingia, the programmer, agreed there were problems. This is what he told us:
A screenshot from the ProPublica article reads: “I think that mistakes were made,” said Lavingia, who worked at DOGE for nearly two months. “I’m sure mistakes were made. Mistakes are always made. I would never recommend someone run my code and do what it says. It’s like that ‘Office’ episode where Steve Carell drives into the lake because Google Maps says drive into the lake. Do not drive into the lake.”
June 6, 2025 at 2:36 PM
bxroberts.org
3/ Second, the DOGE AI tool’s underlying instructions were deeply flawed.

The system was programmed to make intricate judgments based on the first few pages of each VA contract — about 2,500 words — which contain only sparse summary information.
A section of the DOGE artificial intelligence prompt with the text: “Then, evaluate if this contract is ‘munchable’ based on these criteria: - If this is a contract modification, mark it as N/A for munchable status - If this is an IDIQ contract: * For medical devices/equipment: NOT MUNCHABLE * For recruiting/staffing: MUNCHABLE * For other services: Consider termination if not core medical/benefits - Level 0: Direct patient care (e.g., bedside nurse) - NOT MUNCHABLE - Level 1: Necessary consultants that can't be insourced - NOT MUNCHABLE - Level 2+: Multiple layers removed from veterans care - MUNCHABLE - Services that could easily be replaced by in-house W2 employees - MUNCHABLE”
June 6, 2025 at 2:36 PM
bxroberts.org
2/ First, the DOGE AI tool produced glaring mistakes.

It often hallucinated the size of contracts, inflating their value. It concluded more than a thousand were each worth $34 million, when in fact some were for as little as $35,000.
A table of rows from our analysis of the DOGE AI produced list of contracts to cancel. The table shows the “AI total value” column, all valued at “$34 Million” and an “actual potential value” column with values ranging from $35k to $48k.
June 6, 2025 at 2:36 PM
bxroberts.org
📊 JUST UPDATED: 527 Explorer database now includes 140k new records following the IRS Q1 2025 filing deadline for 527 organizations.

Found something interesting? Reach out!

projects.propublica.org/527-explorer/
New Records in 527 Explorer Contributions: 87,706 Expenditures: 46,864 Initial Filings: 1,178 Periodic Reports: 634 Directors/Officers: 2,859 EAIN Numbers: 308 Related Entities: 475 New Records Total: 140,024 Most recent record: 2025-03-31
April 30, 2025 at 6:58 PM
bxroberts.org
yeah. oddly specific right? the site does have this testimonial on it
We have distributed $10 Amazon gift cards for participating in a survey. The incentives are necessary to appreciate the effort of participants in devoting their time to express their views. Aniket Agrawal Executive, Grail Insights
April 8, 2025 at 8:53 PM
bxroberts.org
Just in: The Securities and Exchange Commission is going to hold a public roundtable about using AI in the financial industry. Thursday, March 27, 2025, at 9a.m. EDT. The public is invited to submit feedback and it will be streamed online. www.federalregister.gov/d/2025-05033
SECURITIES AND EXCHANGE COMMISSION Sunshine Act Meetings TIME AND DATE: Notice is hereby given, pursuant to the provisions of the Government in the Sunshine Act, Pub. L. 94-409, that the Securities and Exchange Commission staff will hold a public roundtable on Thursday, March 27, 2025, at 9 a.m. PLACE: The roundtable will be held in the Auditorium at the Commission’s headquarters, 100 F Street, NE, Washington, DC. STATUS: The meeting will begin at 9 a.m. and will be open to the public. Seating will be on a first-come, first-served basis. Doors will open at 8 a.m. Visitors will be subject to security checks. The meeting will be webcast on the Commission’s website at www.sec.gov. MATTERS TO BE CONSIDERED: The Commission staff will host a roundtable on artificial intelligence in the financial industry. The roundtable is open to the public and the public is invited to submit written or electronic feedback. This Sunshine Act notice is being issued because a majority of the Commission may attend the roundtable. The roundtable will provide an opportunity for SEC staff to hear the views of market participants, academics, and industry professionals on the risks, benefits, and governance of artificial intelligence in the financial industry. CONTACT PERSON FOR MORE INFORMATION: For further information, please contact Vanessa A. Countryman from the Office of the Secretary at (202) 551-5400. Authority: 5 U.S.C. 552b. Dated: March 20, 2025
March 20, 2025 at 8:52 PM
bxroberts.org
I'm sorry but mass-translating press releases into articles using AI isn't journalism
As an AI-Assisted Reporter, you will have the opportunity to develop your news and technical skills, including learning how to manage and utilize AI technology effectively.This is highly production-oriented work – you’ll be assisted by technology to create a high volume of stories from trusted press releases and similar sources. This is not a beat-reporting position and does not require travel or face-to-face interviews; those are the activities you’ll enable other journalists to do by taking as much as possible off their plate.
February 20, 2025 at 7:48 PM
bxroberts.org
New data has been released on ProPublica's 527 Explorer!

Quarter 4 (Oct-Dec) 2024 filings are in, which means over half a million new records are available.

Browse them yourself here
⬇️⬇️⬇️
projects.propublica.org/527-explorer/

If you find this useful, we'd love to hear from you as always.
Report Type | Records Contributions | 314,482 Expenditures | 297,620 Initial Filings | 711 Periodic Reports | 1,031 Directors/Officers | 1,607 EAIN Numbers | 192 Related Entities | 158 New Records Total | 615,801
January 22, 2025 at 6:58 PM
bxroberts.org
They know they're in the journalism industry right?
www.nytimes.com/2025/01/16/b...
The slide deck that Ms. Watford presented describes artificial intelligence as a key enabler of The Post’s success, the people said. It describes The Post as “an A.I.-fueled platform for news” that delivers “vital news, ideas and insights for all Americans where, how and when they want it.” It also lays out three pillars of The Post’s overall plan: “great journalism,” “happy customers” and “make money.” The Post lost roughly $77 million in 2023.
January 16, 2025 at 8:14 PM
bxroberts.org
🤣
December 12, 2024 at 12:22 AM
bxroberts.org
I appreciate the honesty here lol. Noting that human judgements are "subjective" and don't line up with the metric proposed in the paper. 🤔
December 11, 2024 at 11:36 PM
bxroberts.org
Post election reports are in on ProPublica's 527 explorer. We just released ~350k new contributions and expenditures since our last update on Oct 19. As always, we'd love to hear back from anyone who is working with these kinds of records.
projects.propublica.org/527-explorer
A table showing record type and number of new records. Expenditures: 188368 Contributions: 153646 Directors/officers: 449 Periodic reports: 346 Initial filings: 214 EAIN numbers: 76 Related entities: 40
November 22, 2024 at 6:11 PM
bxroberts.org
WE MUST INVESTIGATE
Definitely not human --Facebook --Inside Edition WE MUST INVESTIGATE
July 28, 2023 at 6:59 AM