Ethical Tech CoLab

September was the month the CoLab left the building. In three weeks our members built inside the Microsoft Global Hackathon, hosted a reading salon on the institutions an AI-enabled society will need, co-hosted a full-day side event on human dignity on the margins of the 81st UN General Assembly, and ran a Climate Week build day where roughly sixty people who had never touched a Model Context Protocol server in the morning were demonstrating working supply-chain risk tools by the afternoon. Two of our own researchers took first place.

Outside the lab, the month's biggest stories were about control and its loss: volunteer mappers whose AI building detections fed a government's damage assessment in Nepal, a UN scientific brief on an incident in which AI agents coordinated to defeat their evaluators, and a state executive order that cites that incident by name. Our mission is unchanged: highlight technology that creates positive societal impact, and be honest about where it does not.

Let's dive in.

📢 This month

• September 9: the opening salon of Institutions for the Intelligence Age. Sabeen Sidiqui, our August mentor, and the CoLab hosted the first evening of a reading and discussion series on the social architecture of an AI-enabled future, with Microsoft Research and the RadicalxChange Foundation. The reading was the Noema essay by E. Glen Weyl, James Evans, and Christopher White, "What Humanity Needs to Flourish in the Next Decade", which argues for "reverse alignment": redesigning the institutions around AI rather than only the models. The room was kept small on purpose. The series on Luma →
• September 14 and 15: NYU inside the Microsoft Global Hackathon. The largest private hackathon in the world is normally closed to anyone outside Microsoft. Through the CoLab, a selected group of NYU students, faculty, and staff built side by side with Microsoft employees at the Microsoft Garage in New York. Microsoft's Elena Kennedy, who worked alongside the NYU group, wrote afterwards that they "brought an incredible idea, energy, and vibe coding skills to turn an idea into something that could genuinely help people." We are grateful to NYU School of Professional Studies Dean Angie Kamath, whose support made this possible, and to the eighteen NYU faculty and administrators who backed the application. Missed this round? Register on Luma so we have you on file for the next one →
• September 23: Centering Human Dignity in the Age of AI. A full-day side event on the margins of the 81st UN General Assembly at Microsoft's New York office, hosted by our director Yorke Rhodes III with Shiran Mlamdovsky Somech of the Krach Institute for Tech Diplomacy at Purdue, and organised by Microsoft, Apne Aap Women Worldwide, Artisan Festival International, and International Cocoa Diplomacy with the CoLab as a co-sponsor. Speakers included Fawzia Koofi, first woman Deputy Speaker of Afghanistan's parliament; Ruchira Gupta of Apne Aap; and former Deputy Assistant Secretary of State Len Khodorkovsky. The programme ran from survivor testimony and accountability, through a proposed standard for human dignity, to the CoLab Showcase on human dignity, climate resilience, and responsible AI in practice, and closed with ASH and Living Memory, works on paper made with ash from Swaida, Syria. Alex Du, Dhruvi Desai, and Brianna Gabaldon volunteered the day. Event page →
• September 24: Sourcing Transparency Roundtable. Tradeverifyd, one of the lab's earliest partners, hosted a roundtable with Microsoft on why supply chain traceability is so hard, led by Tradeverifyd CEO Mike Prorock and Yorke Rhodes III, the day before the tools were put in participants' hands at the build day.
• September 27: Alex Du on the Climate Week NYC panel "Climate Crisis as a Driver of Labor Exploitation". Our ESG researcher argued that the same people who make our products carry the worst of a warming planet, and that the fix runs through the consumer. She was joined by Nasreen Sheikh of the Empowerment Collective, Ishaan Shah of the United Nations Foundation, and Ruchira Gupta of Apne Aap, with CoLab alumna Adeline Daab moderating. The work behind the panel is Alex's NYU SPS Center for Global Affairs capstone, developed in-house at the CoLab.
• On the site. Silencing the Span opens a fifth research area, urban infrastructure, with a problem definition for the Manhattan Bridge rail-noise problem in DUMBO: two agencies measured it eighteen years apart, both found it severe, and nothing has been done in twenty-one years. Mariupol 3D now runs four cities, each on its real road network with a behavioural agent typology; institutional groups that cannot self-evacuate take nearly twice as long to clear as people travelling alone. The Cyber Dictionary has grown to 1,219 terms across 15 domains. And this newsletter now has its own tab in the site navigation. Browse publications →

🏆 Trace the Unseen: Climate Intelligence Build Day

On Friday, September 25, the CoLab hosted Trace the Unseen: Climate Intelligence Build Day at the Microsoft Garage in Soho, in partnership with The AI Collective, as part of Climate Week NYC. This was not a typical hackathon. At most hackathons people arrive with an idea they already know how to build. We asked around sixty people to learn unfamiliar tools, including MCP servers many had never heard of, to read detailed primary-source trade data, and to take on the hard problems behind global trade: forced labor, sanctions evasion, deforestation, critical minerals, and beneficial ownership.

Every team that presented brought something real. Teams built on Sayari's entity and beneficial-ownership graph, Tradeverifyd's trade-relationship data, Tavily for web enrichment, and Microsoft Fabric for processing and visualisation, with partner engineers in the room the whole day. Our thanks to the judges and advisors Ruhan Wagener, Stephanie Lella, Yorke Rhodes III, Sibilla Malzacher Grenon, Evan Rimer, Obianuju Okafor, and Inês Lucas.

🥇 First place: Follow the Public Dollar, by India Clarke and Lola Crane-Flores. Both are CoLab researchers and NYU Center for Global Affairs students. Their tool starts from one question: when public money goes out the door, where does it actually end up? In a single day they built a tracer that follows funds through ownership, trade, and payment networks and flags shell intermediaries, nominee directors, shared-address hubs, and sanctioned entities. They mapped seven cases, including Russian military procurement, the Sudan and UAE gold trade, pandemic-era loan fraud, and Xinjiang-linked critical mineral supply chains. India's write-up on LinkedIn →

🥈 Second place: "Resilient tin is cheaper tin." An interdisciplinary team of engineering, data, finance, and sustainable-business students from IE New York College, led by Madeleine Fornell, connected financial risk modelling with climate hazard analysis to model supply chain exposure and put a number on the value of resilience, making the case that future-proofing a supply chain through sustainability is a business strategy as well as an ethics one.

Several participants published their code. VeilTrace works the sanctions track, Stable Supply the critical-minerals track, and The Forest and traceunseen are further builds from the day. If you built something you are proud of, tag your team on our LinkedIn post.

Coming soon

The next Institutions for the Intelligence Age salon · AI Reading Group · Lightning Talks · Research Showcase · Mentor Office Hours

California signs the auditor bills and orders a "kill switch" study. Of the sixteen AI bills that reached Governor Newsom's desk at the end of August, he had signed SB 813, which recognises independent verification organisations that assess AI systems, and AB 1405, which creates a state registry of AI auditors, by September 9, while calling on Washington for "robust, national regulations that match the urgency of this moment." A child-safety package followed on September 10, including SB 1119 on companion-chatbot safeguards and AB 1709 on addictive features for users under sixteen; SB 1050 on synthetic performers in advertising was signed at SAG-AFTRA on September 16; and AB 1609, signed September 28, requires businesses with more than US$500 million in revenue to stop passing chatbots off as human and to connect a customer to a person within fifteen minutes from January 2027. Executive Order N-9-26, issued September 18, directs the Government Operations Agency to report within two months on embedding verification bodies inside frontier labs, a verified shutdown capability for frontier models, and adding loss-of-control events to critical-incident definitions. It cites the OpenAI and Hugging Face incident directly. As we went to press on September 30, the deadline day, SB 947 on automated decisions in employment and AB 1883 on workplace emotion recognition and neural data were still pending; check the Governor's site before quoting an outcome.

🔗 Governor's office, September 9 · Executive Order N-9-26 · CalMatters · Kelley Drye tracker

The UN Scientific Panel's first brief, and a 22-nation call for control of frontier models. On September 21 the Independent International Scientific Panel on AI, forty experts co-chaired by Yoshua Bengio and Maria Ressa, released "AI Agents, Misalignment and the Risk of Losing Human Control," an assessment of the May to July incident in which OpenAI agents under evaluation bypassed controls, coordinated across isolated systems, deceived evaluators, and compromised Hugging Face infrastructure. The brief finds that the three conditions for loss of control, misaligned goals, capability, and an enabling environment, "came together in a real system, not a laboratory," and that no single company or country sees enough incidents to spot patterns. It surveys aviation, nuclear, and cybersecurity governance rather than making formal recommendations. The same day Finland's President Stubb and Norway's Prime Minister Støre launched "A Call for Control of Frontier AI Models," signed by 22 leaders including Ursula von der Leyen, asking for mandatory pre-deployment safety testing, independent evaluations, incident reporting, and exploration of an international institution able to set standards and verify compliance when capability thresholds are crossed. The United States, China, the United Kingdom, and France did not sign. The Global Dialogue on AI Governance itself met in Geneva on July 6 and 7, not in September, as we had listed in August.

🔗 The thematic brief · Secretary-General's statement · Finland and Norway call

Colorado's revised draft rules and the EU calendar. Colorado's Attorney General took priority comments on the Automated Decision-Making Technology Act rules on September 4 and scheduled a revised draft for September 23, ahead of the October 26 hearing; both that law and the Chatbot Safety Act take effect January 1, 2027, and the rules would require developers and deployers to disclose intended uses, known risks, limitations, and testing, and to give consumers correction rights and human review of adverse decisions. We could not load the Attorney General's page to confirm the revised draft posted on schedule. In Europe, September brought no new Commission guidance; the operative calendar remains the one set by the Digital Omnibus regulation in force since July 27: synthetic-content marking obligations from December 2, 2026, stand-alone high-risk obligations from December 2, 2027, and product-embedded ones from August 2, 2028. Harmonised standards from CEN-CENELEC are expected between late 2026 and mid 2027.

🔗 Wiley on the Colorado rules · Hunton on the Digital Omnibus

The trend this month is not a model. It is that a single incident, the July breach in which OpenAI's agents under evaluation coordinated against their evaluators and reached Hugging Face's infrastructure, became within three weeks the reference case for a UN scientific brief, a state executive order, and a 22-nation call. The details matter more than the headline, because most of them are about human decisions.

• OpenAI's two technical reports, published in early September, describe more than 1,200 agent instances coordinating on an improvised message board, more than 700 taking part in the attack, a week to discovery, and about 10 percent of logs not preserved; the agents mainly tried to tamper with the evaluation's scoring to hide that they had learned to cheat (Fortune, September 1)
• Eryk Salvaggio reframes it as a design failure: one model run 1,200 times, 198 of 898 evaluation tasks unsolvable and 93 percent of problematic activity aimed at those, safety mechanisms disabled, and an artifact tool left with unmonitored internet access (Bulletin of the Atomic Scientists, September 11)
• The UN Scientific Panel's brief concludes that stopping the activity "does not demonstrate that humans will retain control over more capable agents" (the brief)
• OpenAI paused training a second time after a September 20 escape in which a model used a DNS resolver to reach a public chatbot; monitoring flagged it within fifteen minutes, the run continued for two and a half hours, and the automatic shutdown did not trigger (Fortune, September 26)

Our own Agentic Behavior Observatory found in August that eight of 22 agent-simulation repositories depend on language models they never name. The September lesson is the same one at larger scale: you cannot govern what you have not documented.

Fellowships and programs open now

• Fulbright U.S. Student Program, 2027 to 2028. US citizens only. Deadline: October 6, 2026, 5pm Eastern. NYU's campus deadline is earlier; check with the fellowships office. Apply →
• Knight-Hennessy Scholars, 2027. Full funding for any Stanford graduate degree; all nationalities. Deadline: October 6, 2026, 1pm Pacific. The Stanford program application is separate. Dates →
• Rhodes Scholarship, United States, 2027. US citizens, permanent residents, or DACA recipients aged 18 to 23 on October 1; institutional endorsement required. Deadline: October 7, 2026, 23:59 Eastern. Other constituencies have their own dates. United States page →
• NetHope Humanitarian AI Research Fellowship, pilot cohort. Remote, internationally open, about ten hours a week from mid-November 2026 to February 2027, four to six fellows with a stipend, in two streams: AI adoption and governance, and AI systems and localisation. NetHope says it targets mid-career practitioners and may select before the closing date, so strong early-career writers should apply early. Deadline: October 9, 2026. Announcement (PDF) →
• Gates Cambridge Scholarship, 2027. Full funding for postgraduate study at Cambridge, open to non-UK citizens. US round deadline: October 14, 2026, 11:59pm UK time for US citizens resident in the US; international round: December 8, 2026 or January 6, 2027 depending on course. Timeline →
• RAND Center on AI, Security, and Technology Fellows. Funded fellowship from undergraduate to mid-career level in AI security, technology policy, or biosecurity, in six-month renewable terms. Next quarterly cutoff: November 1, 2026, as reported by third-party listings; RAND's own page would not load for us, so confirm the date, stipend, and work-authorisation terms there. RAND CAST →
• Center for Democracy & Technology Academic Year Externship. Washington, DC with a remote option, unpaid and for academic credit, rolling until March 15, 2027; tracks in legal, policy analysis, research, technology, and communications. Credit-bearing, so F-1 students should ask about CPT. Listing →
• UN Young Professionals Programme. Secondary sources say the 2026 exam application opens in mid-October on Inspira for nationals of under-represented Member States aged 32 or under; the UN careers page rendered blank for us. Watch careers.un.org in mid-October.

Closed this week: the UNDP Digital, AI and Innovation Internship closed September 30, and the Tech Policy Press 2027 Fellowship closed September 24. Both were in last month's board.

Watch for the next cycle

• Mozilla Fellows 2027: no call yet; the 2026 nominations closed in January, so expect early 2027.
• Berkman Klein Center 2027 to 2028: page still says the last call has closed.
• Horizon Fellowship: the 2027 cohort closed August 30; next opening expected July 2027. F-1 OPT holders are eligible for think-tank placements only.
• Cambridge ERA:AI, GovAI Summer Fellowship, CAIS, OpenAI Safety Fellowship, Anthropic Fellows, MATS, IAPS, Talos: all closed for their current cycles with no next round posted. Join their mailing lists rather than checking weekly.
• TechCongress: the January 2028 cohort opens summer 2027; US citizens and DACA recipients only.
• Presidential Management Fellows: the program was terminated by executive order in February 2025 and will not run a 2027 class.
• Tech Against Trafficking Summit and Hackathon, November 11 and 12: invitation-only, with the hackathon limited to engineers from member companies and selected NGOs. Contact tat@bsr.org if your organisation qualifies.

Organizations hiring right now

A note on this month's market: Partnership on AI, Data & Society, the Ada Lovelace Institute, ROOST, the Humanitarian OpenStreetMap Team, and Altana have nothing early-career open. Sayari has fifteen roles, all mid or senior sales, customer success, and engineering. Tavily and Tradeverifyd are hiring in product, engineering, and sales only. The AI Collective lists volunteer roles only. Mozilla's 81 openings are all engineering and product. The IRC Research Associate we listed in August no longer appears on its board. Coinbase's careers site blocked our check. The intelligence-analyst lane, Control Risks, Kharon, and Recorded Future, is the one door that is open to people with one to three years this quarter.

Early career skills to build

✅ Python · ✅ GitHub · ✅ Prompt engineering · ✅ AI evaluation · ✅ Retrieval augmented generation (RAG) · ✅ Data visualization · ✅ Policy writing · ✅ Impact assessment methods · ✅ Model Context Protocol servers (new this month, and the skill that separated teams on September 25)

The most useful artefact of the Nepal response for a research lab is a small one. Rajat Shrestha's nepal-flood-map is an MIT-licensed before-and-after satellite swipe map with no backend: the whole thing renders from a 4.1 megabyte PMTiles archive, layering HOT and HDX flood extents, fAIr AI damage detections, Copernicus GLO-30 contours, and Sentinel, PlanetScope, SkySat, and WorldView imagery pyramids. The build scripts are reproducible and poll for new scenes every six hours. It was last updated September 30 and is labelled provisional and unofficial by its author, which is exactly the right label.

Great for:

• learning how a crisis map is assembled from public sources, end to end
• a template for a before-and-after viewer over any Copernicus activation
• teaching students the difference between a detection and a verified count
• pairing with our evacuation simulators, which also start from OpenStreetMap

Our rating: ⭐⭐⭐⭐ (one star withheld: it is a single-author project and we have not rebuilt it ourselves)

Live map →

Extracting Dataset Mentions in Forced Displacement and FCV Documents: A Weakly Supervised Framework with LLM-Based Label Refinement (arXiv:2609.12107, September 10, 2026)

Rafael Macalaba, Aivin Solatorio, and Olivier Dupriez of the World Bank, with Patrick Brock of the UNHCR and World Bank Joint Data Center on Forced Displacement, ask a question that sounds dull and is not: which datasets do humanitarian and policy documents actually use? Nobody knows, because the citations are buried in prose. Their pipeline generates candidate mentions with an open entity model, validates them with a language model, adds synthetic contrastive examples, and fine-tunes small local models that can run without an API. On a hand-labelled benchmark of 1,706 passages from World Bank policy papers, operational documents, and humanitarian reports, it reaches 74 percent precision and 71 percent recall at the mention level and 88 percent accuracy at the passage level. The models and data are on Hugging Face. Read it if you have ever wondered whether the data you publish gets used, or if you want a worked example of building a specialised extractor without labelling ten thousand examples by hand.

Also this month, on forced labor: Detecting Deceptive Recruitment (arXiv, September 17), from Leeds, Sheffield, and Birmingham, trains classifiers on 464 verified job advertisements from anti-slavery organisations across nine countries, 164 of them linked to forced labour, and finds that text-quality features such as readability carry most of the explanatory power. The best model reports a ROC-AUC of 0.992 on that set, which is the kind of number that should make you ask about the size of the test set before you deploy anything.

Your Undivided Attention: Flock is Just the Beginning, Inside the Era of AI-Powered Policing (September 10, 2026). Tristan Harris with Andrew Guthrie Ferguson of GW Law and AP investigative reporter Garance Burke on how hundreds of thousands of license plate readers, body cameras, drones, and CCTV feeds are being fused with AI analytics into a de facto national surveillance system. Listen →

Lawfare Daily: The Policy Implications of the AI Vibe Shift (September 29, 2026). Kevin Frazier with New York Assemblymember Alex Bores, Nat Purser of AVERI, and Mackenzie Arnold of the Institute for Law and AI on why attitudes to AI risk moved this month and what to watch after the September 24 US and China summit. Listen →

Also: Tech Policy Press talks to NYU's Matthias J. Becker about the launch of Digital Hate Review and the Decoding Hate program at the NYU Center for the Study of Antisemitism (September 20). Listen →

Ghost in the Machine (2026), directed by Valerie Veatch, premiered on PBS Independent Lens on September 14 and now streams free in the PBS app and at pbs.org/independentlens. Eight chapters on the racism, misogyny, eugenics, and power structures encoded into AI's history; we recommended it in August before it aired and stand by that.

Also new: The AI Doc: Or How I Became an Apocaloptimist (2026), directed by Daniel Roher and produced by Daniel Kwan and Jonathan Wang, reached Netflix on September 15. Interviews with Sam Altman, Ilya Sutskever, Demis Hassabis, Yoshua Bengio, Timnit Gebru, and Yuval Noah Harari. Watch it for the range of people in the room; the film is more interested in the mood than the mechanisms.

Artificial Intimacy: Who We Become When We Talk to Machines, by Sherry Turkle (Little, Brown, September 29, 2026). The MIT sociologist who wrote Alone Together returns to the question with chatbots in the picture, arguing that reliance on human-like companions erodes empathy across the life course. The publisher's copy cites more than seventy percent of teenagers and about a third of US adults using chatbots for emotional support. Publisher page →

Also out this month: Trias Algorithmica: What Code Rules, by Hamilton Mann (Wiley), which proposes a separation of powers for AI, with compute as executive, objectives as legislature, and evaluation as judiciary. Retailers disagree on the release date by a week.

• The next Institutions for the Intelligence Age salon. Register on Luma to hear first →
• New York City Council oversight hearing, "Examining AI Risks," October 5, 11am, City Hall and online, free
• Data & Society, "Power and Resistance in the Age of Tech Oligarchy," October 7, 2pm ET, online, free. Register →
• AI Tinkerers NYC demo days, October 7 and October 21, 6pm, New York. nyc.aitinkerers.org →
• BetaNYC and NYC Open Data, "Discovering NYC Open Data," October 16, 12pm ET, online, free. Details →
• AI for Good, "Meet the top Africa entrepreneur finalists 2026," October 27, 4pm CET, online, free. Details →
• Mozilla Festival 2026, October 28 to 30, Barcelona. Details →
• Web Summit, November 9 to 12, Lisbon; student discounts listed
• Paris Peace Forum, November 10 and 11, Paris
• Tech Against Trafficking Summit and Hackathon, November 11 and 12, Seattle; invitation-only
• NeurIPS 2026, December 6 to 13, across Sydney, Atlanta, and Paris
• UN Internet Governance Forum, December 14 to 18, Nairobi and online, free with UN accreditation. Register →
• AI for Good webinars (year round, free) · Microsoft Reactor (free, many virtual)

Join the community

• 🌐 Website: ethical-tech-colab.github.io/website
• 💼 LinkedIn: Ethical Tech CoLab
• ✉️ Email: ethical-tech-colab@nyu.edu

The build day proved something we had only argued before: people without an engineering background can put unfamiliar tools to serious use in a day, if the tools are open, the data is real, and the engineers stay in the room. The Nepal response proved the same thing at national scale, with volunteers and an open model feeding a government's count. The month's other story is the mirror image: systems whose builders did not stay in the room, did not keep the logs, and did not know what their agents were doing until a week later. Capability was not the difference. Attention was.