Testimony of David A. Bader, Ph.D., New Jersey General Assembly Science, Innovation and Technology Committee

Testimony of David A. Bader, Ph.D.
Distinguished Professor of Data Science
Director, Institute for Data Science
Associate Dean for Research, Ying Wu College of Computing
New Jersey Institute of Technology

Before the New Jersey General Assembly Science, Innovation and Technology Committee on October 19, 2026

Artificial Intelligence and Society

1. Opening

Chairman Tully, Vice Chairman Singh, and members of the committee, thank you for inviting me to participate on behalf of the New Jersey Institute of Technology.

I am David Bader, Distinguished Professor of Data Science at NJIT in Newark, where I direct the Institute for Data Science and serve as Associate Dean for Research in the Ying Wu College of Computing.

You will hear today from researchers who study what AI models can do. My vantage point is different. I have spent nearly thirty years building the machines underneath AI: the processors, networks, power, and software that make large-scale computing work. In 1998 I built the first Linux supercomputer from off-the-shelf parts, and that architecture now sits under essentially every cloud region and AI data center running today. I also served on Governor Murphy’s AI Task Force with three NJIT colleagues, and in the interest of full disclosure, I co-founded Adapter, an AI infrastructure startup.

You asked about developments in AI, its use across society and the economy, and the concerns it raises. Our view is that AI is already here, it runs on physical infrastructure this state can shape, and most of the harms worth worrying about are specific enough to fix. We do take the near-term risks seriously. What I don’t accept is that the answer is to stop.

2. Current State

That history shapes how I see AI today. In 1998, the hard part of supercomputing was getting a machine at all. Today, the hard part is plugging it in. AI is physical. Every answer from a chatbot draws power from the same grid that serves your constituents, and real-time AI strains that grid more than steady batch work. The binding constraint on frontier AI right now is electricity, transformers, substations, interconnect queues, and trained staff. We are already power limited.

That makes the buildout a long-term bet. We estimate a GPU trains AI models for about two years, then answers queries for roughly five more. And the substations and transmission lines we build to feed it will be here for decades. So, the better answer is often smarter software running on the hardware we already have.

AI has also changed from something that answers questions into something that acts. These systems write and run code and carry out multistep tasks with access to email, files, and accounts. I use one. Every Monday it scans grants.gov and drafts funding alerts for NJIT’s 66 computing faculty, and nothing goes out until I review it. With these agents, the risk is often the connection, not the model. And people are adopting AI faster than we are building ways to check it. Generative AI reached 53 percent of the population within three years, faster than the PC or the internet, while documented AI incidents rose from 233 in 2024 to a record 362 last year. Nearly every leading developer reports how well its models perform, but the reporting on safety is spotty.

3. AI in New Jersey and Beyond

New Jersey is out in front. A Rutgers survey last year found that 74 percent of residents had used an AI tool, versus 68 percent nationally. Forty-one percent of working New Jerseyans use it on the job every week. Forty-five percent of students use AI often for schoolwork. State government uses it too, with more than 7,000 state workers on the NJ AI Assistant by late 2024. And 57 percent of residents expect AI to eliminate more jobs than it creates.

Access to AI changes what people are capable of. Two New Jersey high school students recently worked with my group at NJIT on Princeton’s FlyWire challenges, mapping the wiring of a fruit fly’s brain, and helped us take first place in one challenge and second in another. The AI turned what would typically be weeks of literature review into minutes. High schoolers competed at the frontier of neuroscience and won.

Other governments around the world are learning how to deal with potential costs of AI as well. The European Union, usually the strictest regulator, has delayed its rules for high-risk uses such as hiring until December 2027, while its requirement that AI-generated content be marked takes effect December 2 of this year. The EU’s emphasis is shifting to specific, provable harms.[MB1.1]

4. Concerns

It is the specific, provable harms that concern us most. Deepfake voice and video are being used for fraud and election manipulation. Spear phishing runs at industrial scale, automated tools probe the networks of utilities, hospitals, and schools, and model-written code goes into production without review. Researchers have documented hallucination and bias in systems deciding loans, hiring, and medical triage. Data pipelines go unaudited, and brittle automation is quietly trusted in power, water, and transportation.

New Jersey is already responding. Since April 2025, using a deceptive deepfake to commit a crime has been a third-degree offense, and 79 percent of residents polled by Rutgers say creating AI images of real people without permission should be illegal.

These are engineering, procurement, and liability problems. They are solvable, and they are not our extinction. My test for a real risk is whether you can name the mechanism, the failure path, and the mitigation. An aerospace engineer can tell you exactly how a turbine fails and what the inspection interval is. Most extinction talk fails that test; it is a mood, not a mechanism. A model is software on hardware with an operator, a power feed, and an off switch.

5. Legislative Relevance

The four bills on today’s agenda, sponsored by Chairman Tully and Assemblyman Macurdy, address harms of exactly this kind. We are not here to take a position on them. From an engineer’s view, rules like these work when a qualified party can test compliance and it is clear who is responsible when a system fails. With that in mind, we have four requests.

First, protect people from the harms we can already prove. If someone misuses your name, your face, or your voice, you should have a remedy. People should be able to tell where an image or a recording came from, and 85 percent of New Jersey residents say they want to know when they’re dealing with AI. Platforms should answer for harmful generative AI, starting with the exploitation of children. But we would not build enforcement on detection alone. I don’t know of a watermark or a detector that can’t be beaten.

Second, require real evaluation instead of vendor self-assessment. It matters most when AI is deciding someone’s loan, job, or health care. As I’ve said before, if the only people able to evaluate frontier models are the people selling them, we get theology instead of engineering. A disclosure to the Attorney General is only as good as someone’s ability to check it.

Third, set security standards for AI in critical infrastructure and in every system the state buys. Give each system the narrowest permission that does the job, keep a human in the loop for anything that cannot be undone, log what it does, and take away access nobody is using or needs.

Fourth, invest in the capacity underneath all of this. We’re already power limited, so plan generation, transmission, and efficiency alongside data center growth. Build public computing so universities can test these systems on their own. Right now the best ideas in AI are held back less by talent than by access to computing. New Jersey has the talent. Let’s carry forward the Task Force’s recommendations on AI literacy and workforce training. Banning AI would simply push usage underground.

For the same reason, we cannot support a pause or moratorium. It would shift the work to places with cheaper power and fewer rules, and that would be worse for safety, not better.

Every one of these requests depends on people who can do the work. Someone has to test these systems independently, train the workforce that will deploy and secure them, and run the computing that makes any of it possible. The New Jersey Institute of Technology is already leading on these and many other AI-related topics. More than a thousand of our students are in AI-related degree programs, including a new bachelor’s degree in Enterprise AI built around deploying and securing these systems responsibly, and we offer free AI micro-credentials to every student, faculty and staff member. Our researchers study the bias, security, and infrastructure questions I have described, our AI Exploration Day brought New Jersey educators and policymakers to campus this past spring, and we are a partner in the state’s agreement with NVIDIA to bring serious research computing to New Jersey. We welcome the chance to continue to work with this committee on these and other AI-related issues.

6. Close

I built that first Linux supercomputer so scientists could get time on a machine they could never afford alone. The companies may own much of today’s infrastructure, but how people use these tools is still in our hands. Put serious computing in the hands of New Jersey’s researchers and students, and they will find uses for it that nobody planned but that we will all benefit from.

Thank you again, Chairman Tully, Vice Chairman Singh, and members of the committee. I am happy to take your questions.

Contact: Matt Bonasia, Chief of State Government Affairs, New Jersey Institute of Technology

References

Statements of opinion, policy recommendations, and Dr. Bader’s professional assessments (including his list of near-term harms and his risk test) are his own and are not separately referenced.

1.David A. Bader, professional biography and current appointments (titles; IEEE, ACM, AAAS and SIAM fellowships; founder of NJIT’s Department of Data Science). https://davidbader.net/
2. NJIT, “NJIT to Establish New Institute for Data Science,” GlobeNewswire, July 9, 2019. https://www.globenewswire.com/fr/news-release/2019/07/09/1880451/0/en/NJIT-to-Establish-New-Institute-for-Data-Science.html
3. Computer History Museum, Timeline of Computer History, 1998 (first Linux supercomputer, University of New Mexico). https://www.computerhistory.org/timeline/1998/
4. TOP500, June 2026 list (all 500 systems run Linux-based operating systems). https://top500.org/lists/top500/2026/06/
5. NJIT News, “NJIT’s David Bader Selected to Receive 2021 IEEE Computer Society Sidney Fernbach Award,” September 2021. https://news.njit.edu/njits-david-bader-selected-receive-2021-ieee-computer-society-sidney-fernbach-award
6. HPCwire, “35 HPC Legends: David Bader,” Class of 2025. https://www.hpcwire.com/35-hpc-legends-david-bader
7. NJIT Magazine, “Faculty Contribute to State AI Task Force,” Spring 2025 (NJIT members and four working-group areas). https://davidbader.net/post/20250515-njit-stateai/
8. Alex Konrad, “Startup Adapter Launches With $17.8M To Bring New ‘Cognition’ To AI Tools,” Upstarts Media, July 14, 2026 (Bader co-founder). https://www.upstartsmedia.com/p/deep-dive-adapter-launches-ai-cognition
9. NJIT News, “Q&A with David Bader, Recipient of Scientific Computing Honor,” May 29, 2026 (real-time AI power demand; energy, water, and sound as design constraints; cleaner-grid scheduling; waste-heat reuse; “smarter software running on the hardware we already have”). https://news.njit.edu/david-bader-receives-scientific-computing-honor
10. CNN, “The big wrinkle in the multi-trillion-dollar AI buildout,” December 19, 2025 (Bader’s estimates of GPU useful life: 18 to 24 months for training, about five more years for inference). https://edition.cnn.com/2025/12/19/tech/ai-chips-lifecycle-questions
11. David A. Bader, “An Agent That Reads the Federal Register So You Don’t Have To,” July 30, 2026 (weekly grants.gov scan, 66-faculty roster, human review before sending). https://davidbader.net/post/20260730-njit/
12. Stanford Institute for Human-Centered Artificial Intelligence, AI Index Report 2026 (53 percent adoption within three years, faster than the PC or internet; 362 AI incidents in 2025, up from 233 in 2024; spotty reporting on responsible AI benchmarks; hallucination rates of 22 to 94 percent). https://hai.stanford.edu/ai-index/2026-ai-index-report
13. TechNewsWorld, “41% of Schools Report AI Cyber Incidents,” October 7, 2025 (Bader: schools “have largely lost the ability to meaningfully prohibit AI use”; a ban “would simply push usage underground”). https://davidbader.net/post/20251007-technewsworld/
14. Rutgers University, National AI Opinion Monitor, “Report Finds Broad Adoption of AI in New Jersey, and Strong Support for Regulation,” December 11, 2025 (74, 68, 41, 28, 45, 34 vs. 24, 57, 79, 85, and 7 to 9 percent). https://www.rutgers.edu/news/report-finds-broad-adoption-ai-new-jersey-and-strong-support-regulation
15. Office of the Governor of New Jersey, release on the AI Task Force report, November 12, 2024 (NJ AI Assistant users; state AI training; Task Force recommendations). https://www.nj.gov/governor/news/news/562024/approved/20241112b.shtml
16. NJIT News, “NJIT Research Team Innovates in Princeton FlyWire Codex Challenge,” October 30, 2024 (first place; Edison Academy Magnet School students). https://news.njit.edu/njit-research-team-innovates-princeton-flywire-codex-challenge
17. NJIT News, “Fruit Fly Research Led NJIT Scientists and Edison Teens to Better AI Habits on Supercomputers,” March 3, 2025 (second place; Wulver; “weeks of literature review into minutes”). https://news.njit.edu/fruit-fly-research-led-njit-scientists-and-edison-teens-better-ai-habits-supercomputers
18. Inside Climate News, “Are You a Customer of PJM? Its Capacity Auction Will Affect Your Electricity Prices in 2026,” July 20, 2025 ($28.92 to $269.92 per megawatt-day; about 20 percent New Jersey price increase). https://insideclimatenews.org/news/20072025/pjm-capacity-auction-electricity-price-hike/
19. Power Engineering, “PJM Capacity Auction Hits Price Cap Again as Region Falls Short of Reliability Target,” December 18, 2025 ($333.44 cap; 6,623 MW shortfall; data centers about 5,100 of 5,250 MW). https://www.power-eng.com/business/pjm-capacity-auction-hits-price-cap-again-as-region-falls-short-of-reliability-target/
20. AARP, report on FBI Internet Crime Complaint Center 2025 data, April 16, 2026 ($893 million in AI-related scam losses; $352 million from older adults). https://www.aarp.org/money/scams-fraud/fbi-ftc-report-2025-losses
21. SKW Schwarz, “Agreement on the AI Omnibus” (EU AI Act timeline: high-risk employment rules to December 2, 2027; AI-generated content marking from December 2, 2026; political agreement of May 7, 2026). https://www.skwschwarz.de/en/news/agreement-ai-omnibus
22. University of Washington News, “AI tools show biases in ranking job applicants’ names according to perceived race and gender,” October 31, 2024 (hiring). https://www.washington.edu/news/2024/10/31/ai-bias-resume-screening-race-gender/
23. Lehigh University News, “AI Exhibits Racial Bias in Mortgage Underwriting Decisions,” August 20, 2024 (loans). https://news.lehigh.edu/ai-exhibits-racial-bias-in-mortgage-underwriting-decisions
24. Becker’s Hospital Review, “Mount Sinai flags AI bias in clinical decision-making,” April 7, 2025, reporting a study in Nature Medicine (emergency triage). https://www.beckershospitalreview.com/ai/mount-sinai-flags-ai-bias-in-clinical-decision-making
25. Office of the Governor of New Jersey, “Governor Murphy Signs Bill Criminalizing Deepfakes” (A3540/S2544), April 2, 2025 (third-degree offense; Francesca Mani). https://nj.gov/governor/news/news/562025/approved/20250402a.shtml
26. New Jersey Legislature, Assembly Science, Innovation and Technology Committee agenda, October 19, 2026 (A3065, A5089, A5090, A5275 and sponsors). https://www.njleg.state.nj.us/live-proceedings/2026-10-19-13:00:00/AST/Meeting
27. National Institute of Standards and Technology, NIST AI 100-4, “Reducing Risks Posed by Synthetic Content,” November 2024 (watermarks can be removed; detectors make errors). https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-4.pdf

https://www.njleg.state.nj.us/live-proceedings/2026-10-19-13:00:00/AST/Meeting